openmerit 0.3.11 → 0.4.0

package/README.md

@@ -1,42 +1,94 @@
 # OpenMerit
 
-OpenMerit provides Scout guidance for application-model choices: in the browser, from the `openmerit scout` command, and through MCP for Claude Code and Pi.
+OpenMerit finds the right model for your task from measured results on your own labelled examples, not from leaderboards alone. It runs locally against your own provider connections and your own Langfuse project. There is no OpenMerit account or login.
 
-Scout recommends candidates to evaluate. It is not a request router or proof of application-quality improvements.
+This package contains:
 
-## Connect Scout to Claude Code
+- **Scout**, a model-recommendation agent built on the [Pi coding agent](https://pi.dev). It reads your task description or codebase, finds or drafts labelled examples, and runs Scout Model Binary Search. It recommends the winner from Langfuse experiments.
+- **Pi with OpenMerit's harness changes**: `openmerit` starts Pi with Scout's tools, a `/scout` command and Scout's instructions. Pi gets its own config directory, and a Claude Code subscription token works as an Anthropic credential.
+- **The Langfuse connector**, with no dependencies. It works with Langfuse Cloud or self-hosted Langfuse.
 
-Run `npx -y openmerit@latest install --client claude-code`, then approve the device in your browser. The command installs the MCP server and shared Scout guidance and checks the connection. New Claude Code sessions can consult Scout before choosing application models. Requires the `claude` CLI on PATH; see [setup, project policy and removal](docs/scout.md#connect-a-coding-agent). Installing the npm package alone makes no agent configuration changes.
+Braintrust and the Tuner come later.
 
-## Connect Scout to Pi
+## Install and connect
 
-The thin Scout adapter supports Pi **1.5.0** through native MCP:
+```sh
+npm install -g openmerit        # Node.js 22.19 or newer
+openmerit status                # connected providers, Scout's model, Langfuse
+```
+
+**Models.** Connect either API keys or a subscription:
+- **API keys:** use any key Pi supports in the environment, such as `OPENROUTER_API_KEY`, `ANTHROPIC_API_KEY` or `OPENAI_API_KEY`.
+- **Subscription:** run `openmerit login`, which is Pi's `/login` for Claude, ChatGPT, Copilot and other subscriptions. You can also set `CLAUDE_CODE_OAUTH_TOKEN`.
+
+The connected providers decide both which model Scout itself runs on and which candidates it can test.
+
+**Langfuse.** Set `LANGFUSE_PUBLIC_KEY`, `LANGFUSE_SECRET_KEY` and `LANGFUSE_BASE_URL`. A self-hosted URL works, and `LANGFUSE_HOST` is accepted as an alias. Scout records every evaluation as a Langfuse experiment and reads the finals back from Langfuse to choose the winner.
+
+OpenMerit keeps Pi's logins and settings in `~/.openmerit/agent`. Set `OPENMERIT_HOME` or `PI_CODING_AGENT_DIR` to move them.
+
+## Use it
 
 ```sh
-npx -y openmerit@latest install --client pi --scope user
+# Ask Scout about a codebase. It reads the code and plans without spending until you add --execute.
+openmerit scout "Which model should our enrollment agent use?" --codebase ./my-app --execute
+
+# Describe a task and give labelled items (JSONL of {id?, input, expectedOutput}).
+openmerit scout "Extract chronic conditions as ICD-10-CM categories from health declarations" \
+  --dataset examples/health-declaration-codes/dataset.jsonl --providers openrouter --max-price 30 --execute
+
+# The search alone, without the agent: deterministic, for CI and other coding agents (JSON output with --json).
+openmerit run --langfuse-dataset sahra-enrollments --evaluator ./evaluator.ts --execute
+openmerit run --dataset items.jsonl --system-prompt "$(cat prompt.txt)" --grader json --execute
+
+# Interactive: Pi with Scout. Ask "which model should I use for …" or type /scout.
+openmerit
 ```
 
-Approve the matching device code in your browser, then start a new Pi session. Scope defaults to `user`; run with `--scope project` from a project's root for project-only setup. Each client/scope installation has its own key; project credentials stay outside the repository. The installed `openmerit-scout.ts` extension references private configuration and uses Pi's native MCP with the shared Scout rule and skill. Pi owns tools, model, permissions and sessions.
+From another coding agent (Claude Code, Codex, Amp and others), run `openmerit scout "<task>" --codebase . --json` or `openmerit run … --json`. A Pi user can also add Scout to their own Pi with `pi install npm:openmerit`.
+
+Every run writes `.openmerit/runs/<run-id>/report.md` and `report.json` in the project directory: the recommendation, finals, every probe, Langfuse links, spend and limitations.
 
-There are no `/openmerit` lifecycle commands, evaluation setup, scheduler, LaunchAgent, or automatic update hooks. To update, explicitly rerun the installer with the same scope. Shell `login`, `logout`, `status`, `remove` and `doctor` support `--client pi --scope user|project`. See [Pi setup and migration](docs/pi-extension.md) for paths, conflicts, old background-job removal, and SDK embedding with `createScoutPiExtension` plus Pi's explicit `createMcpExtension()`.
+## How Scout decides
 
-Package publication and website deployment are separate releases. Pi setup requires OpenMerit 0.1.12 or newer; check the resolved package version when using `@latest`.
+1. **Candidates.** Scout takes every model your connections can run that has an [Epoch Capabilities Index](https://epoch.ai/benchmarks/eci) (CC BY 4.0) and orders them by ECI, weakest first:
+   - **Published:** Epoch's own score, downloaded at run time.
+   - **Estimated or calibrated:** for models Epoch does not score, ScoutBM's estimate, shipped in [`data/eci-estimates.csv`](data/eci-estimates.csv) so no Cloudflare or OpenMerit service is needed. Estimated values come from the model's scores on Epoch's component benchmarks under Epoch's fitted curves. Calibrated values come from other benchmarks that track ECI closely (at least 12 anchor models, |r| ≥ 0.6). Both carry 2–4 points of error, shown as ± in `openmerit candidates`. Maintainers regenerate the file from a Scout catalogue snapshot with `node scripts/build-eci-estimates.ts <catalogue.json>`.
 
-## Evaluation CLI retirement — npm 0.3.0
+   Scout drops aliases, batch and free-tier variants. Each Epoch row or estimate is used once: the first connected provider, then the cheapest listing. With `--tools`, models OpenRouter lists without tool calls are dropped. Models with neither a published nor an estimated ECI are listed as unranked, never guessed.
+2. **Labelled items.** Scout needs at least five. It can use a Langfuse dataset or a local JSONL, which it pushes to Langfuse with stable item IDs. If you have neither, Scout can draft items from your fixtures and tests. It marks them as drafts for you to check.
+3. **Binary search.** Scout draws five items once. It probes the middle model: if that model passes all five, Scout moves to weaker models, and otherwise to stronger ones. The search pick is the weakest model that passed.
+4. **Finalists.** The nearest neighbours by ECI that also pass the same five make up five finalists. Once a weaker neighbour fails, Scout tries only stronger ones.
+5. **Challengers.** Up to three untested models with a higher ECI and no higher list price than the leading finalist are screened on the same five items; those that pass join the finals.
+6. **Finals in Langfuse.** Each finalist runs every item three times (`--repeats`). Each probe is one Langfuse experiment on the dataset and each final run is its own experiment, since Langfuse keeps one result per item per experiment. Scout reads the finals back from Langfuse. The winner has the most passes over all runs, then the lower measured cost per item run, then the lower ECI.
 
-OpenMerit 0.3.0 removes the standalone `openmerit eval` commands, including planning. Existing local evaluation files are not deleted. The old Scout JSON recommendation API went offline with the Scout serving deploy of 5 November 2026. Scout's conversational agent, MCP, coding-agent setup and catalogue publication remain available; see [retirement details](docs/agent-evaluations.md).
+**Errors are not verdicts.** Provider, auth, billing and harness errors are never counted as quality failures:
+- A model the provider rejects, or one your balance can't cover, is skipped and listed as untested.
+- Any other error stops the run, which keeps its completed results.
 
