llm-orchestrator 1.2.7 → 1.3.0

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "llm-orchestrator",
   "description": "Write /task once — it plans the work, shards it across parallel subagents, gates every phase and verifies before claiming done. Claude Code, Codex, OpenCode, Kilo.",
-  "version": "1.2.7",
+  "version": "1.3.0",
   "author": {
     "name": "Bogdan-Gabriel Torcescu",
     "url": "https://www.linkedin.com/in/bogdantorcescu/"

package/README.md

@@ -365,8 +365,8 @@ more instructions:
   it, runs opened without a PlanShard count, runs that planned several shards but started no
   subagents, and runs still open.
 - Trivial is the narrow exception — a one-line change such as a typo or a version bump. A run
-  declared trivial that then edits a second file, touches tests, or keeps working past eight calls
-  (an investigation is not trivial) gets one reminder to reopen it as
+  declared trivial that then edits a second file, touches tests, queries a live source twice, or keeps
+  working past eight calls (an investigation is not trivial) gets one reminder to reopen it as
   a typed run, and counts as `trivial_overreach` in the audit. Trivial runs close at the end of the
   turn (`Stop` hook), so single-prompt sessions still reach the history. Typed runs stay open across
   turns and close when the session ends (`SessionEnd`), so an unclosed run is never lost.
@@ -429,7 +429,7 @@ This resolves (task type, phase, role, risk, complexity, context size, harness)
 
 ## Cost-aware routing (built in)
 
-Routing is data, not prose: `registries/routing-matrix.json` encodes the W/S/X/F tiers, the T0–T5 thinking levels with their per-provider controls (`effort` on Claude, `reasoning_effort` on Codex, budget-only on Haiku), the tier × thinking resolution table, default routing, the per-task-type flows (phase → pair → roles → gate), 21 risk-floor areas with implementation and independent-review floors, agent defaults, both escalation ladders, fan-out minimums (2 MODERATE / 3 COMPLEX / 4 CRITICAL), the target tier distribution and the quota degradation ladder. `models/top-models.json` carries the curated top 20 models — ten ladder incumbents and ten measured candidates — with their supported thinking levels, prices, measured $/task per effort and thinking-cost indices (Artificial Analysis snapshot, copied verbatim, never invented). `lib/router.mjs` resolves a dispatch; `bin/route.mjs` exposes it:
+Routing is data, not prose: `registries/routing-matrix.json` encodes the W/S/X/F tiers, the T0–T5 thinking levels with their per-provider controls (`effort` on Claude, `reasoning_effort` on Codex, budget-only on Haiku), the tier × thinking resolution table, default routing, the per-task-type flows (phase → pair → roles → gate), 21 risk-floor areas with implementation and independent-review floors, agent defaults, both escalation ladders, fan-out minimums (2 MODERATE / 3 COMPLEX / 4 CRITICAL), the target tier distribution and the quota degradation ladder. `models/top-models.json` carries the curated shortlist — thirteen ladder incumbents (three of them fallbacks for their successors) and ten measured candidates — with their supported thinking levels, prices, measured $/task per effort and thinking-cost indices (Artificial Analysis snapshot, copied verbatim, never invented). `lib/router.mjs` resolves a dispatch; `bin/route.mjs` exposes it:
 
 ```sh
 node bin/llm-orchestrator.mjs route --list                                   # vocab: task types, phases, roles, areas
@@ -448,20 +448,22 @@ Resolution order: default routing → agent default → task-flow phase → comp
 
 The CLI resolves one dispatch. Inside a flow, the same resolution runs **for every PlanShard, at dispatch time, against the live inventory** — `lib/dispatch-contract.mjs` exposes `buildShardRouting(shard, options)` for one shard, `buildShardContracts(flow, { inventory, harness, includeCandidates })` for all of them (returning the flow ledger: tier histogram against the target distribution, mean `$/task`, blocked shards, warnings), and `rerouteRemaining(flow, inventory)` when a model-not-found, a rejected effort or a quota change invalidates the inventory mid-flow — remaining shards only, so the ledger stays true. Each shard carries a `routing` block with exactly these thirteen fields: `pair`, `tier`, `thinking_level`, `model_requested`, `effort_requested`, `model_effective`, `effort_effective`, `review_floor`, `independent_review`, `selection_reason`, `inventory_revision`, `price_source`, `est_usd_per_task`. An inventory that exposes nothing eligible for the resolved tier yields `blocked: "no eligible model"` — the floor is never lowered to fit what happens to be available.
 
-### Top 20 models and thinking levels (snapshot 2026-09-22)
+### Shortlisted models and thinking levels (snapshot 2026-09-23)
 
 Incumbents hold a seat on a provider ladder. Candidates are measured but unseated: their tier is a `tier_bands` placement from the best measured Artificial Analysis score (W ≤ 37, S 38–44, X 45–50, F ≥ 51), and they rank only with `--include-candidates` or an inventory that exposes them.
