llm-orchestrator 1.0.0

package/models/top-models.json

@@ -0,0 +1,1299 @@
+{
+  "_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",
+  "purpose": "Curated top-20 LLM shortlist for agentic coding, with the facts the router needs: tier on a provider ladder or a measured score band, thinking controls, list prices, measured cost per completed task where a measurement exists, and the thinking-cost indices the router ranks by.",
+  "measurement_source": {
+    "dataset": "models/model-thinking-data.json",
+    "benchmark": "Artificial Analysis Intelligence Index",
+    "version": "v4.3.2",
+    "leaderboard_url": "https://artificialanalysis.ai/leaderboards/models",
+    "note": "Every `measured` entry is copied verbatim from models/model-thinking-data.json. Unmeasured efforts are null; nothing here is interpolated or estimated."
+  },
+  "harness_id_note": "`api_ids.codex` and `api_ids.claude` carry the canonical provider IDs. `api_ids.opencode` and `api_ids.kilo` are null on purpose: on OpenCode and Kilo the dispatchable `provider/model` identifier is resolved from the active runtime (policies/routing.md, 'Discovery before dispatch'). A convention is not evidence of exposure.",
+  "tier_note": "`tier` is the model's seat on its own provider ladder (W/S/X/F) for an incumbent, or its `tier_bands` placement from the best measured Artificial Analysis score for a candidate. Models with `tier: null` are available-but-unrated: they are never auto-promoted into a tier seat and never rank as an eligible model without a local measurement.",
+  "admission_note": "`admission: \"incumbent\"` means the model holds a seat on a harness provider ladder and may be dispatched anywhere its tier allows. `admission: \"candidate\"` means the model is measured but unseated: it is ranked only when the caller passes `--include-candidates` / `includeCandidates: true`, or when a runtime inventory marks it `exposed` or `verified`, and even then only for noncritical lanes. A candidate is never eligible for an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+  "ladder_note": "`ladder` names the dispatch ladder a model sits on: `codex` and `claude` are the two first-party harness ladders; `open` is a published open-weight lineage reachable through third-party or self-hosted endpoints; `third-party` is a hosted API outside both harness ladders whose weight availability is not verified here. `null` means unrated and unplaced. Weight availability is recorded at lab level and is not verified per checkpoint.",
+  "tier_bands": {
+    "basis": "Artificial Analysis Intelligence Index v4.3.2, best measured effort per model (models/model-thinking-data.json).",
+    "applies_to": "candidate",
+    "bands": {
+      "W": {"min": null, "max": 37},
+      "S": {"min": 38, "max": 44},
+      "X": {"min": 45, "max": 50},
+      "F": {"min": 51, "max": null}
+    },
+    "note": "Bands are calibrated on the incumbents' measured range: Luna tops W at 37; Terra spans 38–42 in S; Sol 33–47 and Opus 39–50 straddle into X, whose seats sit at 45–50; Astra 46–53 and Fable 5.1 47–53 cover F from 51 up. They place *candidates*, which sit on no provider ladder. An incumbent's `tier` is its ladder seat and is never re-derived from a score — which is why two incumbent points fall outside their own band: Opus 5 `max` (51, above its `xhigh` policy cap) and Fable 5 `max` (50, its only measured point and above the unmeasured `high` that is the actual F default)."
+  },
+  "indices_note": "`indices.thinking_cost_index` is 100 × measured $/task ÷ $0.50, the dataset's benchmark cost baseline (`index_baselines.benchmark_cost_usd`). GPT-5.6 Sol `medium` costs exactly $0.50/task, so it is index 100 by construction. `indices.marginal_thinking` holds, for each adjacent pair of fully measured efforts, the displayed-score delta and the cost multiplier of the step — the price of one more notch of thinking. Efforts with no published cost carry a null index and open no marginal pair; nothing is interpolated.",
+  "models": [
+    {
+      "key": "gpt-5-6-luna",
+      "display_name": "GPT-5.6 Luna",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "W",
+      "eligible_tiers": ["W"],
+      "api_ids": {
+        "codex": "gpt-5.6-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": "Luna is exempt from the Codex `high` ceiling: thinking is cheap on a small model and xhigh/max are both cheaper and higher-scoring than Terra high."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.2,
+        "output_usd_per_mtok": 1.2,
+        "cache_read_usd_per_mtok": 0.02,
+        "cache_write_multiplier": 1.25
+      },
+      "long_context_surcharge": {
+        "threshold_input_tokens": 272000,
+        "input_multiplier": 2,
+        "output_multiplier": 1.5,
+        "effective_input_usd_per_mtok": 0.4,
+        "effective_output_usd_per_mtok": 1.8,
+        "note": "Crossing the threshold reprices the entire request, not just the overflow."