-## Inference Terminal retirement — npm 0.3.0
+## Evaluators
 
-OpenMerit 0.3.0 also removes the local Inference Terminal: `openmerit dash`, its `--demo` sample, collectors, repository index and browser interface. Existing `.openmerit/terminal.json`, `.openmerit/inference.jsonl` and application logs are not deleted. Older installed packages keep the command until they are upgraded. See [retirement details](docs/inference-terminal.md).
+- **Built-in Pi evaluator** (`--system-prompt`): one model call per item through Pi, so any connected provider can be a candidate. It has four graders:
+  - `json`: every expected field must match; arrays are compared as sets.
+  - `exact` and `contains`: for short answers.
+  - `judge`: a rubric checked by Scout's model.
+- **Your harness** (`--evaluator module.ts`): default-export an `EvaluatorPlugin` that runs one item through your real system and returns `{ pass, output, costUsd? }`. When a model cannot run, throw an error whose `name` is `"CandidateUnavailable"`; Scout then skips it as untested. Throw any other error to stop the run. [`examples/sahra-enrollment/evaluator.ts`](examples/sahra-enrollment/evaluator.ts), which you can copy into your project, drives a deployed multi-step tool-calling agent over HTTP. It reads cost and provider errors from that agent's own Langfuse traces.
 
-## Documentation
+## Limitations
 
-The [documentation overview](docs/index.md) links the maintained Markdown guides. The unpublished `/docs/` website shell is retired; the live Scout data policy and protected memo have separate builds.
+- Binary search assumes task quality rises with ECI, so models it skips are untested.
+- Screens run each sample once, so one unlucky run can screen a model out. Three runs per final item narrow chance differences but do not remove them.
+- Estimated and calibrated ECIs are only as fresh as the bundled CSV and carry 2–4 points of error, more than the gap between neighbouring models.
+- ECI reflects the settings Epoch evaluated, not your provider's defaults.
+- The built-in Pi evaluator makes one call per item without tools. Use an evaluator module for agents.
+- Langfuse ingestion is asynchronous, so reading the finals back can take a minute or two.
+- Langfuse Cloud's rate limit, about 30 requests a minute, slows large runs.
 
-- [Getting started](docs/getting-started.md) and [troubleshooting](docs/troubleshooting.md)
-- [Scout](docs/scout.md) and [Pi adapter](docs/pi-extension.md)
-- [Close the model-selection loop](docs/scout-close-loop.md): consult, measure, import and choose
-- Retired: [Inference Terminal](docs/inference-terminal.md) and [agent evaluations](docs/agent-evaluations.md)
-- [Security and data handling](docs/security.md)
-- [Testing](docs/testing.md) and [current limitations](docs/roadmap.md)
+## Development
+
+```sh
+npm install
+npm run check      # strict typecheck plus mocked tests (no model or Langfuse calls)
+OPENMERIT_LANGFUSE_LIVE=1 node --test src/connectors/langfuse/live.test.ts   # live Langfuse check (writes a scratch experiment)
+npm run build && npm pack
+```