 
 | Model | Provider | Ladder | Admission | Tier | Thinking levels | $ in / out per MTok |
 |---|---|---|---|---|---|---|
-| GPT-5.6 Luna | openai | codex | incumbent | W | low, medium, high, xhigh, max (reasoning_effort) | $0.2 / $1.2 |
+| GPT-6 Luna | openai | codex | incumbent | W | low, medium, high, xhigh, max (reasoning_effort) | $0.1 / $0.5 |
+| GPT-5.6 Luna | openai | codex | incumbent (fallback for GPT-6 Luna) | W | low, medium, high, xhigh, max (reasoning_effort) | $0.2 / $1.2 |
 | GPT-5.6 Terra | openai | codex | incumbent | S | low, medium, high, xhigh, max (reasoning_effort) | $2 / $12 |
-| GPT-5.6 Sol | openai | codex | incumbent | X | low, medium, high, xhigh, max (reasoning_effort) | $4 / $20 |
+| GPT-6 Sol | openai | codex | incumbent | X | low, medium, high, xhigh, max (reasoning_effort) | $2 / $10 |
+| GPT-5.6 Sol | openai | codex | incumbent (fallback for GPT-6 Sol) | X | low, medium, high, xhigh, max (reasoning_effort) | $4 / $20 |
 | GPT-6 Astra | openai | codex | incumbent | F | low, medium, high, xhigh, max (reasoning_effort) | $10 / $50 |
 | Claude Haiku 4.5 | anthropic | claude | incumbent | W | disabled, enabled (budget_tokens) | $1 / $5 |
 | Claude Sonnet 5 | anthropic | claude | incumbent | S | low, medium, high, xhigh, max (effort) | $2 / $10 |
 | Claude Opus 5 | anthropic | claude | incumbent (fallback for Opus 5.5) | X | low, medium, high, xhigh, max (effort) | $5 / $25 |
-| Claude Opus 5.5 | anthropic | claude | incumbent | X | max measured; low–xhigh supported, unmeasured (effort) | $4 / $20 |
+| Claude Opus 5.5 | anthropic | claude | incumbent | X | low, medium, high, xhigh, max (effort) | $4 / $20 |
 | Claude Fable 5 | anthropic | claude | incumbent | F | low, medium, high, xhigh, max (effort) | $10 / $50 |
 | Claude Fable 5.1 | anthropic | claude | incumbent | F | low, medium, high, xhigh, max (effort) | $10 / $50 |
 | Grok 4.7 | xai | — | incumbent | — (unrated) | low, medium, high, xhigh (reasoning_effort) | $2 / $6 |

package/lib/flow-gate.mjs

@@ -52,6 +52,11 @@ export function overreachWorkNudgeFor(cli = 'llm-orchestrator', calls = 9) {
   return `This task was declared trivial, but it has now made ${calls} work calls — trivial is a one-line change, not an investigation. Reopen it as a typed run (\`${cli} run start --type <TYPE> --shards <n>\`) so it is classified, planned and verified like one.`;
 }
 
+/** A run declared trivial that is querying live sources. */
+export function overreachLiveNudgeFor(cli = 'llm-orchestrator', live = 2) {
+  return `This task was declared trivial, but it has now queried live sources ${live} times — an investigation over production evidence is not trivial. Reopen it as a typed run (\`${cli} run start --type INCIDENT|INVESTIGATION --shards <n>\`), where independent live sources are shards for collectors.`;
+}
+
 // Edit-shaped tools, per harness. Only a short hash of each path is kept.
 const EDIT_TOOLS = new Set(['edit', 'write', 'multiedit', 'notebookedit', 'str_replace_based_edit_tool', 'apply_patch', 'patch']);
 const TEST_PATH = /(^|\/)(test|tests|__tests__|spec|specs)\/|[._-](test|spec)\.[a-z0-9]+$|(^|\/)test_[^/]+\.py$|_spec\.rb$/i;
@@ -122,6 +127,8 @@ export const INLINE_REASON = /^stateful:\s*\S/;
 
 // A: a trivial run is a one-line change, not an investigation.
 export const TRIVIAL_WORK_LIMIT = 8;
+// Querying a live source more than once is an investigation, however few calls it takes.
+export const TRIVIAL_LIVE_LIMIT = 2;
 
 // C: sources whose reads cost real time and live remotely — the reads that pay to fan out.
 const LIVE_SOURCE = /(^|[\s;&|(])(rtk\s+)?(ssh|scp|psql|mysql|mongosh|mongo|redis-cli|kubectl|docker\s+(exec|logs)|aws|gcloud|az|curl|wget|http|httpie)(\s|$)/;
@@ -428,20 +435,24 @@ export function decide(previous, event, now) {
   if (session.run) {
     const current = session.run;
     current.main_work_calls = (current.main_work_calls ?? 0) + 1;
+    // Live reads count on every run, trivial ones included — they are what fan-out is for.
+    if (event.live) current.live_calls = (current.live_calls ?? 0) + 1;
     if (current.trivial) {
       current.edited_files = [...new Set([...(current.edited_files ?? []), ...event.edits.map((entry) => entry.hash)])];
       current.touched_tests = Boolean(current.touched_tests) || event.edits.some((entry) => entry.test);
       const grewByEdits = current.edited_files.length >= 2 || current.touched_tests;
       const grewByWork = current.main_work_calls > TRIVIAL_WORK_LIMIT;
-      if (!current.overreach_nudged && (grewByEdits || grewByWork)) {
+      const grewByLive = (current.live_calls ?? 0) >= TRIVIAL_LIVE_LIMIT;
+      if (!current.overreach_nudged && (grewByEdits || grewByWork || grewByLive)) {
         current.overreach_nudged = true;
         output = grewByEdits
           ? { additionalContext: overreachNudgeFor(undefined, current.edited_files.length), kind: 'overreach', files: current.edited_files.length }
-          : { additionalContext: overreachWorkNudgeFor(undefined, current.main_work_calls), kind: 'overreach', calls: current.main_work_calls };
+          : grewByLive
+            ? { additionalContext: overreachLiveNudgeFor(undefined, current.live_calls), kind: 'overreach', live: current.live_calls }
+            : { additionalContext: overreachWorkNudgeFor(undefined, current.main_work_calls), kind: 'overreach', calls: current.main_work_calls };
       }
       return { session, output, history };
     }
-    if (event.live) current.live_calls = (current.live_calls ?? 0) + 1;
     const evidence = EVIDENCE_TYPES.has(current.task_type);
     const step = (evidence ? EVIDENCE_CALLS_PER_SHARD : DISPATCH_CALLS_PER_SHARD) * (current.planned_shards ?? 0);
     const nudges = current.dispatch_nudges ?? (current.dispatch_nudged ? 1 : 0);
@@ -593,7 +604,8 @@ export async function handleHook({ payload, project, now = Date.now(), cli = 'll
   // decide() speaks in the default CLI spelling; the hook swaps in the runnable path.
