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
| Model | Excels at | Weak at / quirks |
|---|---|---|
| Claude Fable 5 | Long-horizon autonomy, large migrations, multi-service debugging, final review | Credit-metered — keystrokes on it are an economics failure |
| Claude Opus 4.8 | Deep single-problem reasoning, architecture, security | Overkill for scoped edits |
| Claude Sonnet 5 | Default executor: scoped multi-file edits, solid tool use | Hands multi-hour autonomy up a tier |
| Claude Haiku 4.5 | Classification, summaries, spec-tuning | Multi-file reasoning, subtle bugs |
| GPT-5.6 Sol | Agentic 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 pick | Same system-card caveats as Sol |
| GPT-5.6 Luna | Frontier-adjacent at $1/$6 — tuner/light executor | Keep off architecture and review |
| ChatGPT (GPT-5.5 + Instant Mini fallback) | Conversational spec-shaping, piloted one-step turns | No execution; fallback varies the tier mid-session |
| Grok 4.5 | Terminal/CLI tasks (≈GPT-5.5 class), long tool-use runs, token efficiency, price; Preferred catalog tier = mid | Measurably 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-class | Long-context ingestion, multimodal | Same external-verification rules as everyone |
| Nemotron (NIM) | Local planner/critic stand-in; clean thinking toggle | Fabricates unfamiliar APIs under pressure |
Local models
Model names link to the official quick start. See also Local Models for Anchor quirks and serve notes.
| Model | Excels at | Weak at / quirks |
|---|---|---|
| Qwen3 32B / 30B-A3B | Spec-driven edits; 32B /think checklist critic | Small plans only as planner; never greedy while thinking |
| Gemma 3 27B | Best instruction following per size | No system role; agreeable — needs the BLOCKED guardrail |
| Mistral Small 3.x | Fast executor, best local function calling | Terse — drops footers under load; won't push back |
| DeepSeek-R1 distills | Best local critic per GB; hard single problems | Never an executor; no system prompt; greedy breaks it |
| Llama 3.3 70B | Generalist executor+critic | Confident 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.