+      },
+      "measured": {
+        "low": {"score": 21, "usd_per_task": 0.01},
+        "medium": {"score": 25, "usd_per_task": 0.02},
+        "high": {"score": 32, "usd_per_task": 0.04},
+        "xhigh": {"score": 35, "usd_per_task": 0.09},
+        "max": {"score": 37, "usd_per_task": 0.18}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 2,
+          "medium": 4,
+          "high": 8,
+          "xhigh": 18,
+          "max": 36
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 4, "cost_multiplier": 2},
+          "medium->high": {"delta_score": 7, "cost_multiplier": 2},
+          "high->xhigh": {"delta_score": 3, "cost_multiplier": 2.25},
+          "xhigh->max": {"delta_score": 2, "cost_multiplier": 2}
+        }
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "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."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-5-6-luna"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "gpt-5-6-terra",
+      "display_name": "GPT-5.6 Terra",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": "gpt-5.6-terra",
+        "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": "medium",
+        "note": "Policy ceiling is `high`. `xhigh` and `max` exist on the API but every measured point above `high` is dominated by a cheaper Sol config — escalate the model, never the dial."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 12,
+        "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": 18,
+        "note": "Crossing the threshold reprices the entire request, not just the overflow."
+      },
+      "measured": {
+        "low": {"score": 27, "usd_per_task": 0.14},
+        "medium": {"score": 30, "usd_per_task": 0.18},
+        "high": {"score": 34, "usd_per_task": 0.34},
+        "xhigh": {"score": 38, "usd_per_task": 0.63},
+        "max": {"score": 42, "usd_per_task": 1.4}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 28,
+          "medium": 36,
+          "high": 68,
+          "xhigh": 126,
+          "max": 280
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 3, "cost_multiplier": 1.29},
+          "medium->high": {"delta_score": 4, "cost_multiplier": 1.89},
+          "high->xhigh": {"delta_score": 4, "cost_multiplier": 1.85},
+          "xhigh->max": {"delta_score": 4, "cost_multiplier": 2.22}
+        }
+      },
+      "caps": {
+        "max_effort": "high",
+        "share_max": null
+      },
+      "notes": [
+        "Terra never above `high` — firm rule in policies/routing.md.",
+        "Terra is the squeezed tier: every measured Terra config is dominated by a Luna or Sol config. Score is not capacity, so the S tier is not indicted as a whole, but do not raise Terra's dial to buy quality."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-5-6-terra"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "gpt-5-6-sol",
+      "display_name": "GPT-5.6 Sol",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": "gpt-5.6-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": "Policy ceiling is `high`. T4 on the gpt-5.6 family is expressed as `high` plus an independent second Sol `high` reviewer with no shared history — a risk-floor rule, not a capability workaround."
+      },
+      "price": {
+        "input_usd_per_mtok": 4,
+        "output_usd_per_mtok": 20,
+        "cache_read_usd_per_mtok": 0.4,
+        "cache_write_multiplier": 1.25
+      },
+      "long_context_surcharge": {
+        "threshold_input_tokens": 272000,
+        "input_multiplier": 2,
+        "output_multiplier": 1.5,
+        "effective_input_usd_per_mtok": 8,
+        "effective_output_usd_per_mtok": 30,
+        "note": "Crossing the threshold reprices the entire request, not just the overflow."