   const text = result.output.kind === 'dispatch' ? dispatchNudgeFor(cli, result.output.planned, result.output.taskType)
     : result.output.kind === 'dispatch_followup' ? dispatchFollowupFor(cli, result.output.planned, result.output.taskType)
-    : result.output.kind === 'overreach' ? (result.output.calls ? overreachWorkNudgeFor(cli, result.output.calls) : overreachNudgeFor(cli, result.output.files))
+    : result.output.kind === 'overreach' ? (result.output.live ? overreachLiveNudgeFor(cli, result.output.live)
+      : result.output.calls ? overreachWorkNudgeFor(cli, result.output.calls) : overreachNudgeFor(cli, result.output.files))
       : nudgeFor(cli);
   return { hookSpecificOutput: { hookEventName: 'PreToolUse', additionalContext: text } };
 }

package/models/model-thinking-data.json

@@ -1,7 +1,7 @@
 {
   "_attribution": "llm-orchestrator · created by Bogdan-Gabriel Torcescu · https://www.linkedin.com/in/bogdantorcescu/ · keep this credit when copying or deriving",
   "schema_version": 1,
-  "observed_at": "2026-09-22",
+  "observed_at": "2026-09-23",
   "benchmark": {
     "name": "Artificial Analysis Intelligence Index",
     "version": "v4.3.2",
@@ -54,6 +54,42 @@
         ]
       ]
     },
+    {
+      "key": "gpt-6-luna",
+      "display_name": "GPT-6 Luna",
+      "provider": "OpenAI",
+      "api_id": "gpt-6-luna",
+      "input_usd_per_mtok": 0.1,
+      "output_usd_per_mtok": 0.5,
+      "source": "https://artificialanalysis.ai/models/gpt-6-luna",
+      "measurements": [
+        [
+          "low",
+          21,
+          0.0045
+        ],
+        [
+          "medium",
+          29,
+          0.02
+        ],
+        [
+          "high",
+          32,
+          0.03
+        ],
+        [
+          "xhigh",
+          34,
+          0.04
+        ],
+        [
+          "max",
+          37,
+          0.07
+        ]
+      ]
+    },
     {
       "key": "gpt-5-6-terra",
       "display_name": "GPT-5.6 Terra",
@@ -126,6 +162,42 @@
         ]
       ]
     },
+    {
+      "key": "gpt-6-sol",
+      "display_name": "GPT-6 Sol",
+      "provider": "OpenAI",
+      "api_id": "gpt-6-sol",
+      "input_usd_per_mtok": 2,
+      "output_usd_per_mtok": 10,
+      "source": "https://artificialanalysis.ai/models/gpt-6-sol",
+      "measurements": [
+        [
+          "low",
+          34,
+          0.13
+        ],
+        [
+          "medium",
+          40,
+          0.25
+        ],
+        [
+          "high",
+          43,
+          0.37
+        ],
+        [
+          "xhigh",
+          44,
+          0.53
+        ],
+        [
+          "max",
+          48,
+          1.06
+        ]
+      ]
+    },
     {
       "key": "gpt-6-astra",
       "display_name": "GPT-6 Astra",
@@ -242,8 +314,28 @@
       "input_usd_per_mtok": 4,
       "output_usd_per_mtok": 20,
       "source": "https://artificialanalysis.ai/models/claude-opus-5-5",
-      "measurement_note": "Only the max-effort configuration is published (Adaptive Reasoning, Max Effort, Default Fallback); low–xhigh are not measured. The score includes the benchmark default fallback.",
+      "measurement_note": "All displayed scores include the benchmark default fallback and are not standalone scores.",
       "measurements": [
+        [
+          "low",
+          42,
+          0.55
+        ],
+        [
+          "medium",
+          51,
+          1.34
+        ],
+        [
+          "high",
+          54,
+          1.82
+        ],
+        [
+          "xhigh",
+          56,
+          3.46
+        ],
         [
           "max",
           58,

package/models/model-thinking-matrix.md

@@ -1,7 +1,7 @@
 <!-- llm-orchestrator · created by Bogdan-Gabriel Torcescu · https://www.linkedin.com/in/bogdantorcescu/ · keep this credit when copying or deriving -->
 # Model thinking matrix — AA v4.3.2 snapshot
 
-Observed: 2026-09-22. Scores and weighted USD/task costs come from the Artificial Analysis Intelligence Index v4.3.2 snapshot, not live account tariffs or a guarantee of repository outcomes. [Overall leaderboard](https://artificialanalysis.ai/leaderboards/models).
+Observed: 2026-09-23. Scores and weighted USD/task costs come from the Artificial Analysis Intelligence Index v4.3.2 snapshot, not live account tariffs or a guarantee of repository outcomes. [Overall leaderboard](https://artificialanalysis.ai/leaderboards/models).
 
 This report compares measurements. It cannot automatically select a model, establish account availability, or replace security, payment, migration, concurrency, or compatibility review floors.