package/data/eci-estimates.csv

@@ -0,0 +1,179 @@
+# ScoutBM ECI estimates. Catalogue 357d7378c58637dd44e7de07c3371f7b checked 2026-11-10T01:05:22.911Z; Epoch ECI fetched 2026-11-10T03:15:23.201Z (https://epoch.ai/data/eci_scores.csv, https://epoch.ai/data/eci_benchmarks.csv; CC BY 4.0). Regenerate with scripts/build-eci-estimates.ts.
+catalogue_model,name,openrouter_id,eci,eci_source,standard_error,benchmarks
+alibaba/qwen-max,Qwen Max,,132.68,estimated,0.94,Aider polyglot; Fiction.LiveBench; GPQA diamond; MATH level 5; OTIS Mock AIME 2024-2025
+alibaba/qwen3-next-80b-a3b-thinking,Qwen3-Next 80B-A3B (Thinking),,144.13,estimated,2.04,GPQA diamond
+alibaba/qwen3-vl-235b-a22b-thinking,Qwen3 VL 235B A22B Thinking,,146.84,estimated,12.66,MMLU
+alibaba/qwen3.5-plus,Qwen3.5 Plus,,147.58,estimated,1.25,CL-bench; CL-bench Life; Chess Puzzles; DTBench; GPQA diamond; LMCA; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; SimpleQA Verified
+alibaba/qwen3.8-flash-next,Qwen3.8 Flash Next,,154.54,estimated,4.89,GPQA diamond
+amazon/nova-2-lite-v1,Nova 2 Lite,amazon/nova-2-lite-v1,145.33,estimated,2.22,GPQA diamond
+amazon/nova-lite-v1,Nova Lite 1.0,amazon/nova-lite-v1,119.96,calibrated,4.44,arena:creative_writing; arena:expert; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision
+amazon/nova-premier-v1,Nova Premier 1.0,amazon/nova-premier-v1,136.10,estimated,1.47,GPQA diamond
+amazon/nova-pro-v1,Nova Pro 1.0,amazon/nova-pro-v1,124.03,calibrated,4.44,arena:creative_writing; arena:expert; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision
+anthropic/claude-haiku-4-5-20251001,Claude Haiku 4.5,,138.82,estimated,0.75,ARC-AGI; ARC-AGI-2; Balrog; Chess Puzzles; DTBench; DeepResearch Bench; ExploitBench; GPQA diamond; LMCA; MATH level 5; OTIS Mock AIME 2024-2025; SimpleQA Verified; Terminal Bench; WeirdML
+anthropic/claude-haiku-4.5,Claude Haiku 4.5,anthropic/claude-haiku-4.5,145.84,calibrated,3.22,arena:creative_writing; arena:document; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:webdev; arena:webdev:webdev-html; arena:webdev:webdev-react; seal:mcp_atlas
+anthropic/claude-opus-4-5,Claude Opus 4.5 (latest),,154.49,calibrated,3.22,arena:creative_writing; arena:document; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:webdev; arena:webdev:webdev-html; arena:webdev:webdev-react; seal:mcp_atlas; tau:banking_knowledge
+anthropic/claude-opus-4-5-20251101,Claude Opus 4.5,,148.83,estimated,0.52,ARC-AGI; ARC-AGI-2; Balrog; CL-bench; Chess Puzzles; Cybench; DTBench; DeepResearch Bench; EBR-bench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; Furniture Assembly; GDPval; GPQA diamond; GSO-Bench; GeoBench; LMCA; METR Time Horizons; Mystery Game Puzzles; OSWorld; OTIS Mock AIME 2024-2025; ProofBench; Remote Labor Index; SimpleBench; SimpleQA Verified; Terminal Bench; VPCT; WeirdML; WeirdML v3
+anthropic/claude-opus-4.1,Claude Opus 4.1,anthropic/claude-opus-4.1,146.03,estimated,2.33,GPQA diamond
+anthropic/claude-sonnet-4-5,Claude Sonnet 4.5 (latest),,151.55,calibrated,3.22,arena:creative_writing; arena:document; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:webdev; arena:webdev:webdev-html; arena:webdev:webdev-react; seal:mcp_atlas
+anthropic/claude-sonnet-4-5-20250929,Claude Sonnet 4.5,,144.28,estimated,0.55,ARC-AGI; ARC-AGI-2; Chess Puzzles; Cybench; DTBench; DeepResearch Bench; EBR-bench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GDPval; GPQA diamond; GSO-Bench; LMCA; MATH level 5; METR Time Horizons; Mystery Game Puzzles; OSWorld; OTIS Mock AIME 2024-2025; ProofBench; Remote Labor Index; SimpleBench; SimpleQA Verified; Terminal Bench; VPCT; WeirdML
+arcee-ai/trinity-large-thinking,Trinity Large Thinking,,142.62,estimated,1.75,GPQA diamond; Surface Evolver Bench
+baidu/ernie-4.5-300b-a47b,ERNIE 4.5 300B A47B,,146.15,estimated,2.35,GPQA diamond
+bytedance-seed/seed-1-8,Seed 1.8,,151.16,estimated,15.17,MMLU
+bytedance-seed/seed-1.6,Seed 1.6,bytedance-seed/seed-1.6,142.40,estimated,1.37,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-2-1-turbo,Seed 2.1 Turbo,bytedance-seed/seed-2-1-turbo,151.21,estimated,2.53,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-2.0-lite,Seed-2.0-Lite,bytedance-seed/seed-2.0-lite,148.96,estimated,2.22,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-2.0-mini,Seed-2.0-Mini,bytedance-seed/seed-2.0-mini,145.02,estimated,2.17,GPQA diamond
+bytedance-seed/seed-2.0-pro,Seed 2.0 Pro,,150.53,estimated,2.44,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-2.1-pro,Seed 2.1 Pro,,151.52,estimated,2.57,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-evolving,Seed Evolving,,156.46,estimated,2.95,Chess Puzzles; GPQA diamond; OTIS Mock AIME 2024-2025; SimpleQA Verified
+bytedance-seed/seed-oss-36b-instruct,Seed-OSS-36B-Instruct,,140.99,estimated,1.27,GPQA diamond; OTIS Mock AIME 2024-2025
+cohere/c4ai-aya-expanse-32b,Aya Expanse 32B,,122.11,calibrated,4.44,arena:creative_writing; arena:expert; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn
+cohere/command-a,Command A,cohere/command-a,134.68,estimated,1.44,GPQA diamond
+cohere/command-a-plus,Command A+,cohere/command-a-plus,143.58,estimated,1.98,GPQA diamond
+cohere/command-a-reasoning-08-2025,Command A Reasoning,,138.86,calibrated,4.44,bfcl:memoryKV; bfcl:memorySummary; bfcl:memoryVector
+cohere/command-r-08-2024,Command R (08-2024),cohere/command-r-08-2024,116.72,calibrated,4.44,arena:creative_writing; arena:expert; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; vals:legal_bench
+cohere/command-r-plus-08-2024,Command R+ (08-2024),cohere/command-r-plus-08-2024,124.47,estimated,5.86,MMLU
+cohere/command-r7b-12-2024,Command R7B (12-2024),cohere/command-r7b-12-2024,123.79,calibrated,4.44,bfcl:memoryKV; bfcl:memorySummary; bfcl:memoryVector