+      },
+      "measured": {
+        "low": {"score": 33, "usd_per_task": 0.26},
+        "medium": {"score": 39, "usd_per_task": 0.5},
+        "high": {"score": 42, "usd_per_task": 0.81},
+        "xhigh": {"score": 44, "usd_per_task": 1.18},
+        "max": {"score": 47, "usd_per_task": 1.99}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 52,
+          "medium": 100,
+          "high": 162,
+          "xhigh": 236,
+          "max": 398
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 6, "cost_multiplier": 1.92},
+          "medium->high": {"delta_score": 3, "cost_multiplier": 1.62},
+          "high->xhigh": {"delta_score": 2, "cost_multiplier": 1.46},
+          "xhigh->max": {"delta_score": 3, "cost_multiplier": 1.69}
+        }
+      },
+      "caps": {
+        "max_effort": "high",
+        "share_max": null,
+        "t4_substitution": {
+          "effort": "high",
+          "requires_independent_second_reviewer": true
+        }
+      },
+      "notes": [
+        "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)."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-5-6-sol"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "gpt-6-astra",
+      "display_name": "GPT-6 Astra",
+      "provider": "openai",
+      "ladder": "codex",
+      "admission": "incumbent",
+      "tier": "F",
+      "eligible_tiers": ["F"],
+      "api_ids": {
+        "codex": "gpt-6-astra",
+        "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": "Exempt from the Codex `high` ceiling because F needs a T4/T5 expression. `none` exists in prerelease evaluations only and is not publicly callable — never dispatch it."
+      },
+      "price": {
+        "input_usd_per_mtok": 10,
+        "output_usd_per_mtok": 50,
+        "cache_read_usd_per_mtok": 1,
+        "cache_write_multiplier": 1.25
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 46, "usd_per_task": 0.82},
+        "medium": {"score": 50, "usd_per_task": 1.54},
+        "high": {"score": 51, "usd_per_task": 1.73},
+        "xhigh": {"score": 52, "usd_per_task": 2.31},
+        "max": {"score": 53, "usd_per_task": 3.26}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 164,
+          "medium": 308,
+          "high": 346,
+          "xhigh": 462,
+          "max": 652
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 4, "cost_multiplier": 1.88},
+          "medium->high": {"delta_score": 1, "cost_multiplier": 1.12},
+          "high->xhigh": {"delta_score": 1, "cost_multiplier": 1.34},
+          "xhigh->max": {"delta_score": 1, "cost_multiplier": 1.41}
+        }
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "A real single-agent frontier tier: no decomposition workaround needed.",
+        "Astra `low`/`medium` are priced like a strong X tier but are not a substitute for Sol on X work.",
+        "Long-context repricing is not documented for this model; assume it may apply and prefer a known-flat-priced option on very large working sets."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gpt-6-astra"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "claude-4-5-haiku",
+      "display_name": "Claude Haiku 4.5",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "W",
+      "eligible_tiers": ["W"],
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-haiku-4-5",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 200000,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "budget_tokens",
+        "levels": ["disabled", "enabled"],
+        "always_on": false,
+        "default_for_tier": "disabled",
+        "note": "No `effort` parameter at all — only `thinking: enabled` + `budget_tokens`. T0/T1 = thinking off, T2 = small budget, T3 = moderate budget; anything that would need T4+ escalates to S instead of inflating the budget."