 
@@ -18,6 +18,11 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | gpt-5-6-luna | GPT-5.6 Luna | high | 32 | $0.04 | 8 | 6 | +$0.02 (2.00×); +7 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-luna | GPT-5.6 Luna | xhigh | 35 | $0.09 | 18 | 6 | +$0.05 (2.25×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-luna | GPT-5.6 Luna | max | 37 | $0.18 | 36 | 6 | +$0.09 (2.00×); +2 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-luna | GPT-6 Luna | low | 21 | $0.00 | 1 | 3 | — | First measured effort for this model; no internal effort comparison. |
+| gpt-6-luna | GPT-6 Luna | medium | 29 | $0.02 | 4 | 3 | +$0.02 (4.44×); +8 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-luna | GPT-6 Luna | high | 32 | $0.03 | 6 | 3 | +$0.01 (1.50×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-luna | GPT-6 Luna | xhigh | 34 | $0.04 | 8 | 3 | +$0.01 (1.33×); +2 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-luna | GPT-6 Luna | max | 37 | $0.07 | 14 | 3 | +$0.03 (1.75×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-terra | GPT-5.6 Terra | low | 27 | $0.14 | 28 | 56 | — | First measured effort for this model; no internal effort comparison. |
 | gpt-5-6-terra | GPT-5.6 Terra | medium | 30 | $0.18 | 36 | 56 | +$0.04 (1.29×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-terra | GPT-5.6 Terra | high | 34 | $0.34 | 68 | 56 | +$0.16 (1.89×); +4 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
@@ -28,6 +33,11 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | gpt-5-6-sol | GPT-5.6 Sol | high | 42 | $0.81 | 162 | 100 | +$0.31 (1.62×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-sol | GPT-5.6 Sol | xhigh | 44 | $1.18 | 236 | 100 | +$0.37 (1.46×); +2 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-5-6-sol | GPT-5.6 Sol | max | 47 | $1.99 | 398 | 100 | +$0.81 (1.69×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-sol | GPT-6 Sol | low | 34 | $0.13 | 26 | 50 | — | First measured effort for this model; no internal effort comparison. |
+| gpt-6-sol | GPT-6 Sol | medium | 40 | $0.25 | 50 | 50 | +$0.12 (1.92×); +6 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-sol | GPT-6 Sol | high | 43 | $0.37 | 74 | 50 | +$0.12 (1.48×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-sol | GPT-6 Sol | xhigh | 44 | $0.53 | 106 | 50 | +$0.16 (1.43×); +1 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
+| gpt-6-sol | GPT-6 Sol | max | 48 | $1.06 | 212 | 50 | +$0.53 (2.00×); +4 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-6-astra | GPT-6 Astra | low | 46 | $0.82 | 164 | 250 | — | First measured effort for this model; no internal effort comparison. |
 | gpt-6-astra | GPT-6 Astra | medium | 50 | $1.54 | 308 | 250 | +$0.72 (1.88×); +4 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | gpt-6-astra | GPT-6 Astra | high | 51 | $1.73 | 346 | 250 | +$0.19 (1.12×); +1 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
@@ -43,7 +53,11 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | claude-opus-5 | Claude Opus 5 | high | 48 | $3.61 | 722 | 125 | +$1.42 (1.65×); +3 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | claude-opus-5 | Claude Opus 5 | xhigh | 50 | $4.88 | 976 | 125 | +$1.27 (1.35×); +2 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
 | claude-opus-5 | Claude Opus 5 | max | 51 | $5.86 | 1172 | 125 | +$0.98 (1.20×); +1 score | Compare only with adjacent measured effort; higher effort raises observed cost. |
-| claude-opus-5-5 | Claude Opus 5.5 | max | 58 | $5.98 | 1196 | 100 | — | Segregate from standalone cross-model comparisons. |
+| claude-opus-5-5 | Claude Opus 5.5 | low | 42 | $0.55 | 110 | 100 | — | Segregate from standalone cross-model comparisons. |
+| claude-opus-5-5 | Claude Opus 5.5 | medium | 51 | $1.34 | 268 | 100 | +$0.79 (2.44×); +9 score | Segregate from standalone cross-model comparisons. |
+| claude-opus-5-5 | Claude Opus 5.5 | high | 54 | $1.82 | 364 | 100 | +$0.48 (1.36×); +3 score | Segregate from standalone cross-model comparisons. |
+| claude-opus-5-5 | Claude Opus 5.5 | xhigh | 56 | $3.46 | 692 | 100 | +$1.64 (1.90×); +2 score | Segregate from standalone cross-model comparisons. |
+| claude-opus-5-5 | Claude Opus 5.5 | max | 58 | $5.98 | 1196 | 100 | +$2.52 (1.73×); +2 score | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | low | 47 | $2.37 | 474 | 250 | — | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | medium | 49 | $2.98 | 596 | 250 | +$0.61 (1.26×); +2 score | Segregate from standalone cross-model comparisons. |
 | claude-fable-5-1 | Claude Fable 5.1 | high | 51 | $3.91 | 782 | 250 | +$0.93 (1.31×); +2 score | Segregate from standalone cross-model comparisons. |
@@ -84,8 +98,10 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | Model family | Input USD/MTok | Output USD/MTok | Price note |
 | --- | ---: | ---: | --- |
 | GPT-5.6 Luna | $0.200 | $1.200 | AA-observed price; not a live account tariff. |
+| GPT-6 Luna | $0.100 | $0.500 | AA-observed price; not a live account tariff. |
 | GPT-5.6 Terra | $2.000 | $12.000 | AA-observed price; not a live account tariff. |
 | GPT-5.6 Sol | $4.000 | $20.000 | AA-observed price; not a live account tariff. |
+| GPT-6 Sol | $2.000 | $10.000 | AA-observed price; not a live account tariff. |
 | GPT-6 Astra | $10.000 | $50.000 | AA-observed price; not a live account tariff. |
 | Claude Sonnet 5 | $2.000 | $10.000 | AA-observed price; not a live account tariff. |
 | Claude Opus 5 | $5.000 | $25.000 | AA-observed price; not a live account tariff. |
@@ -111,8 +127,10 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 | Model family | Canonical API ID | Availability evidence |
 | --- | --- | --- |
 | GPT-5.6 Luna | `gpt-5.6-luna` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
+| GPT-6 Luna | `gpt-6-luna` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | GPT-5.6 Terra | `gpt-5.6-terra` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | GPT-5.6 Sol | `gpt-5.6-sol` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