+cohere/north-mini-code-1-0,North Mini Code,,143.40,estimated,1.95,GPQA diamond
+deepseek/deepseek-chat-v3-0324,DeepSeek V3 0324,deepseek/deepseek-chat-v3-0324,137.10,estimated,0.79,Aider polyglot; DTBench; GPQA diamond; LMCA; MATH level 5; METR Time Horizons; OTIS Mock AIME 2024-2025; SimpleBench; WeirdML
+deepseek/deepseek-r1-0528,R1 0528,deepseek/deepseek-r1-0528,140.91,estimated,0.73,ARC-AGI; ARC-AGI-2; Aider polyglot; DeepResearch Bench; GPQA diamond; MATH level 5; METR Time Horizons; OTIS Mock AIME 2024-2025; SimpleBench; WeirdML
+deepseek/deepseek-r1-0528-qwen3-8b,DeepSeek R1 0528 Qwen3 8B,,118.30,estimated,2.93,GPQA diamond; OTIS Mock AIME 2024-2025
+deepseek/deepseek-r1-distill-qwen-32b,DeepSeek-R1-Distill-Qwen-32B,,137.62,estimated,1.08,Balrog; GPQA diamond; OTIS Mock AIME 2024-2025
+deepseek/deepseek-v3.1-terminus,DeepSeek V3.1 Terminus,deepseek/deepseek-v3.1-terminus,144.50,estimated,1.96,DTBench; LMCA
+deepseek/deepseek-v4-flash-vision-exp,DeepSeek V4 Flash Vision Exp,deepseek/deepseek-v4-flash-vision-exp,155.52,calibrated,1.66,fireworks:frontierswe-v2; vendor:agents-last-exam; vendor:nl2repo; vendor:terminal-bench-2.1
+deepseek/deepseek-v4.1-flash-1017,DeepSeek V4.1 Flash 1017,deepseek/deepseek-v4.1-flash-1017,159.14,estimated,0.73,APEX-Agents; ARC-AGI; ARC-AGI-2; Chess Puzzles; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; SimpleQA Verified; WeirdML v3
+fireworks/ember-1,Ember-1,fireworks/ember-1,158.01,calibrated,2.25,fireworks:bedside-bench; fireworks:big-finance-benchmark; fireworks:healthbench-professional; vals:biomysterybench; vals:code-migration; vals:emb; vals:fabv2; vals:ioi; vals:legal_research; vals:mysterymechanism; vals:public-benefits-bench; vals:sage; vals:vibe-code; vendor:deepswe-1.1; vendor:swe-bench-verified; vendor:terminal-bench-2.1
+google/gemini-2.5-flash,Gemini 2.5 Flash,google/gemini-2.5-flash,143.31,estimated,1.07,Balrog; DTBench; GPQA diamond; LMCA; SimpleBench; Terminal Bench
+google/gemini-2.5-flash-lite,Gemini 2.5 Flash Lite,google/gemini-2.5-flash-lite,138.87,estimated,1.57,GPQA diamond
+google/gemini-2.5-pro,Gemini 2.5 Pro,google/gemini-2.5-pro,143.98,estimated,0.63,ARC-AGI; ARC-AGI-2; DTBench; DeepResearch Bench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GDPval; GPQA diamond; GSO-Bench; LMCA; OTIS Mock AIME 2024-2025; Terminal Bench; WeirdML
+google/gemini-2.5-pro-preview,Gemini 2.5 Pro Preview 06-05,google/gemini-2.5-pro-preview,142.68,estimated,0.76,ARC-AGI; ARC-AGI-2; Aider polyglot; DeepResearch Bench; Fiction.LiveBench; GPQA diamond; GSO-Bench; METR Time Horizons; Remote Labor Index; SimpleBench; VPCT
+google/gemini-3-flash-preview,Gemini 3 Flash Preview,google/gemini-3-flash-preview,148.90,estimated,0.58,ARC-AGI; ARC-AGI-2; Balrog; Chess Puzzles; DTBench; DeepResearch Bench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; GSO-Bench; GeoBench; LMCA; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; ProofBench; SimpleBench; SimpleQA Verified; Terminal Bench; VPCT; WeirdML
+google/gemini-3.1-flash-lite-preview,Gemini 3.1 Flash Lite Preview,google/gemini-3.1-flash-lite-preview,149.85,estimated,3.16,GPQA diamond
+google/gemini-3.1-pro-preview,Gemini 3.1 Pro Preview,google/gemini-3.1-pro-preview,155.79,estimated,0.49,APEX-Agents; ARC-AGI; ARC-AGI-2; Balrog; CL-bench; CL-bench Life; Chess Puzzles; DTBench; DeepResearch Bench; EBR-bench; ExploitBench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; Furniture Assembly; GPQA diamond; GSO-Bench; LMCA; METR Time Horizons; MirrorCode; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; PostTrainBench; ProofBench; SimpleBench; SimpleQA Verified; Terminal Bench; WeirdML
+google/gemma-3-12b-it,Gemma 3 12B,google/gemma-3-12b-it,129.93,estimated,1.11,DTBench; GPQA diamond; LMCA; OTIS Mock AIME 2024-2025
+google/gemma-3-4b-it,Gemma 3 4B,google/gemma-3-4b-it,126.68,estimated,1.32,DTBench; GPQA diamond; LMCA; OTIS Mock AIME 2024-2025
+google/gemma-4-12b-it,Gemma 4 12B IT,,144.92,estimated,2.15,GPQA diamond
+ibm-granite/granite-4.2-8b,Granite 4.2 8B,ibm-granite/granite-4.2-8b,138.51,estimated,1.55,GPQA diamond
+inception/mercury-2,Mercury 2,,142.54,estimated,1.67,GPQA diamond; WeirdML
+inception/mercury-2.5,Mercury 2.5,,144.66,estimated,12.30,ProofBench
+inception/mercury-3,Mercury 3,,151.74,estimated,3.74,GPQA diamond
+inclusionai/ling-3.0-flash,Ling 3.0 Flash,inclusionai/ling-3.0-flash,148.16,calibrated,2.33,vendor:browsecomp; vendor:ifbench; vendor:livecodebench-v6; vendor:swe-bench-multilingual; vendor:swe-bench-pro; vendor:terminal-bench-2.1
+inclusionai/ling-3.0-flash-fin,Ling 3.0 Flash Fin,,147.90,calibrated,2.78,aa:analystAgent; aa:automationbench; aa:critpt; aa:gdpPdf; aa:gdpvalRating; aa:hle; aa:hleDiamond; aa:lcr; aa:mlcr; aa:scicode; aa:tauBanking; aa:terminalBench21; aa:terminalBench40
+inclusionai/ling-3.0-flash-vl,Ling 3.0 Flash VL,inclusionai/ling-3.0-flash-vl,149.31,estimated,3.02,GPQA diamond
+inclusionai/ling-3.0-tiny,Ling 3.0 Tiny,inclusionai/ling-3.0-tiny,142.40,estimated,1.84,GPQA diamond
+inclusionai/ling-3.1-flash,Ling 3.1 Flash,inclusionai/ling-3.1-flash,156.79,calibrated,4.51,aa:automationbench; aa:critpt; aa:gdpPdf; aa:gdpvalRating; aa:hle; aa:lcr; aa:mlcr; aa:scicode; aa:terminalBench40; opencompass:abstraction; opencompass:agent; opencompass:knowledge
+liquid/lfm-2.5-2.6b,LFM2.5-2.6B,liquid/lfm-2.5-2.6b,135.72,estimated,1.46,GPQA diamond
+liquid/lfm-3-1.7b,LFM3-1.7B,liquid/lfm-3-1.7b,138.41,estimated,1.55,GPQA diamond
+meituan/longcat-2.0,LongCat 2.0,meituan/longcat-2.0,144.51,estimated,2.10,GPQA diamond
+meta-llama/llama-3.1-70b-instruct,Llama 3.1 70B Instruct,meta-llama/llama-3.1-70b-instruct,128.54,estimated,0.94,Balrog; DTBench; GPQA diamond; LMCA; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; The Agent Company; WeirdML