+      },
+      "price": {
+        "input_usd_per_mtok": 1,
+        "output_usd_per_mtok": 5,
+        "cache_read_usd_per_mtok": 0.1,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "disabled": {"score": null, "usd_per_task": null},
+        "enabled": {"score": 17, "usd_per_task": 0.21}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "disabled": null,
+          "enabled": 42
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "enabled",
+        "share_max": null,
+        "context_cap_tokens": 200000
+      },
+      "notes": [
+        "200K context is a routing floor, not a difficulty signal: a mechanical task over a >150K working set is not a Haiku dispatch — shard it or run it S at T0–T1.",
+        "The only measured point (thinking enabled, 17 @ $0.21) is dominated by every measured Luna config. Keep it for single-provider Claude flows, not because it is cheap per task.",
+        "The `disabled` score in the dataset is incomplete and intentionally excluded."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-4-5-haiku"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "claude-sonnet-5",
+      "display_name": "Claude Sonnet 5",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-sonnet-5",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1000000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "effort",
+        "levels": ["low", "medium", "high", "xhigh", "max"],
+        "always_on": false,
+        "default_for_tier": "medium",
+        "note": "Never run Sonnet at `max` — its only measured point and its worst value. Lower efforts are unrated, not proven bad."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 10,
+        "cache_read_usd_per_mtok": 0.2,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 24, "usd_per_task": 0.51},
+        "medium": {"score": 28, "usd_per_task": 1},
+        "high": {"score": 32, "usd_per_task": 1.79},
+        "xhigh": {"score": 34, "usd_per_task": 2.87},
+        "max": {"score": 38, "usd_per_task": 5.09}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 102,
+          "medium": 200,
+          "high": 358,
+          "xhigh": 574,
+          "max": 1018
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 4, "cost_multiplier": 1.96},
+          "medium->high": {"delta_score": 4, "cost_multiplier": 1.79},
+          "high->xhigh": {"delta_score": 2, "cost_multiplier": 1.6},
+          "xhigh->max": {"delta_score": 4, "cost_multiplier": 1.77}
+        }
+      },
+      "caps": {
+        "max_effort": "xhigh",
+        "share_max": null
+      },
+      "notes": [
+        "Sonnet 4.6 ($3/$15) is more expensive than Sonnet 5 — never pick it for cost."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-sonnet-5"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "claude-opus-5",
+      "display_name": "Claude Opus 5",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-opus-5",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1000000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "effort",
+        "levels": ["low", "medium", "high", "xhigh", "max"],
+        "always_on": false,
+        "default_for_tier": "high",
+        "note": "X defaults to `high`; `xhigh` is the T4 expression on Claude."
+      },
+      "price": {
+        "input_usd_per_mtok": 5,
+        "output_usd_per_mtok": 25,
+        "cache_read_usd_per_mtok": 0.5,
+        "cache_write_multiplier": null,
+        "fast_mode_note": "`speed: \"fast\"` reprices Opus to $10/$50 — Fable-tier cost for Opus-tier output. Off by default; it never replaces an F dispatch."
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 39, "usd_per_task": 1.1},
+        "medium": {"score": 45, "usd_per_task": 2.19},
+        "high": {"score": 48, "usd_per_task": 3.61},
+        "xhigh": {"score": 50, "usd_per_task": 4.88},
+        "max": {"score": 51, "usd_per_task": 5.86}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 220,
+          "medium": 438,
+          "high": 722,
+          "xhigh": 976,
+          "max": 1172
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 6, "cost_multiplier": 1.99},
+          "medium->high": {"delta_score": 3, "cost_multiplier": 1.65},
+          "high->xhigh": {"delta_score": 2, "cost_multiplier": 1.35},
+          "xhigh->max": {"delta_score": 1, "cost_multiplier": 1.2}
+        }
+      },
+      "caps": {
+        "max_effort": "xhigh",
+        "share_max": null
+      },
+      "notes": [
+        "Opus 4.8/4.7/4.6 share the $5/$25 price and are fallback-only.",
+        "Every measured Opus config is dominated on $/task by a Sol or Astra config; it stays the Claude X seat for single-provider flows and retention-constrained work."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-opus-5"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "claude-fable-5",
+      "display_name": "Claude Fable 5",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "F",
+      "eligible_tiers": ["F"],
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-fable-5",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1000000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "effort",
+        "levels": ["low", "medium", "high", "xhigh", "max"],
+        "always_on": false,
+        "default_for_tier": "high",
+        "note": "Start F at `high`, not xhigh/max."