+| GPT-6 Sol | `gpt-6-sol` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | GPT-6 Astra | `gpt-6-astra` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Sonnet 5 | `claude-sonnet-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
 | Claude Opus 5 | `claude-opus-5` | Canonical ID recorded in this dataset; exposure in a specific harness remains unverified. |
@@ -136,12 +154,14 @@ Benchmark cost index: `100 × observed benchmark cost / $0.50`; GPT-5.6 Sol medi
 ## Source and interpretation limits
 
 - [GPT-5.6 Luna](https://artificialanalysis.ai/models/gpt-5-6-luna)
+- [GPT-6 Luna](https://artificialanalysis.ai/models/gpt-6-luna)
 - [GPT-5.6 Terra](https://artificialanalysis.ai/models/gpt-5-6-terra)
 - [GPT-5.6 Sol](https://artificialanalysis.ai/models/gpt-5-6-sol)
+- [GPT-6 Sol](https://artificialanalysis.ai/models/gpt-6-sol)
 - [GPT-6 Astra](https://artificialanalysis.ai/models/gpt-6-astra)
 - [Claude Sonnet 5](https://artificialanalysis.ai/models/claude-sonnet-5)
 - [Claude Opus 5](https://artificialanalysis.ai/models/claude-opus-5)
-- [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5) — Only the max-effort configuration is published (Adaptive Reasoning, Max Effort, Default Fallback); low–xhigh are not measured. The score includes the benchmark default fallback.
+- [Claude Opus 5.5](https://artificialanalysis.ai/models/claude-opus-5-5) — All displayed scores include the benchmark default fallback and are not standalone scores.
 - [Claude Fable 5.1](https://artificialanalysis.ai/models/claude-fable-5-1) — All displayed scores include the benchmark default fallback and are not standalone scores.
 - [Claude Fable 5](https://artificialanalysis.ai/models/claude-fable-5) — AA v4.3.2 max result uses Opus 4.8 default fallback. This legacy catalog model is not in the current leaderboard filter; segregate it from standalone cross-model comparisons.
 - [Claude Haiku 4.5](https://artificialanalysis.ai/models/claude-4-5-haiku) — Its API thinking budget is not an effort enum. The disabled score 15 is incomplete and intentionally excluded.

package/models/top-models.json

@@ -91,6 +91,7 @@
         "share_max": null
       },
       "notes": [
+        "Fallback for GPT-6 Luna on its Codex seat: dispatched when GPT-6 Luna is not exposed by the inventory or is rolled back.",
         "Default W-tier config is `xhigh`; `max` when it stalls.",
         "Escalate to S when the task needs capacity (judgment, cross-file scope, ambiguity), not merely more reasoning depth."
       ],
@@ -99,6 +100,115 @@
       ],
       "observed_at": "2026-09-22"
     },
+    {
+      "key": "gpt-6-luna",
+      "display_name": "GPT-6 Luna",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "W",
+      "eligible_tiers": [
+        "W"
+      ],
+      "supersedes": "gpt-5-6-luna",
+      "api_ids": {
+        "codex": "gpt-6-luna",
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1050000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "reasoning_effort",
+        "levels": [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        "always_on": false,
+        "default_for_tier": "xhigh",
+        "note": "Same W policy as GPT-5.6 Luna: exempt from the Codex `high` ceiling, thinking is cheap on a small model. The API also accepts `none`, which is not measured and never routed."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.1,
+        "output_usd_per_mtok": 0.5,
+        "cache_read_usd_per_mtok": 0.01,
+        "cache_write_multiplier": 1.25
+      },
+      "long_context_surcharge": {
+        "threshold_input_tokens": 272000,
+        "input_multiplier": 2,
+        "output_multiplier": 1.5,
+        "effective_input_usd_per_mtok": 0.2,
+        "effective_output_usd_per_mtok": 0.75,
+        "note": "Crossing the threshold reprices the entire request, not just the overflow."
+      },
+      "measured": {
+        "low": {
+          "score": 21,
+          "usd_per_task": 0.0045
+        },
+        "medium": {
+          "score": 29,
+          "usd_per_task": 0.02
+        },
+        "high": {
+          "score": 32,
+          "usd_per_task": 0.03
+        },
+        "xhigh": {
+          "score": 34,
+          "usd_per_task": 0.04
+        },
+        "max": {
+          "score": 37,
+          "usd_per_task": 0.07
+        }
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 0.9,
+          "medium": 4,
+          "high": 6,
+          "xhigh": 8,
+          "max": 14
+        },
+        "marginal_thinking": {
+          "low->medium": {
+            "delta_score": 8,
+            "cost_multiplier": 4.44
+          },
+          "medium->high": {
+            "delta_score": 3,
+            "cost_multiplier": 1.5
+          },
+          "high->xhigh": {
+            "delta_score": 2,
+            "cost_multiplier": 1.33
+          },
+          "xhigh->max": {
+            "delta_score": 3,
+            "cost_multiplier": 1.75
+          }
+        }
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "Successor to GPT-5.6 Luna on the Codex W seat: $0.10/$0.50 vs $0.20/$1.20 per MTok, and cheaper per completed task at every measured effort (max 37 @ $0.07 vs 37 @ $0.18). GPT-5.6 Luna stays routable as its fallback.",
+        "Default W-tier config is `xhigh`; `max` when it stalls. Escalate to S for capacity, not depth."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-6-luna",
+        "https://developers.openai.com/api/docs/models/gpt-6-luna"
+      ],
+      "observed_at": "2026-09-23"
+    },
     {
       "key": "gpt-5-6-terra",
       "display_name": "GPT-5.6 Terra",
@@ -239,6 +349,7 @@
         }
       },
       "notes": [
+        "Fallback for GPT-6 Sol on its Codex seat: dispatched when GPT-6 Sol is not exposed by the inventory or is rolled back.",
         "Sol never above `high`; `max` is dominated and never correct.",
         "X T4 on Codex = Sol `high` + an independent Sol `high` reviewer (different agent, no forked history)."