+meta-llama/llama-3.1-8b-instruct,Llama 3.1 8B Instruct,meta-llama/llama-3.1-8b-instruct,123.18,estimated,1.18,Balrog; DTBench; GPQA diamond; GSM8K; LMCA; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; PIQA; WeirdML
+meta-llama/llama-3.2-1b-instruct,Llama 3.2 1B Instruct,meta-llama/llama-3.2-1b-instruct,122.38,estimated,2.03,Balrog; GPQA diamond; OTIS Mock AIME 2024-2025
+meta-llama/llama-3.2-3b-instruct,Llama 3.2 3B Instruct,meta-llama/llama-3.2-3b-instruct,125.70,estimated,1.67,Balrog; GPQA diamond; MMLU
+meta-llama/llama-3.3-70b-instruct,Llama 3.3 70B Instruct,meta-llama/llama-3.3-70b-instruct,130.09,estimated,0.91,Balrog; DTBench; Fiction.LiveBench; GPQA diamond; LMCA; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; SimpleBench; WeirdML
+meta/llama-3.2-11b-vision-instruct,Llama-3.2-11B-Vision-Instruct,,108.55,estimated,4.21,Balrog; MMLU
+meta/llama-3.2-90b-vision-instruct,Llama-3.2-90B-Vision-Instruct,,127.09,estimated,1.11,Balrog; GPQA diamond; GeoBench; MATH level 5; MMLU; OTIS Mock AIME 2024-2025
+meta/muse-ember,Muse Ember,meta/muse-ember,160.08,calibrated,3.12,aa:automationbench; aa:critptV2Hard; aa:frontiercode20; aa:gdpPdf; aa:hleDiamond; aa:mlcr; aa:omniscienceAccuracy; aa:terminalBench40; aa:terminalBenchScience; vals:emb-v1-1; vals:fabv2-1; vals:legal_research; vals:vibe-code
+meta/muse-glimmer-30b,Muse Glimmer 30B,meta/muse-glimmer-30b,147.54,estimated,2.61,GPQA diamond
+microsoft/wizardlm-2-8x22b,WizardLM 2 8x22B,microsoft/wizardlm-2-8x22b,126.30,estimated,1.24,GPQA diamond; MATH level 5
+minimax/MiniMax-M2,MiniMax-M2,,144.27,estimated,1.35,GPQA diamond; Terminal Bench
+minimax/MiniMax-M2.1,MiniMax-M2.1,,145.34,estimated,1.34,GPQA diamond; Terminal Bench
+minimax/minimax-m1-80k,MiniMax M1 80K,,140.97,estimated,1.63,Fiction.LiveBench; GPQA diamond
+minimax/minimax-m2.5,MiniMax M2.5,minimax/minimax-m2.5,146.86,estimated,0.75,ARC-AGI; ARC-AGI-2; CL-bench; CL-bench Life; ProofBench; Terminal Bench
+minimax/minimax-m2.7,MiniMax M2.7,minimax/minimax-m2.7,146.52,estimated,1.37,ExploitBench; ProofBench; Terminal Bench; WeirdML
+minimax/minimax-m3.1-flash,MiniMax M3.1 Flash,minimax/minimax-m3.1-flash,157.54,calibrated,1.66,aa:automationbench; aa:critpt; aa:critptV2Hard; aa:frontiercode; aa:frontiercode20; aa:gdpPdf; aa:gdpvalRating; aa:hle; aa:hleDiamond; aa:lcr; aa:mlcr; aa:mmmuPro; aa:omniscienceAccuracy; aa:scicode; aa:terminalBench40; aa:terminalBenchScience; arena:webdev; arena:webdev:webdev-html; arena:webdev:webdev-react; fireworks:o11y-bench; livebench:browser; livebench:coding; livebench:design; livebench:long_agentic; livebench:reasoning; livebench:research; opencompass:abstraction; opencompass:agent; opencompass:knowledge; seal:prbench-finance; vals:emb-v1-1; vals:fabv2-1; vals:legal_research; vals:sage; vals:vibe-code; vendor:agents-last-exam; vendor:charxiv-r; vendor:cybergym; vendor:deepswe-1.1; vendor:nl2repo; vendor:terminal-bench-4.0
+mistral/mistral-7b-instruct-v0.3,Mistral 7B Instruct v0.3,,112.98,estimated,2.74,GPQA diamond; MATH level 5; MMLU; OTIS Mock AIME 2024-2025
+mistral/mistral-small-2506,Mistral Small 3.2,,140.42,estimated,2.68,DTBench
+mistral/mixtral-8x7b-instruct-v0.1,Mixtral 8x7B Instruct v0.1,,120.61,estimated,1.63,GPQA diamond; MATH level 5
+mistral/pixtral-12b,Pixtral 12B,,118.00,estimated,5.06,MMLU
+mistralai/devstral-2512,Devstral 2 2512,mistralai/devstral-2512,138.40,calibrated,3.22,arena:webdev; arena:webdev:webdev-html
+mistralai/ministral-3b-latest,Ministral 3B (latest),,134.91,estimated,1.45,GPQA diamond
+mistralai/ministral-8b-latest,Ministral 8B (latest),,139.70,estimated,1.62,GPQA diamond
+mistralai/mistral-large-2407,Mistral Large 2407,mistralai/mistral-large-2407,130.98,estimated,0.96,DTBench; GPQA diamond; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; SimpleBench
+mistralai/mistral-large-2512,Mistral Large 3 2512,mistralai/mistral-large-2512,133.80,estimated,1.19,DTBench; GPQA diamond; LMCA
+mistralai/mistral-large-latest,Mistral Large (latest),,131.67,estimated,1.46,GPQA diamond
+mistralai/mistral-medium-3.1,Mistral Medium 3.1,mistralai/mistral-medium-3.1,142.80,calibrated,4.44,arena:creative_writing; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision; arena:vision:creative_writing_vision; arena:vision:diagram; arena:vision:homework; arena:vision:ocr
+mistralai/mistral-medium-latest,Mistral Medium (latest),,145.90,calibrated,7.87,seal:prbench-finance
+mistralai/mistral-small-24b-instruct-2501,Mistral Small 3,mistralai/mistral-small-24b-instruct-2501,132.15,estimated,1.45,GPQA diamond
+mistralai/mistral-small-2603,Mistral Small 4,mistralai/mistral-small-2603,139.93,estimated,1.26,DTBench; GPQA diamond; LMCA
+mistralai/mistral-small-3.1-24b-instruct,Mistral Small 3.1 24B,mistralai/mistral-small-3.1-24b-instruct,132.29,estimated,1.42,GPQA diamond; MMLU
+mistralai/mistral-small-3.2-24b-instruct,Mistral Small 3.2 24B,mistralai/mistral-small-3.2-24b-instruct,132.34,estimated,1.42,GPQA diamond; MMLU
+mistralai/mistral-small-latest,Mistral Small (latest),,141.45,estimated,1.75,GPQA diamond
+mistralai/mixtral-8x22b-instruct,Mixtral 8x22B Instruct,mistralai/mixtral-8x22b-instruct,133.09,estimated,2.09,Cybench; DTBench; WeirdML
+moonshotai/kimi-k2-0905,Kimi K2 0905,moonshotai/kimi-k2-0905,143.54,estimated,1.94,GPQA diamond; MMLU
+moonshotai/kimi-k2-instruct,Kimi K2 Instruct,,142.16,estimated,1.11,Aider polyglot; Fiction.LiveBench; GPQA diamond; GSO-Bench; MMLU; SimpleBench; Terminal Bench; WeirdML
+nvidia/llama-3.1-nemotron-70b-instruct,Llama 3.1 Nemotron 70B Instruct,,129.67,estimated,6.79,MMLU
+nvidia/llama-3.3-nemotron-super-49b-v1,Llama 3.3 Nemotron Super 49B v1,,139.65,estimated,1.62,GPQA diamond
+nvidia/nemotron-3-nano-30b-a3b,Nemotron 3 Nano 30B A3B,nvidia/nemotron-3-nano-30b-a3b,139.54,estimated,1.61,GPQA diamond