+      },
+      "price": {
+        "input_usd_per_mtok": 10,
+        "output_usd_per_mtok": 50,
+        "cache_read_usd_per_mtok": 1,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": null, "usd_per_task": null},
+        "medium": {"score": null, "usd_per_task": null},
+        "high": {"score": null, "usd_per_task": null},
+        "xhigh": {"score": null, "usd_per_task": null},
+        "max": {"score": 50, "usd_per_task": 8.75}
+      },
+      "measurement_note": "The single measured point (max) uses an Opus 4.8 default fallback in AA v4.3.2 and is segregated from standalone cross-model comparisons. `high` — the actual default F config — is unmeasured.",
+      "indices": {
+        "thinking_cost_index": {
+          "low": null,
+          "medium": null,
+          "high": null,
+          "xhigh": null,
+          "max": 1750
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "Default F on Claude. Choose it over the cheaper cross-provider frontier config when the flow is single-provider Claude, when long-context repricing elsewhere is a risk, or when the other model's retention terms make the material unwise to send."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-fable-5"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "claude-fable-5-1",
+      "display_name": "Claude Fable 5.1",
+      "provider": "anthropic",
+      "ladder": "claude",
+      "admission": "incumbent",
+      "tier": "F",
+      "eligible_tiers": ["F"],
+      "api_ids": {
+        "codex": null,
+        "claude": "claude-fable-5-1",
+        "opencode": null,
+        "kilo": null
+      },
+      "context_window": 1000000,
+      "max_output_tokens": 128000,
+      "thinking": {
+        "control": "effort",
+        "levels": ["low", "medium", "high", "xhigh", "max"],
+        "always_on": true,
+        "default_for_tier": null,
+        "note": "Thinking is always on: omit `thinking` or send `adaptive`; `disabled` and `budget_tokens` return 400. Forced `tool_choice` any/tool returns 400 — use `auto` + instruction, `strict: true`, or structured outputs. No assistant prefill; thinking blocks are model-bound and the transcript must stay append-only. Per-message `effort` via a mid-conversation system message avoids cache resets."
+      },
+      "price": {
+        "input_usd_per_mtok": 10,
+        "output_usd_per_mtok": 50,
+        "cache_read_usd_per_mtok": 0.25,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 47, "usd_per_task": 2.37},
+        "medium": {"score": 49, "usd_per_task": 2.98},
+        "high": {"score": 51, "usd_per_task": 3.91},
+        "xhigh": {"score": 53, "usd_per_task": 5.98},
+        "max": {"score": 53, "usd_per_task": 7.63}
+      },
+      "measurement_note": "All displayed scores include the benchmark default fallback and are not standalone scores.",
+      "indices": {
+        "thinking_cost_index": {
+          "low": 474,
+          "medium": 596,
+          "high": 782,
+          "xhigh": 1196,
+          "max": 1526
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 2, "cost_multiplier": 1.26},
+          "medium->high": {"delta_score": 2, "cost_multiplier": 1.31},
+          "high->xhigh": {"delta_score": 2, "cost_multiplier": 1.53},
+          "xhigh->max": {"delta_score": 0, "cost_multiplier": 1.28}
+        }
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": 0.02,
+        "requires_explicit_flag": true
+      },
+      "notes": [
+        "Hard-capped exception: ≤2% of all dispatches, on explicit user request or after a documented Fable 5 F-T4 failure only. Never a default at any tier.",
+        "Shares Fable 5's list price but costs significantly more per completed task — always-on thinking, longer turns, more output tokens.",
+        "Retention caveat: frontier models under mandatory 30-day retention never receive secrets, PII, production tokens or credential-bearing code."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/claude-fable-5-1"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "grok-4-7",
+      "display_name": "Grok 4.7",
+      "provider": "xai",
+      "ladder": null,
+      "admission": "incumbent",
+      "tier": null,
+      "eligible_tiers": [],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": "grok-4.7",
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "reasoning_effort",
+        "levels": ["low", "medium", "high", "xhigh"],
+        "always_on": false,
+        "default_for_tier": null,
+        "note": "`high` is the cheapest measured configuration; `xhigh` shows the same displayed score at greater cost. low/medium are unmeasured — evaluate locally before using them."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 6,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": {
+        "threshold_input_tokens": 200000,
+        "input_multiplier": 2,
+        "output_multiplier": 2,
+        "effective_input_usd_per_mtok": 4,
+        "effective_output_usd_per_mtok": 12,
+        "note": "$2/$6 below 200K input; $4/$12 above."