       ],
@@ -247,6 +358,119 @@
       ],
       "observed_at": "2026-09-22"
     },
+    {
+      "key": "gpt-6-sol",
+      "display_name": "GPT-6 Sol",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "X",
+      "eligible_tiers": [
+        "X"
+      ],
+      "supersedes": "gpt-5-6-sol",
+      "api_ids": {
+        "codex": "gpt-6-sol",
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1050000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "reasoning_effort",
+        "levels": [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
+        "always_on": false,
+        "default_for_tier": "high",
+        "note": "Same Codex policy as GPT-5.6 Sol: ceiling `high`; T4 is `high` plus an independent second Sol `high` reviewer with no shared history. The API also accepts `none`, which is not measured and never routed."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 10,
+        "cache_read_usd_per_mtok": 0.2,
+        "cache_write_multiplier": 1.25
+      },
+      "long_context_surcharge": {
+        "threshold_input_tokens": 272000,
+        "input_multiplier": 2,
+        "output_multiplier": 1.5,
+        "effective_input_usd_per_mtok": 4,
+        "effective_output_usd_per_mtok": 15,
+        "note": "Crossing the threshold reprices the entire request, not just the overflow."
+      },
+      "measured": {
+        "low": {
+          "score": 34,
+          "usd_per_task": 0.13
+        },
+        "medium": {
+          "score": 40,
+          "usd_per_task": 0.25
+        },
+        "high": {
+          "score": 43,
+          "usd_per_task": 0.37
+        },
+        "xhigh": {
+          "score": 44,
+          "usd_per_task": 0.53
+        },
+        "max": {
+          "score": 48,
+          "usd_per_task": 1.06
+        }
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 26,
+          "medium": 50,
+          "high": 74,
+          "xhigh": 106,
+          "max": 212
+        },
+        "marginal_thinking": {
+          "low->medium": {
+            "delta_score": 6,
+            "cost_multiplier": 1.92
+          },
+          "medium->high": {
+            "delta_score": 3,
+            "cost_multiplier": 1.48
+          },
+          "high->xhigh": {
+            "delta_score": 1,
+            "cost_multiplier": 1.43
+          },
+          "xhigh->max": {
+            "delta_score": 4,
+            "cost_multiplier": 2.0
+          }
+        }
+      },
+      "caps": {
+        "max_effort": "high",
+        "share_max": null,
+        "t4_substitution": {
+          "effort": "high",
+          "requires_independent_second_reviewer": true
+        }
+      },
+      "notes": [
+        "Successor to GPT-5.6 Sol on the Codex X seat at half the per-token price ($2/$10 vs $4/$20, permanent pricing). It is cheaper per completed task at every measured effort and scores equal or higher at each (high 43 @ $0.37 vs 42 @ $0.81); GPT-5.6 Sol stays routable as its fallback.",
+        "Its `medium` (40 @ $0.25) also dominates GPT-5.6 Terra `xhigh` (38 @ $0.63); the S seat stays Terra because an incumbent's tier is its ladder seat, not a score — see the routing note."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-6-sol",
+        "https://developers.openai.com/api/docs/models/gpt-6-sol"
+      ],
+      "observed_at": "2026-09-23"
+    },
     {
       "key": "gpt-6-astra",
       "display_name": "GPT-6 Astra",
@@ -506,7 +730,9 @@
       "ladder": "claude",
       "admission": "incumbent",
       "tier": "X",
-      "eligible_tiers": ["X"],
+      "eligible_tiers": [
+        "X"
+      ],
       "supersedes": "claude-opus-5",
       "api_ids": {
         "codex": null,
@@ -518,8 +744,13 @@
       "max_output_tokens": 128000,
       "thinking": {
         "control": "effort",
-        "levels": ["low", "medium", "high", "xhigh", "max"],
-        "unmeasured_levels": ["low", "medium", "high", "xhigh"],
+        "levels": [
+          "low",
+          "medium",
+          "high",
+          "xhigh",
+          "max"
+        ],
         "always_on": true,
         "default_for_tier": "high",
         "note": "Thinking is always on: omit `thinking` or send `adaptive`; `disabled` and `budget_tokens` return 400 at every effort. The API default effort is `medium` (one below Opus 5), so X always sets effort explicitly — `high` for T3, `xhigh` for T4. Forced `tool_choice` any/tool returns 400 — use `auto` + instruction, `strict: true`, or structured outputs. Thinking blocks are model-bound; keep transcripts append-only. Computer use only via `computer_toolset_20260801`."