+nvidia/nemotron-3-nano-omni-30b-a3b-reasoning,Nemotron 3 Nano Omni 30B A3B Reasoning,,132.69,estimated,1.44,GPQA diamond
+nvidia/nemotron-3-super-120b-a12b,Nemotron 3 Super,nvidia/nemotron-3-super-120b-a12b,145.01,estimated,1.88,GPQA diamond; WeirdML
+nvidia/nemotron-3.5-lightning,Nemotron 3.5 Lightning,nvidia/nemotron-3.5-lightning,143.05,estimated,1.91,GPQA diamond
+nvidia/nemotron-3.5-super-vl,Nemotron 3.5 Super VL,nvidia/nemotron-3.5-super-vl,152.00,estimated,3.83,GPQA diamond
+nvidia/nemotron-nano-12b-v2-vl,Nemotron Nano 12B v2 VL,,133.94,estimated,1.44,GPQA diamond
+nvidia/nemotron-nano-9b-v2,Nemotron Nano 9B v2,,138.65,estimated,1.56,GPQA diamond
+openai/gpt-3.5-turbo-0613,GPT-3.5 Turbo (older v0613),openai/gpt-3.5-turbo-0613,113.67,estimated,2.60,BBH; GSM8K; MMLU; Winogrande
+openai/gpt-4,GPT-4,openai/gpt-4,129.18,estimated,1.51,GPQA diamond; MMLU
+openai/gpt-4o,GPT-4o,openai/gpt-4o,132.63,calibrated,4.44,arena:creative_writing; arena:expert; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision
+openai/gpt-4o-2024-05-13,GPT-4o (2024-05-13),openai/gpt-4o-2024-05-13,131.13,estimated,0.85,Balrog; DTBench; GPQA diamond; LMCA; MATH level 5; METR Time Horizons; MMLU; OTIS Mock AIME 2024-2025
+openai/gpt-4o-2024-08-06,GPT-4o (2024-08-06),openai/gpt-4o-2024-08-06,131.51,estimated,0.93,Aider polyglot; CadEval; Chess Puzzles; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; SimpleBench; SimpleQA Verified
+openai/gpt-4o-2024-11-20,GPT-4o (2024-11-20),openai/gpt-4o-2024-11-20,129.76,estimated,0.94,ARC-AGI; ARC-AGI-2; Aider polyglot; Cybench; GDPval; GPQA diamond; GSO-Bench; GeoBench; MATH level 5; MMLU; OTIS Mock AIME 2024-2025; The Agent Company; VPCT; WeirdML
+openai/gpt-4o-mini-2024-07-18,GPT-4o-mini (2024-07-18),openai/gpt-4o-mini-2024-07-18,128.64,estimated,0.93,ARC-AGI-2; Aider polyglot; Balrog; Chess Puzzles; DTBench; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; GSM8K; GeoBench; LMCA; MATH level 5; MMLU; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; PIQA; SimpleBench; SimpleQA Verified; VPCT; WeirdML
+openai/gpt-5.1-codex,GPT-5.1-Codex,openai/gpt-5.1-codex,150.24,estimated,1.53,Terminal Bench
+openai/gpt-5.1-codex-max,GPT-5.1-Codex-Max,openai/gpt-5.1-codex-max,151.26,estimated,1.18,METR Time Horizons; ProofBench; Terminal Bench
+openai/gpt-5.1-codex-mini,GPT-5.1-Codex-Mini,openai/gpt-5.1-codex-mini,147.65,estimated,1.55,Terminal Bench
+openai/gpt-5.2-chat,GPT-5.2 Chat,openai/gpt-5.2-chat,154.94,calibrated,4.44,arena:creative_writing; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision; arena:vision:creative_writing_vision; arena:vision:diagram; arena:vision:homework; arena:vision:ocr
+openai/gpt-5.2-codex,GPT-5.2-Codex,openai/gpt-5.2-codex,153.66,estimated,1.71,Terminal Bench
+openai/gpt-5.3-codex,GPT-5.3-Codex,openai/gpt-5.3-codex,156.34,estimated,1.53,GPQA diamond; METR Time Horizons; Terminal Bench; WeirdML
+openai/gpt-5.6,GPT-5.6,,159.76,calibrated,1.80,fireworks:frontierswe-v2; seal:mcp_atlas
+particle-ai/particle-1,Particle-1,particle-ai/particle-1,151.71,estimated,3.63,GPQA diamond; MMLU
+qwen/qwen-2.5-72b-instruct,Qwen2.5 72B Instruct,qwen/qwen-2.5-72b-instruct,130.52,estimated,0.83,Balrog; DTBench; GPQA diamond; LMCA; MATH level 5; METR Time Horizons; MMLU; OTIS Mock AIME 2024-2025; The Agent Company; WeirdML
+qwen/qwen-2.5-7b-instruct,Qwen2.5 7B Instruct,qwen/qwen-2.5-7b-instruct,127.57,estimated,1.22,Balrog; DTBench; GPQA diamond; LMCA; MMLU; OTIS Mock AIME 2024-2025
+qwen/qwen-2.5-coder-32b-instruct,Qwen2.5 Coder 32B Instruct,qwen/qwen-2.5-coder-32b-instruct,125.13,estimated,4.11,Aider polyglot; GSM8K; MMLU
+qwen/qwen-plus,Qwen-Plus,qwen/qwen-plus,131.32,estimated,1.03,GPQA diamond; MATH level 5; OTIS Mock AIME 2024-2025
+qwen/qwen2.5-vl-72b-instruct,Qwen2.5 VL 72B Instruct,qwen/qwen2.5-vl-72b-instruct,126.67,estimated,11.48,GeoBench; OSWorld
+qwen/qwen3-235b-a22b-2507,Qwen3 235B A22B Instruct 2507,qwen/qwen3-235b-a22b-2507,140.07,estimated,0.82,ARC-AGI; ARC-AGI-2; Aider polyglot; DTBench; GPQA diamond; LMCA; WeirdML
+qwen/qwen3-235b-a22b-thinking-2507,Qwen3 235B A22B Thinking 2507,qwen/qwen3-235b-a22b-thinking-2507,144.97,estimated,1.21,DTBench; GPQA diamond; LMCA; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; SimpleQA Verified; WeirdML
+qwen/qwen3-30b-a3b-instruct-2507,Qwen3 30B A3B Instruct 2507,qwen/qwen3-30b-a3b-instruct-2507,137.90,estimated,0.97,DTBench; GPQA diamond; LMCA; OTIS Mock AIME 2024-2025
+qwen/qwen3-coder,Qwen3 Coder 480B A35B,qwen/qwen3-coder,142.37,estimated,1.78,GSO-Bench; Terminal Bench; WeirdML
+qwen/qwen3-coder-next,Qwen3 Coder Next,qwen/qwen3-coder-next,141.59,estimated,1.60,GPQA diamond; WeirdML
+qwen/qwen3-next-80b-a3b-instruct,Qwen3 Next 80B A3B Instruct,qwen/qwen3-next-80b-a3b-instruct,142.37,estimated,1.84,GPQA diamond
+qwen/qwen3-vl-235b-a22b-instruct,Qwen3 VL 235B A22B Instruct,qwen/qwen3-vl-235b-a22b-instruct,142.97,estimated,10.84,MMLU
+qwen/qwen3-vl-30b-a3b-instruct,Qwen3 VL 30B A3B Instruct,qwen/qwen3-vl-30b-a3b-instruct,140.97,estimated,1.69,GPQA diamond; MMLU
+qwen/qwen3-vl-30b-a3b-thinking,Qwen3 VL 30B A3B Thinking,qwen/qwen3-vl-30b-a3b-thinking,142.76,estimated,1.85,GPQA diamond; MMLU
+qwen/qwen3-vl-8b-instruct,Qwen3 VL 8B Instruct,qwen/qwen3-vl-8b-instruct,130.30,estimated,6.92,MMLU
+qwen/qwen3.5-122b-a10b,Qwen3.5-122B-A10B,qwen/qwen3.5-122b-a10b,147.45,estimated,1.55,DTBench; GPQA diamond; LMCA; Mystery Game Puzzles
+qwen/qwen3.5-27b,Qwen3.5-27B,qwen/qwen3.5-27b,146.28,estimated,1.41,DTBench; GPQA diamond; LMCA; WeirdML
+qwen/qwen3.5-flash-02-23,Qwen3.5-Flash,qwen/qwen3.5-flash-02-23,144.29,estimated,1.04,Chess Puzzles; DTBench; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; LMCA; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; SimpleQA Verified