+      },
+      "measured": {
+        "low": {"score": null, "usd_per_task": null},
+        "medium": {"score": null, "usd_per_task": null},
+        "high": {"score": 46, "usd_per_task": 2.73},
+        "xhigh": {"score": 46, "usd_per_task": 3.74}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": null,
+          "medium": null,
+          "high": 546,
+          "xhigh": 748
+        },
+        "marginal_thinking": {
+          "high->xhigh": {"delta_score": 0, "cost_multiplier": 1.37}
+        }
+      },
+      "caps": {
+        "max_effort": "high",
+        "share_max": null
+      },
+      "notes": [
+        "Best-measured non-incumbent with a canonical API ID. It holds no tier seat: available-but-unrated models get neither an invented score nor automatic promotion to a critical reviewer seat.",
+        "Unknown data retention or endpoint eligibility blocks sending sensitive material there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/grok-4-7",
+        "https://docs.x.ai/developers/release-notes"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "muse-spark-1-3",
+      "display_name": "Muse Spark 1.3",
+      "provider": "muse",
+      "ladder": "third-party",
+      "admission": "candidate",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["xhigh", "max"],
+        "always_on": null,
+        "default_for_tier": "xhigh",
+        "note": "Only `xhigh` and `max` are measured and the provider's real thinking control is unverified — dispatch the provider default and never send an invented effort enum."
+      },
+      "price": {
+        "input_usd_per_mtok": 1.25,
+        "output_usd_per_mtok": 4.25,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "xhigh": {"score": 45, "usd_per_task": 1.37},
+        "max": {"score": 48, "usd_per_task": 1.6}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "xhigh": 274,
+          "max": 320
+        },
+        "marginal_thinking": {
+          "xhigh->max": {"delta_score": 3, "cost_multiplier": 1.17}
+        }
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "`xhigh` is the lower-cost of the two measured points; `max` buys ~3 displayed points for ~17% more — use it only when those capabilities matter.",
+        "Best measured point 48 lands in the X band (45–50). That is a band placement, not a provider-ladder seat.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/muse-spark-1-3"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "grok-4-6",
+      "display_name": "Grok 4.6",
+      "provider": "xai",
+      "ladder": "third-party",
+      "admission": "candidate",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "reasoning_effort",
+        "levels": ["low", "medium", "high", "xhigh"],
+        "always_on": null,
+        "default_for_tier": "medium",
+        "note": "The xAI family exposes `reasoning_effort` (same control documented for Grok 4.7). Whether thinking can be turned off on this checkpoint is not documented — `always_on` stays null."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 6,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 35, "usd_per_task": 0.48},
+        "medium": {"score": 43, "usd_per_task": 1.5},
+        "high": {"score": 44, "usd_per_task": 1.86},
+        "xhigh": {"score": 44, "usd_per_task": 2.32}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": 96,
+          "medium": 300,
+          "high": 372,
+          "xhigh": 464
+        },
+        "marginal_thinking": {
+          "low->medium": {"delta_score": 8, "cost_multiplier": 3.13},
+          "medium->high": {"delta_score": 1, "cost_multiplier": 1.24},
+          "high->xhigh": {"delta_score": 0, "cost_multiplier": 1.25}
+        }
+      },
+      "caps": {
+        "max_effort": "xhigh",
+        "share_max": null
+      },
+      "notes": [
+        "`xhigh` shows no gain over `high` (44 at both) — never pay for it.",
+        "`low` (35) sits below the S band; `medium` (43 @ $1.50) is the cheapest measured effort that reaches the band and is therefore the default.",
+        "Prefer 4.7 when the evidence justifies it, not because the version number is higher.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/grok-4-6",
+        "https://docs.x.ai/developers/release-notes"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "mimo-v2-6-pro",
+      "display_name": "MiMo V2.6 Pro",
+      "provider": "xiaomi",
+      "ladder": "third-party",
+      "admission": "candidate",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": "mimo-v2.6-pro",
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["default"],
+        "always_on": null,
+        "default_for_tier": "default",
+        "note": "Only the provider default is measured; no effort enum is established. Do not send invented effort labels."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.435,
+        "output_usd_per_mtok": 0.87,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "default": {"score": 46, "usd_per_task": 0.13}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "default": 26
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "default",
+        "share_max": null
+      },
+      "notes": [
+        "Best measured score-per-dollar in the dataset (46 @ $0.13) — but only at the provider default, with no multi-effort curve behind it.",
+        "Tier assigned from the single measured `default` point (46 → X band); nothing confirms it holds at other settings.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/mimo-v2-6-pro",