@@ -533,14 +764,54 @@
       },
       "long_context_surcharge": null,
       "measured": {
-        "max": {"score": 58, "usd_per_task": 5.98}
+        "low": {
+          "score": 42,
+          "usd_per_task": 0.55
+        },
+        "medium": {
+          "score": 51,
+          "usd_per_task": 1.34
+        },
+        "high": {
+          "score": 54,
+          "usd_per_task": 1.82
+        },
+        "xhigh": {
+          "score": 56,
+          "usd_per_task": 3.46
+        },
+        "max": {
+          "score": 58,
+          "usd_per_task": 5.98
+        }
       },
-      "measurement_note": "Only `max` is published (Adaptive Reasoning, Max Effort, Default Fallback); the score includes the benchmark default fallback. low–xhigh are unmeasured, so a dispatch at them carries no $/task estimate until Artificial Analysis publishes them.",
+      "measurement_note": "All displayed scores include the benchmark default fallback and are not standalone scores.",
       "indices": {
         "thinking_cost_index": {
+          "low": 110,
+          "medium": 268,
+          "high": 364,
+          "xhigh": 692,
           "max": 1196
         },
-        "marginal_thinking": {}
+        "marginal_thinking": {
+          "low->medium": {
+            "delta_score": 9,
+            "cost_multiplier": 2.44
+          },
+          "medium->high": {
+            "delta_score": 3,
+            "cost_multiplier": 1.36
+          },
+          "high->xhigh": {
+            "delta_score": 2,
+            "cost_multiplier": 1.9
+          },
+          "xhigh->max": {
+            "delta_score": 2,
+            "cost_multiplier": 1.73
+          }
+        }
       },
       "caps": {
         "max_effort": "xhigh",
@@ -548,7 +819,7 @@
       },
       "notes": [
         "Successor to Claude Opus 5 on the Claude X seat. Every per-token price is lower (Anthropic list: input $4 vs $5, output $20 vs $25, cache read $0.20 vs $0.50, cache write $5 vs $6.25), with the same 1M context, 128K output and tokenizer, so the router ranks it ahead of Opus 5 at the same effort; Opus 5 stays routable as its fallback.",
-        "Only `max` is independently measured (Artificial Analysis 58 @ $5.98, above every other shortlisted config). Anthropic's launch table puts it above Opus 5 on every published benchmark (Terminal-Bench 4.0 66.4 vs 52.3, FrontierCode v1.1 54.4 vs 48.0, CursorBench 4.0 57.8 vs 46.6) and claims ~40% lower cost on typical workloads — vendor figures, kept as notes and never used as $/task estimates.",
+        "Measured at every effort (Artificial Analysis, with default fallback): high 54 @ $1.82 vs Opus 5 high 48 @ $3.61 — higher score at half the cost per task. Anthropic's launch table also puts it above Opus 5 on every published benchmark.",
         "Seat stays X: an incumbent tier is never re-derived from a score."
       ],
       "sources": [

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "llm-orchestrator",
-  "version": "1.2.7",
+  "version": "1.3.0",
   "description": "Write /task once — it plans the work, shards it across parallel subagents, gates every phase and verifies before claiming done. Claude Code, Codex, OpenCode, Kilo.",
   "type": "module",
   "engines": {

package/policies/routing.md

@@ -74,9 +74,9 @@ the incumbents below are role mappings, not guarantees of availability.
 
 | Tier | Role | Claude incumbent | Codex incumbent |
 | --- | --- | --- | --- |
-| **W** worker | local, mechanical, repetitive, well-defined: code search, classification, extraction, small edits, boilerplate, simple tests, consistency checks, scoped transforms | Haiku 4.5 (`claude-haiku-4-5`) | `gpt-5.6-luna` |
+| **W** worker | local, mechanical, repetitive, well-defined: code search, classification, extraction, small edits, boilerplate, simple tests, consistency checks, scoped transforms | Haiku 4.5 (`claude-haiku-4-5`) | `gpt-6-luna`, falling back to `gpt-5.6-luna` |
 | **S** standard | default software-engineering model: normal implementation, frontend/backend, moderate debugging, tests, reasonable multi-file refactors, codebase analysis, tool use | Sonnet 5 (`claude-sonnet-5`) | `gpt-5.6-terra` |
-| **X** senior | hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants | Opus 5.5 (`claude-opus-5-5`), falling back to Opus 5 (`claude-opus-5`) | `gpt-5.6-sol` |
+| **X** senior | hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants | Opus 5.5 (`claude-opus-5-5`), falling back to Opus 5 (`claude-opus-5`) | `gpt-6-sol`, falling back to `gpt-5.6-sol` |
 | **F** frontier | exceptional escalation: very ambiguous, long-horizon, cross-system, major architecture, very large codebase, planning under heavy constraints, or when X fails to produce a solid solution | Fable 5 (`claude-fable-5`) — default F. Fable 5.1 (`claude-fable-5-1`) is a hard-capped exception: **≤2% of all dispatches**, explicit request or documented F-T4 failure on Fable 5 only | GPT-6 Astra (`gpt-6-astra`) — a real single-agent frontier tier, no decomposition workaround needed |
 
 Escalation order within a provider: **W → S → X → F**.