+qwen/qwen3.7-flash,Qwen3.7 Flash,qwen/qwen3.7-flash,144.42,estimated,1.26,FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; Mystery Game Puzzles; OTIS Mock AIME 2024-2025
+qwen/qwen3.8-2.4t-a95b,Qwen3.8 2.4T A95B,qwen/qwen3.8-2.4t-a95b,156.48,estimated,5.95,GPQA diamond
+qwen/qwen3.8-flash,Qwen3.8 Flash,qwen/qwen3.8-flash,154.18,estimated,4.73,GPQA diamond
+reflection/beam-1.0-501b-a23b,Beam 1.0 501B A23B,reflection/beam-1.0-501b-a23b,148.39,estimated,0.85,APEX-Agents; ARC-AGI; ARC-AGI-2; GPQA diamond; ProofBench
+rekaai/reka-flash-3,Reka Flash 3,rekaai/reka-flash-3,134.77,estimated,1.41,Balrog; GPQA diamond
+sakana/sakana-namazu,Sakana Namazu,sakana/sakana-namazu,151.20,calibrated,2.35,vendor:livecodebench-v6; vendor:mmlu-pro
+sarvam/sarvam-105b,Sarvam 105B,,143.77,estimated,1.97,GPQA diamond; MMLU
+stepfun/step-3.5-flash,Step 3.5 Flash,stepfun/step-3.5-flash,145.57,calibrated,2.35,arena:creative_writing; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:factuality:industry_legal_and_government; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; vendor:aime-2025; vendor:browsecomp; vendor:livecodebench-v6; vendor:swe-bench-verified; vendor:terminal-bench-2
+stepfun/step-3.7-flash,Step 3.7 Flash,stepfun/step-3.7-flash,149.22,calibrated,2.46,vendor:swe-bench-pro; vendor:terminal-bench-2.1
+stepfun/step-5-preview,Step 5 Preview,stepfun/step-5-preview,155.57,estimated,1.47,ProofBench
+swiss-ai/apertus-70b,Apertus 70B,,125.37,estimated,1.82,GPQA diamond
+tencent/hy3,Hy3,tencent/hy3,152.49,estimated,4.01,GPQA diamond
+tencent/hy3-preview,Hy3 preview,tencent/hy3-preview,145.69,calibrated,3.41,arena:creative_writing; arena:expert; arena:factuality; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:webdev; arena:webdev:webdev-react
+tencent/hy4,Hy4,tencent/hy4,160.41,calibrated,3.12,aa:automationbench; aa:critptV2Hard; aa:frontiercode20; aa:gdpPdf; aa:hleDiamond; aa:omniscienceAccuracy; aa:terminalBench40; aa:terminalBenchScience; opencompass:abstraction; opencompass:agent; opencompass:knowledge; vals:emb-v1-1; vals:fabv2-1; vals:legal_research; vals:vibe-code
+tencent/hy4-preview,Hy4 preview,tencent/hy4-preview,160.17,estimated,1.79,GPQA diamond; ProofBench
+upstage/solar-mini4,Solar Mini 4,upstage/solar-mini4,148.40,calibrated,4.54,aa:automationbench; aa:critpt; aa:gdpPdf; aa:gdpvalRating; aa:hle; aa:lcr; aa:mlcr; aa:omniscienceAccuracy; aa:scicode; aa:terminalBench40
+upstage/solar-pro4,Solar Pro 4,upstage/solar-pro4,151.57,estimated,3.69,GPQA diamond
+x-ai/grok-4.20-0309-reasoning,Grok 4.20 (Reasoning),,150.32,estimated,1.12,Chess Puzzles; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; OTIS Mock AIME 2024-2025; ProofBench; SimpleQA Verified
+x-ai/grok-4.3,Grok 4.3,x-ai/grok-4.3,149.23,estimated,0.98,Chess Puzzles; DTBench; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; GPQA diamond; LMCA; OTIS Mock AIME 2024-2025; ProofBench; SimpleQA Verified; WeirdML
+x-ai/grok-build-0.1,Grok Build 0.1,x-ai/grok-build-0.1,146.05,calibrated,3.35,vals:vibe-code
+xiaomi/mimo-v2-flash,MiMo-V2-Flash,,147.67,estimated,2.64,GPQA diamond
+xiaomi/mimo-v2.5,MiMo-V2.5,xiaomi/mimo-v2.5,150.95,estimated,2.66,ProofBench
+xiaomi/mimo-v2.5-pro,MiMo-V2.5-Pro,xiaomi/mimo-v2.5-pro,142.66,estimated,1.34,DTBench; GPQA diamond; LMCA; MMLU; ProofBench
+xiaomi/mimo-v2.6-flash,MiMo-V2.6-Flash,xiaomi/mimo-v2.6-flash,158.53,estimated,1.54,ProofBench
+xiaomi/mimo-v2.6-pro,MiMo-V2.6-Pro,xiaomi/mimo-v2.6-pro,159.52,estimated,1.56,APEX-Agents; ProofBench
+z-ai/glm-4.5,GLM 4.5,z-ai/glm-4.5,144.06,estimated,1.79,GPQA diamond; WeirdML
+z-ai/glm-4.5-air,GLM 4.5 Air,z-ai/glm-4.5-air,143.09,estimated,1.88,GPQA diamond; GSO-Bench
+z-ai/glm-4.5v,GLM 4.5V,z-ai/glm-4.5v,134.67,calibrated,4.88,arena:creative_writing; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_life_and_physical_and_social_science; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision; arena:vision:diagram; arena:vision:ocr
+z-ai/glm-4.6,GLM 4.6,z-ai/glm-4.6,144.00,estimated,1.36,GPQA diamond; Terminal Bench
+z-ai/glm-4.6v,GLM 4.6V,z-ai/glm-4.6v,138.68,calibrated,4.88,arena:creative_writing; arena:factuality; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_life_and_physical_and_social_science; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision; arena:vision:diagram; arena:vision:ocr
+z-ai/glm-4.7-flash,GLM 4.7 Flash,z-ai/glm-4.7-flash,136.83,estimated,1.08,GPQA diamond; OTIS Mock AIME 2024-2025
+z-ai/glm-5.3,GLM 5.3,z-ai/glm-5.3,156.86,estimated,0.69,APEX-Agents; Chess Puzzles; DTBench; FrontierCode; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; FrontierSWE; GPQA diamond; LMCA; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; ProofBench; SimpleQA Verified; WeirdML
+z-ai/glm-5.3-flash,GLM 5.3 Flash,z-ai/glm-5.3-flash,152.40,estimated,0.67,APEX-Agents; ARC-AGI; ARC-AGI-2; Chess Puzzles; FrontierCode; FrontierMath-Tier-4-v2-Private; FrontierMath-Tiers-1-3-v2-Private; FrontierSWE; GPQA diamond; Mystery Game Puzzles; OTIS Mock AIME 2024-2025; ProofBench; Surface Evolver Bench
+z-ai/glm-5v-turbo,GLM 5V Turbo,z-ai/glm-5v-turbo,148.70,calibrated,3.14,arena:creative_writing; arena:document; arena:expert; arena:factuality; arena:factuality:industry_business_and_management_and_financial_operations; arena:hard_prompts; arena:industry_business_and_management_and_financial_operations; arena:industry_entertainment_and_sports_and_media; arena:industry_legal_and_government; arena:industry_life_and_physical_and_social_science; arena:industry_medicine_and_healthcare; arena:industry_writing_and_literature_and_language; arena:instruction_following; arena:longer_query; arena:multi_turn; arena:vision; arena:vision:creative_writing_vision; arena:vision:diagram; arena:vision:homework; arena:vision:ocr; arena:webdev; arena:webdev:image_to_webdev; arena:webdev:webdev-react