+        "https://mimo.mi.com/docs/en-US/quick-start/usage-guide/text-generation/batch-api"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "gemini-3-8-flash",
+      "display_name": "Gemini 3.8 Flash",
+      "provider": "google",
+      "ladder": "third-party",
+      "admission": "candidate",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["low", "medium", "high"],
+        "always_on": null,
+        "default_for_tier": "medium",
+        "note": "Three measured points exist, but the parameter name that selects them is not verified here — treat the labels as benchmark rows, not as API values."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.75,
+        "output_usd_per_mtok": 3.75,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": 33, "usd_per_task": null},
+        "medium": {"score": 40, "usd_per_task": 0.93},
+        "high": {"score": 41, "usd_per_task": 1.24}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": null,
+          "medium": 186,
+          "high": 248
+        },
+        "marginal_thinking": {
+          "medium->high": {"delta_score": 1, "cost_multiplier": 1.33}
+        }
+      },
+      "caps": {
+        "max_effort": "high",
+        "share_max": null
+      },
+      "notes": [
+        "`medium` when the task fits; `high` buys 1 displayed point for ~33% more. Multimodal or latency needs may still justify it.",
+        "`low` has a displayed score (33) but no published task cost — it cannot be ranked on $/task and is never the default.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/gemini-3-8-flash"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "qwen3-8-max",
+      "display_name": "Qwen3.8 Max",
+      "provider": "alibaba",
+      "ladder": "third-party",
+      "admission": "candidate",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["default"],
+        "always_on": null,
+        "default_for_tier": "default",
+        "note": "Verified provider default only; no multi-effort curve established. Do not send invented effort enums."
+      },
+      "price": {
+        "input_usd_per_mtok": 2,
+        "output_usd_per_mtok": 6,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "default": {"score": 45, "usd_per_task": 5.41}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "default": 1082
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "default",
+        "share_max": null
+      },
+      "notes": [
+        "Tier assigned from the single measured `default` point (45 → X band).",
+        "At $5.41/task it is the most expensive measured point anywhere in the X band — the score places it, it does not recommend it.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/qwen3-8-max"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "glm-5-3",
+      "display_name": "GLM 5.3",
+      "provider": "z.ai",
+      "ladder": "open",
+      "admission": "candidate",
+      "tier": "X",
+      "eligible_tiers": ["X"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["max"],
+        "always_on": null,
+        "default_for_tier": "max",
+        "note": "Only `max` has a complete measured score/cost pair. Confirm the real control before dispatching anything else."
+      },
+      "price": {
+        "input_usd_per_mtok": 1.4,
+        "output_usd_per_mtok": 4.4,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "max": {"score": 45, "usd_per_task": 2.01}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "max": 402
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "GLM 5.3 only `max` measured — that is not evidence to default to `max`. Evaluate lower settings on noncritical, checkable work first.",
+        "Tier assigned from the single measured `max` point (45 → X band); no curve exists to confirm it.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/glm-5-3"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "glm-5-3-flash",
+      "display_name": "GLM 5.3 Flash",
+      "provider": "z.ai",
+      "ladder": "open",
+      "admission": "candidate",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["default"],
+        "always_on": null,
+        "default_for_tier": "default",
+        "note": "Verified provider default; no multi-effort curve established. Do not send invented effort enums."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.15,
+        "output_usd_per_mtok": 0.5,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "default": {"score": 42, "usd_per_task": 0.25}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "default": 50
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "default",
+        "share_max": null
+      },
+      "notes": [
+        "Tier assigned from the single measured `default` point (42 → S band).",
+        "Cheapest measured configuration that reaches the S band ($0.25/task) — the strongest cost argument among the candidates, and still not a review seat.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/glm-5-3-flash"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "kimi-k3",
+      "display_name": "Kimi K3",
+      "provider": "moonshot",
+      "ladder": "open",
+      "admission": "candidate",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["low", "max"],
+        "always_on": null,
+        "default_for_tier": "max",
+        "note": "The dataset carries a `low` row with no score and no cost; only `max` is measured. Confirm controls before treating `low` as callable."