@@ -85,9 +85,9 @@ Verified prices (Sep 2026, provider pricing pages), USD in/out per MTok:
 
 | Tier | Claude model | Claude $ in/out | Codex model | Codex $ in/out |
 | --- | --- | --- | --- | --- |
-| W | Haiku 4.5 (200K ctx) | $1 / $5 | `gpt-5.6-luna` (1.05M ctx) | $0.20 / $1.20 |
+| W | Haiku 4.5 (200K ctx) | $1 / $5 | `gpt-6-luna` (1.05M ctx); fallback `gpt-5.6-luna` $0.20 / $1.20 | $0.10 / $0.50 |
 | S | Sonnet 5 (1M ctx) | $2 / $10 | `gpt-5.6-terra` (1.05M ctx) | $2 / $12 |
-| X | Opus 5.5 (1M ctx); fallback Opus 5 $5 / $25 | $4 / $20 | `gpt-5.6-sol` (1.05M ctx) | $4 / $20 |
+| X | Opus 5.5 (1M ctx); fallback Opus 5 $5 / $25 | $4 / $20 | `gpt-6-sol` (1.05M ctx); fallback `gpt-5.6-sol` $4 / $20 | $2 / $10 |
 | F | Fable 5 (1M ctx) | $10 / $50 | `gpt-6-astra` (1.05M ctx) | $10 / $50 list — **but the cheapest F per completed task of any model here** |
 | F+ (≤2%) | Fable 5.1 (1M ctx) | $10 / $50 list — **effective cost significantly higher** (always-on thinking, longer turns, more output tokens per task) | — (Astra covers F) | — |
 

package/registries/routing-matrix.json

@@ -8,7 +8,8 @@
     "W": {
       "name": "worker",
       "responsibility": "Local, mechanical, repetitive, well-defined: code search, classification, extraction, small edits, boilerplate, simple tests, consistency checks, scoped transforms.",
-      "incumbents": {"claude": "claude-haiku-4-5", "codex": "gpt-5.6-luna"}
+      "incumbents": {"claude": "claude-haiku-4-5", "codex": "gpt-6-luna"},
+      "fallbacks": {"codex": ["gpt-5.6-luna"]}
     },
     "S": {
       "name": "standard",
@@ -18,8 +19,8 @@
     "X": {
       "name": "senior",
       "responsibility": "Hard debugging, architecture, concurrency, migrations, security, auth, payments, billing, backwards compatibility, critical code review, many invariants.",
-      "incumbents": {"claude": "claude-opus-5-5", "codex": "gpt-5.6-sol"},
-      "fallbacks": {"claude": ["claude-opus-5"]}
+      "incumbents": {"claude": "claude-opus-5-5", "codex": "gpt-6-sol"},
+      "fallbacks": {"claude": ["claude-opus-5"], "codex": ["gpt-5.6-sol"]}
     },
     "F": {
       "name": "frontier",
@@ -74,16 +75,16 @@
   "resolution": {
     "W T0-T1": {
       "claude": {"model": "claude-haiku-4-5", "effort": null, "thinking": "off"},
-      "codex": {"model": "gpt-5.6-luna", "effort": "low"}
+      "codex": {"model": "gpt-6-luna", "effort": "low"}
     },
     "W T2": {
       "claude": {"model": "claude-haiku-4-5", "effort": null, "thinking": "small budget"},
-      "codex": {"model": "gpt-5.6-luna", "effort": "medium"}
+      "codex": {"model": "gpt-6-luna", "effort": "medium"}
     },
     "W T3": {
       "claude": null,
       "claude_note": "No headroom at the Claude W tier — escalate to S.",
-      "codex": {"model": "gpt-5.6-luna", "effort": "xhigh", "alternate_effort": "max"}
+      "codex": {"model": "gpt-6-luna", "effort": "xhigh", "alternate_effort": "max"}
     },
     "S T1": {
       "claude": {"model": "claude-sonnet-5", "effort": "low"},
@@ -99,15 +100,15 @@
     },
     "X T2": {
       "claude": {"model": "claude-opus-5-5", "effort": "medium"},
-      "codex": {"model": "gpt-5.6-sol", "effort": "medium"}
+      "codex": {"model": "gpt-6-sol", "effort": "medium"}
     },
     "X T3": {
       "claude": {"model": "claude-opus-5-5", "effort": "high"},
-      "codex": {"model": "gpt-5.6-sol", "effort": "high"}
+      "codex": {"model": "gpt-6-sol", "effort": "high"}
     },
     "X T4": {
       "claude": {"model": "claude-opus-5-5", "effort": "xhigh"},
-      "codex": {"model": "gpt-5.6-sol", "effort": "high", "independent_second_reviewer": {"model": "gpt-5.6-sol", "effort": "high"}}
+      "codex": {"model": "gpt-6-sol", "effort": "high", "independent_second_reviewer": {"model": "gpt-6-sol", "effort": "high"}}
     },
     "F T3": {
       "claude": {"model": "claude-fable-5", "effort": "high"},
@@ -288,15 +289,15 @@
       {"model": "claude-fable-5-1", "effort": "high", "requires_explicit_flag": true, "note": "≤2% cap, explicit request or documented Fable 5 F-T4 failure only."}
     ],
     "codex": [
-      {"model": "gpt-5.6-luna", "effort": "low"},
-      {"model": "gpt-5.6-luna", "effort": "medium"},
-      {"model": "gpt-5.6-luna", "effort": "xhigh"},
-      {"model": "gpt-5.6-luna", "effort": "max"},
+      {"model": "gpt-6-luna", "effort": "low"},
+      {"model": "gpt-6-luna", "effort": "medium"},
+      {"model": "gpt-6-luna", "effort": "xhigh"},
+      {"model": "gpt-6-luna", "effort": "max"},
       {"model": "gpt-5.6-terra", "effort": "medium"},
       {"model": "gpt-5.6-terra", "effort": "high"},
-      {"model": "gpt-5.6-sol", "effort": "medium"},
-      {"model": "gpt-5.6-sol", "effort": "high"},
-      {"model": "gpt-5.6-sol", "effort": "high", "independent_second_reviewer": true},
+      {"model": "gpt-6-sol", "effort": "medium"},
+      {"model": "gpt-6-sol", "effort": "high"},
+      {"model": "gpt-6-sol", "effort": "high", "independent_second_reviewer": true},
       {"model": "gpt-6-astra", "effort": "high"},
       {"model": "gpt-6-astra", "effort": "xhigh"},
       {"model": "gpt-6-astra", "effort": "max", "note": "Essentially never."}