package/dist/cli.d.ts

@@ -0,0 +1,2 @@
+#!/usr/bin/env node
+export {};

package/dist/cli.js

@@ -0,0 +1,180 @@
+#!/usr/bin/env node
+import { parseArgs } from "node:util";
+import { resolve } from "node:path";
+import { configureHarness, openRuntime, pickScoutModel } from "./pi/harness.js";
+const HELP = `OpenMerit: find the right model for your task.
+
+Usage:
+  openmerit                          Pi coding agent with OpenMerit Scout (/scout, scout_* tools)
+  openmerit login                    Connect a subscription or API key (Pi's /login)
+  openmerit status                   Connected providers, Scout's model and Langfuse
+  openmerit scout "<task>" [options] Scout agent: reads the task and codebase, runs the search, recommends a model
+  openmerit run [options]            The search alone, without the agent (for CI and coding agents)
+  openmerit candidates [options]     ECI-ranked candidates the connected providers can run
+  openmerit pi [pi args]             Pi itself, with the Scout extension
+
+Options:
+  --codebase <dir>          Project Scout reads (default: current directory)
+  --langfuse-dataset <name> Labelled Langfuse dataset
+  --dataset <file.jsonl>    Local labelled items {id?, input, expectedOutput}; pushed to Langfuse
+  --evaluator <module.ts>   Evaluator module that runs your real harness
+  --system-prompt <text>    run: production system prompt for the built-in Pi evaluator
+  --grader <exact|contains|json|judge>  run: how the Pi evaluator checks a reply (default json)
+  --rubric <text>           run: rubric for the judge grader
+  --task <text>             run: task description recorded with the experiments
+  --providers <a,b>         Candidate providers (default: every connected provider)
+  --tools                   Candidates must support tool calls
+  --max-price <usd>         Max list price per million input+output tokens
+  --exclude <a,b>           Drop candidates whose provider/id contains any of these
+  --model <provider/id>     Model Scout itself runs on
+  --sample-size <n>         Items per binary-search probe (default 5)
+  --finalists <n>           Finalists run on every item (default 5)
+  --repeats <n>             Runs of every item per finalist (default 3)
+  --execute                 Allow spending provider credit (otherwise plan only)
+  --json                    Machine-readable output
+
+Credentials come from Pi: API keys in the environment (OPENROUTER_API_KEY, ANTHROPIC_API_KEY,
+OPENAI_API_KEY, ...), subscriptions via \`openmerit login\`, or CLAUDE_CODE_OAUTH_TOKEN. Langfuse:
+LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, LANGFUSE_BASE_URL (Cloud or self-hosted). No OpenMerit account.`;
+const [command = "", ...rest] = process.argv.slice(2);
+configureHarness();
+async function launchPi(args) {
+    const { main } = await import("@earendil-works/pi-coding-agent");
+    const { createScoutExtension } = await import("./pi/extension.js");
+    await main(args, { extensionFactories: [{ name: "openmerit-scout", factory: createScoutExtension({ allowSpend: "ask" }) }] });
+}
+function options(args) {
+    return parseArgs({
+        args, allowPositionals: true,
+        options: {
+            codebase: { type: "string" }, "langfuse-dataset": { type: "string" }, dataset: { type: "string" }, evaluator: { type: "string" },
+            "system-prompt": { type: "string" }, grader: { type: "string" }, rubric: { type: "string" }, task: { type: "string" },
+            providers: { type: "string" }, tools: { type: "boolean" }, "max-price": { type: "string" }, exclude: { type: "string" },
+            model: { type: "string" }, "sample-size": { type: "string" }, finalists: { type: "string" }, repeats: { type: "string" },
+            execute: { type: "boolean", default: false }, json: { type: "boolean", default: false }, help: { type: "boolean" },
+        },
+    });
+}
+const list = (value) => value?.split(",").map(item => item.trim()).filter(Boolean);
+const filterFrom = (values) => ({
+    ...(values.providers ? { providers: list(values.providers) } : {}), ...(values.tools ? { tools: true } : {}),
+    ...(values["max-price"] ? { maxPrice: Number(values["max-price"]) } : {}), ...(values.exclude ? { exclude: list(values.exclude) } : {}),
+});
+const log = (line) => process.stderr.write(`${line}\n`);
+async function hostFor(values) {
+    const { runtime, caller, models } = await openRuntime();
+    const scout = await pickScoutModel(runtime, values.model).catch(() => null);
+    return { runtime, scout, host: { caller, models, cwd: resolve(values.codebase ?? "."), env: process.env, log, ...(scout ? { scoutModel: `${scout.provider}/${scout.id}` } : {}) } };
+}
+async function main() {
+    if (command === "")
+        return launchPi(rest);
+    if (command === "pi")
+        return launchPi(rest);
+    if (command === "login")
+        return launchPi(["/login", ...rest]);
+    if (command === "help" || command === "--help" || command === "-h")
+        return void process.stdout.write(`${HELP}\n`);
+    if (command === "--version" || command === "-v") {
+        const { readFile } = await import("node:fs/promises");
+        const pkg = JSON.parse(await readFile(new URL("../package.json", import.meta.url), "utf8"));
+        return void process.stdout.write(`${pkg.version}\n`);
+    }
+    const { values, positionals } = options(rest);
+    if (values.help)
+        return void process.stdout.write(`${HELP}\n`);
+    const service = await import("./scout/service.js");
+    if (command === "status") {
+        const { host } = await hostFor(values);
+        const status = await service.connections(host);
+        if (values.json)
+            return void process.stdout.write(`${JSON.stringify(status, null, 2)}\n`);
+        process.stdout.write(`Providers: ${Object.entries(status.providers).map(([name, count]) => `${name} (${count} models)`).join(", ") || "none"}\n`);
+        process.stdout.write(`Scout model: ${status.scoutModel ?? "none"}\n`);
+        process.stdout.write(`Langfuse: ${status.langfuse.connected ? `${status.langfuse.project} at ${status.langfuse.baseUrl}; datasets ${status.langfuse.datasets?.join(", ") || "none"}` : `not connected${status.langfuse.error ? ` (${status.langfuse.error})` : ""}`}\n`);
+        return;
+    }
+    if (command === "candidates") {
+        const { host } = await hostFor(values);
+        const ranked = await service.candidates(host, filterFrom(values));
+        if (values.json)
+            return void process.stdout.write(`${JSON.stringify(ranked, null, 2)}\n`);
+        for (const item of ranked.candidates)
+            process.stdout.write(`${item.eci.toFixed(1).padStart(6)}  ${item.eciSource === "published" ? "         " : `${item.eciSource.slice(0, 4)} ±${(item.eciError ?? 0).toFixed(1)}`.padEnd(9)}  ${item.model}  ${item.price === null ? "?" : `$${item.price.toFixed(2)}/M`}\n`);
+        process.stdout.write(`${ranked.candidates.length} candidates (${ranked.counts.published} published ECI, ${ranked.counts.estimated} estimated, ${ranked.counts.calibrated} calibrated from data/eci-estimates.csv), ${ranked.unranked.length} unranked, ${ranked.dropped.length} filtered. ${ranked.eci.attribution}\n`);
+        return;
+    }
+    if (command === "run") {
+        const { host } = await hostFor(values);
+        const evaluator = values.evaluator
+            ? { type: "module", path: values.evaluator }
+            : (() => {
+                if (!values["system-prompt"])
+                    throw new Error("run needs --evaluator <module> or --system-prompt for the built-in Pi evaluator");
+                const kind = values.grader ?? "json";
+                const grader = kind === "judge" ? { type: "judge", rubric: values.rubric ?? "The output must agree with the expected output." }
+                    : kind === "exact" ? { type: "exact" } : kind === "contains" ? { type: "contains" } : { type: "json" };
+                return { type: "pi", systemPrompt: values["system-prompt"], grader };
+            })();
+        const params = {
+            task: values.task ?? positionals.join(" ") ?? "", evaluator, filter: filterFrom(values),
+            ...(values["langfuse-dataset"] ? { langfuseDataset: values["langfuse-dataset"] } : {}), ...(values.dataset ? { datasetPath: values.dataset } : {}),
+            ...(values["sample-size"] ? { sampleSize: Number(values["sample-size"]) } : {}), ...(values.finalists ? { finalists: Number(values.finalists) } : {}),
+            ...(values.repeats ? { repeats: Number(values.repeats) } : {}),
+        };
+        if (!values.execute) {
+            const planned = await service.plan(host, params);
+            const out = { status: "plan_only", candidates: planned.ranked.candidates.length, datasetSize: planned.datasetSize, estimate: planned.estimate, evaluator: planned.plugin?.name ?? "pi", note: "Add --execute to run it." };
+            return void process.stdout.write(`${JSON.stringify(out, null, 2)}\n`);
+        }
+        const result = await service.run(host, params);
+        process.stdout.write(values.json ? `${JSON.stringify(result.report, null, 2)}\n` : `${result.markdown}\nReport: ${result.files.markdown}\n`);
+        return;
+    }
+    if (command === "scout") {
+        const task = positionals.join(" ").trim();
+        if (!task)
+            throw new Error("Describe the task: openmerit scout \"<what the model does>\"");
+        const { createAgentSession, DefaultResourceLoader, SessionManager, getAgentDir } = await import("@earendil-works/pi-coding-agent");
+        const { createScoutExtension } = await import("./pi/extension.js");
+        const { runtime, scout, host } = await hostFor(values);
+        if (!scout)
+            throw new Error("No model provider is connected for Scout. Set an API key or run `openmerit login`.");
+        const cwd = host.cwd;
+        const resourceLoader = new DefaultResourceLoader({
+            cwd, agentDir: getAgentDir(), noSkills: true, noPromptTemplates: true,
+            extensionFactories: [{ name: "openmerit-scout", factory: createScoutExtension({ allowSpend: values.execute, log, scoutModel: host.scoutModel }) }],
+        });
+        await resourceLoader.reload();
+        const { session } = await createAgentSession({
+            cwd, model: scout, thinkingLevel: "medium", modelRuntime: runtime, resourceLoader, sessionManager: SessionManager.inMemory(cwd),
+            tools: ["read", "grep", "find", "ls", "scout_connections", "scout_candidates", "scout_dataset", "scout_save_dataset", "scout_run"],
+        });
+        log(`Scout on ${host.scoutModel} in ${cwd}${values.execute ? "" : " (plan only: add --execute to spend)"}`);
+        session.subscribe((event) => {
+            if (event.type === "tool_execution_start")
+                log(`→ ${event.toolName} ${JSON.stringify(event.args ?? {}).slice(0, 300)}`);
+        });
+        const hints = [
+            values["langfuse-dataset"] && `Use the Langfuse dataset "${values["langfuse-dataset"]}".`,
+            values.dataset && `Use the local dataset ${values.dataset}.`,
+            values.evaluator && `Use the evaluator module ${values.evaluator}.`,
+            (values.providers || values.tools || values["max-price"] || values.exclude) && `Candidate filter: ${JSON.stringify(filterFrom(values))}.`,
+            values["sample-size"] && `sampleSize ${values["sample-size"]}.`, values.finalists && `finalists ${values.finalists}.`, values.repeats && `repeats ${values.repeats}.`,
+        ].filter(Boolean).join(" ");
+        try {
+            await session.prompt(`Recommend the model for this task.\n\nTask: ${task}\n\n${hints}`.trim());
+            const answer = session.getLastAssistantText() ?? "";
+            process.stdout.write(values.json ? `${JSON.stringify({ task, scoutModel: host.scoutModel, answer }, null, 2)}\n` : `${answer}\n`);
+        }
+        finally {
+            session.dispose();
+        }
+        return;
+    }
+    throw new Error(`Unknown command "${command}". Run \`openmerit help\`.`);
+}
+main().catch((error) => {
+    process.stderr.write(`openmerit: ${error instanceof Error ? error.message : String(error)}\n`);
+    process.exitCode = 1;
+});

package/dist/connectors/langfuse/client.d.ts

@@ -0,0 +1,42 @@
+export interface LangfuseConnection {
+    /** Langfuse origin, optionally with a path prefix for a reverse proxy, e.g. https://langfuse.internal/langfuse. */
+    baseUrl: string;
+    publicKey: string;
+    secretKey: string;
+}
+export interface TransportOptions {
+    /** Injected for tests; defaults to global fetch. */
+    fetch?: typeof fetch;
+    /** Per-request timeout. Default 30 seconds. */
+    timeoutMs?: number;
+    /** Retries after 429 (honouring Retry-After) and 502/503/504. Default 5. */
+    retries?: number;
+    /** Longest single wait for a retry. Default 120 seconds. */
+    maxRetryDelayMs?: number;
+    /** Injected for tests; defaults to a real timer. */
+    sleep?: (ms: number) => Promise<void>;
+}
+export declare class LangfuseError extends Error {
+    readonly status: number | null;
+    constructor(message: string, status?: number | null);
+}
+/**
+ * Reads LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY and LANGFUSE_BASE_URL (or LANGFUSE_HOST, the
+ * Langfuse SDK's name). Without a base URL it uses Langfuse Cloud EU, like the Langfuse SDKs.
+ */
+export declare function connectionFromEnv(env?: Record<string, string | undefined>): LangfuseConnection;
+export declare function normalizeBaseUrl(value: string): string;
+export type Query = Record<string, string | number | undefined | null>;
+export interface Transport {
+    readonly baseUrl: string;
+    request<T>(method: "GET" | "POST", path: string, options?: {
+        query?: Query;
+        body?: unknown;
+        headers?: Record<string, string>;
+        allowNotFound?: boolean;
+    }): Promise<{
+        status: number;
+        data: T | null;
+    }>;
+}
+export declare function createTransport(connection: LangfuseConnection, options?: TransportOptions): Transport;