+      },
+      "price": {
+        "input_usd_per_mtok": 3,
+        "output_usd_per_mtok": 15,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "low": {"score": null, "usd_per_task": null},
+        "max": {"score": 44, "usd_per_task": 2}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "low": null,
+          "max": 400
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "Only `max` measured. Confirm controls; evaluate lower settings on noncritical checkable work first.",
+        "`low` is unmeasured, not cheap — it has neither a score nor a cost in the dataset.",
+        "Tier assigned from the single measured `max` point (44 → top of the S band).",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/kimi-k3"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "deepseek-v4-1-flash",
+      "display_name": "DeepSeek V4.1 Flash",
+      "provider": "deepseek",
+      "ladder": "open",
+      "admission": "candidate",
+      "tier": "S",
+      "eligible_tiers": ["S"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["max"],
+        "always_on": null,
+        "default_for_tier": "max",
+        "note": "Only `max` is measured. Confirm the real control before defaulting to it."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.3,
+        "output_usd_per_mtok": 1.2,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "max": {"score": 39, "usd_per_task": 0.27}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "max": 54
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "max",
+        "share_max": null
+      },
+      "notes": [
+        "Only `max` measured. Confirm controls before defaulting to it.",
+        "Tier assigned from the single measured `max` point (39 → bottom of the S band); one point of benchmark noise would move it to W.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/deepseek-v4-1-flash"
+      ],
+      "observed_at": "2026-09-22"
+    },
+    {
+      "key": "minimax-m3",
+      "display_name": "MiniMax M3",
+      "provider": "minimax",
+      "ladder": "open",
+      "admission": "candidate",
+      "tier": "W",
+      "eligible_tiers": ["W"],
+      "api_ids": {
+        "codex": null,
+        "claude": null,
+        "opencode": null,
+        "kilo": null
+      },
+      "canonical_api_id": null,
+      "context_window": null,
+      "max_output_tokens": null,
+      "thinking": {
+        "control": "provider-default",
+        "levels": ["default"],
+        "always_on": null,
+        "default_for_tier": "default",
+        "note": "Verified provider default; no multi-effort curve established. Do not send invented effort enums."
+      },
+      "price": {
+        "input_usd_per_mtok": 0.3,
+        "output_usd_per_mtok": 1.2,
+        "cache_read_usd_per_mtok": null,
+        "cache_write_multiplier": null
+      },
+      "long_context_surcharge": null,
+      "measured": {
+        "default": {"score": 29, "usd_per_task": 0.51}
+      },
+      "indices": {
+        "thinking_cost_index": {
+          "default": 102
+        },
+        "marginal_thinking": {}
+      },
+      "caps": {
+        "max_effort": "default",
+        "share_max": null
+      },
+      "notes": [
+        "Tier assigned from the single measured `default` point (29 → W band).",
+        "Every measured Luna config is both cheaper and higher-scoring — keep this one for flows that cannot reach the Codex ladder, not because it is cheap.",
+        "Candidate seat: ranked only with `--include-candidates` or when the runtime inventory exposes the model; never admitted to an independent-review seat, a T4/T5 pair or a risk-floor review row.",
+        "Data retention and endpoint eligibility are unknown — never send secrets, PII, production tokens or credential-bearing code there."
+      ],
+      "sources": [
+        "https://artificialanalysis.ai/models/minimax-m3"
+      ],
+      "observed_at": "2026-09-22"
+    }
+  ],
+  "candidates": [
+    {
+      "key": "step-5",
+      "display_name": "Step 5",
+      "provider": "stepfun",
+      "canonical_api_id": null,
+      "price": {
+        "input_usd_per_mtok": 1,
+        "output_usd_per_mtok": 2.7
+      },
+      "measured": {
+        "default": {"score": 44, "usd_per_task": 0.72}
+      },
+      "note": "Verified provider default; no multi-effort curve established. Left unpromoted: a single default-only point is enough to place a tier band but the shortlist is capped at 20 routable entries, and Step 5 adds no band coverage that GLM 5.3 Flash (42 @ $0.25) does not already hold more cheaply.",
+      "source": "https://artificialanalysis.ai/models/step-5"
+    }
+  ],
+  "limitations": [
+    "Benchmark cost per task comes from a public suite with its own weighting — it is not any consuming project's workload. Ratios transfer better than absolute dollars.",
+    "Score is not reliability: on a strict end-to-end 'fully resolved' cut the ranking reorders.",
+    "A price or a catalog entry is not evidence that the active harness can dispatch the model. Intersect with a runtime inventory before routing.",
+    "Risk floors, retention terms and reviewer independence are not priced here and are never overridden by a cheaper config."
+  ]
+}