Skip to main content

Model fitness

Where each supported model excels and where it fails — reviewed 2026-07-08 — plus the protocol that makes the list actionable: the fit check. Vendor-reported numbers stay (unverified) until your own benchmark.py run confirms them; your benchmark table, not this page, is your routing policy.

The fit check

Every fleet worker gets mythos-core rule 11: before planning, compare the pending task against your own row. Fit is a gate, not a soft suggestion:

Poor fit → the entire first line is SUGGEST-ESCALATE: <better-suited model> — <reason>, then stop. The operator can insist (orchestrate.py --insist), and the worker then proceeds strictly in scope with shaky output marked (unverified). Scaffolded projects carry the operator's model-priority order in ANCHOR-CONVENTIONS.md, so the suggestion names the nearest better-fitting model from that user's list. Suggesting down-tier is equally required — boilerplate on a frontier model is the mirror-image failure.

orchestrate.py honors a SUGGEST-ESCALATE first line immediately (escalate, or hold in detached mode) without burning retry attempts.

Frontier / API models

ModelExcels atWeak at / quirks
Claude Fable 5Long-horizon autonomy, large migrations, multi-service debugging, final reviewCredit-metered — keystrokes on it are an economics failure
Claude Opus 4.8Deep single-problem reasoning, architecture, securityOverkill for scoped edits
Claude Sonnet 5Default executor: scoped multi-file edits, solid tool useHands multi-hour autonomy up a tier
Claude Haiku 4.5Classification, summaries, spec-tuningMulti-file reasoning, subtle bugs
GPT-5.6 SolAgentic coding + cybersecurity (unverified, vendor)System-card-documented over-eagerness: unrequested actions, claiming unperformed work
GPT-5.6 Terra~GPT-5.5 quality at ~half cost — the executor pickSame system-card caveats as Sol
GPT-5.6 LunaFrontier-adjacent at $1/$6 — tuner/light executorKeep off architecture and review
ChatGPT (GPT-5.5 + Instant Mini fallback)Conversational spec-shaping, piloted one-step turnsNo execution; fallback varies the tier mid-session
Grok 4.5Terminal/CLI tasks (≈GPT-5.5 class), long tool-use runs, token efficiency, price; Preferred catalog tier = midMeasurably weaker at repo-scale issue resolution — decompose to file-scoped specs; reasoning_effort defaults high (use /effort low for mechanical); high effort ≠ frontier promotion; community-reported tool-use flakiness
Gemini 2.5-classLong-context ingestion, multimodalSame external-verification rules as everyone
Nemotron (NIM)Local planner/critic stand-in; clean thinking toggleFabricates unfamiliar APIs under pressure

Local models

Model names link to the official quick start. See also Local Models for Anchor quirks and serve notes.

ModelExcels atWeak at / quirks
Qwen3 32B / 30B-A3BSpec-driven edits; 32B /think checklist criticSmall plans only as planner; never greedy while thinking
Gemma 3 27BBest instruction following per sizeNo system role; agreeable — needs the BLOCKED guardrail
Mistral Small 3.xFast executor, best local function callingTerse — drops footers under load; won't push back
DeepSeek-R1 distillsBest local critic per GB; hard single problemsNever an executor; no system prompt; greedy breaks it
Llama 3.3 70BGeneralist executor+criticConfident fabrication; verbose without caps

The full matrix with pricing, dates, and per-entry sourcing lives in anchor/model-fitness.md in this repo, and is scaffolded into projects as .anchor/model-fitness.md.