Models: the latest strong model per provider¶

This page is part of Harnesses: versions, the contract, the canary and models.

A model: value is either a pinned id (claude-opus-5-5, gpt-6.1-sol), passed through as it is, or a policy:

Environments (model: frontier, or per harness model: {claude-code: frontier}) and repo policies (hangar repo set acme/app --model frontier) accept a policy. A policy in a repo policy applies to every harness, where a pinned id applies only to the policy's own harness. So do submit_job(model=…), send_to_session(model=…) and run_agent(model=…).

Where the ids come from¶

The Anthropic side was checked against the API reference on 2026-10-14 (platform.claude.com/docs/en/api/models/list, which docs.claude.com redirects to). It says "More recently released models are listed first". Each model has id, display_name, created_at, lifecycle (active, deprecated, retired) and line (haiku, sonnet, opus, fable, mythos, or null; "do not infer a line from the id"). Pages use limit (up to 1000), after_id and has_more. Mythos is Fable for verified organizations; it appears only in their lists.

Why Google's tier comes after the version: Gemini's flash line moves faster than pro. Not verified against Google's own pages: that Gemini 3.8 Flash (stable, September 2026) is stronger than Gemini 3.1 Pro Preview (February). Google's model list (ai.google.dev/gemini-api/docs/models, read 2026-10-14) calls 3.8 Flash "our most intelligent Flash model, engineered for long-horizon software engineering" and lists 3.1 Pro as a preview, but compares neither with the other; the ranking rests on third-party indexes. Where it matters, pin a Gemini id. A lite model is never frontier while anything else is listed, so Google's next flash-lite doesn't take over by being newest.

Verified on 2026-10-14 against the vendors' pages: Anthropic's models overview and pricing (Fable 5.1 "for demanding reasoning and long-horizon agentic work", $10/$50 per MTok; Opus 5.5 the general starting point, $4/$20); OpenAI's models and pricing pages (gpt-6-astra "our most capable model", $10/$50, listed for the API and the Codex CLI; gpt-6.1-sol "near-Astra performance at a lower cost", $2/$10). OpenRouter's -pro variants (gpt-6-astra-pro) are not on OpenAI's pages and are never picked.

What frontier costs. On Anthropic it picks Claude Fable 5.1 ($10/$50 per MTok, against Opus 5.5's $4/$20), and on OpenAI gpt-6-astra ($10/$50, against gpt-6.1-sol's $2/$10): roughly 2.5 to 5 times balanced. That is why jobs default to balanced (the owner's decision); frontier stays a line away for an environment, a repo policy or one job. Either way:

• every id a policy resolves to is in hangar's price table, including OpenRouter's dotted ids and agy's thinking-level ids, so job budgets and the daily caps (repo-policy.md) count its cost;
• set a budget: no job budget or daily cap is set by default, and hangar doctor (model spend) and hangar status warn until one is. HANGAR_JOB_BUDGET_MAX_COST_USD, a repo policy's job_budget and max_usd_per_day, and HANGAR_PERSON_MAX_USD_PER_DAY bound it;
• the review packet on each agent PR names the model the job's session ran, the policy it came from and its list price (modelclaude-fable-5-1(frontier; list price $10/$50 per MTok in/out)), and says when the job had no budget; hangar env list shows the same for each environment's harnesses;
• ask for frontier where the last bit of capability matters more than the price.

Concrete ids and floating aliases¶

A pinned id is only reproducible if the provider can't move what it means. hangar calls an id concrete when it is in the provider's cached model list or has a known versioned shape (claude-opus-5-5, claude-sonnet-4-20250514, gpt-6-astra, gpt-*-YYYY-MM-DD, gemini-3.8-flash, openrouter/anthropic/claude-opus-5.5), and a floating alias when it is *-latest, one of OpenRouter's ~vendor/… aliases (or openrouter/auto), or a bare family name (opus, sonnet, haiku, gpt-5, gemini-pro), even when a provider lists it. An id that is neither listed nor of a known shape is never treated as concrete. With no list cached (no list credential, or not fetched yet), only the shape decides and hangar says so (modelpolicy.is_concrete; every built-in id above is of a known shape).

• Experiments refuse aliases. An experiment (and the trials API, #178) won't start when a variant's model, after its policy is resolved, isn't concrete; the error names the alias and the ids balanced and frontier resolve to now. A policy is fine: it is resolved once, at the start, and every trial gets that id (jobs/experiments.md).
• Jobs and sessions keep working with aliases, but their manifests record model.concrete: false (the outbox's run.json, #163, carries it as pinned: false), so evaluation leaves them out.

Pinned, cached, and what happens when a list can't be had¶

• Pinned at creation. A sandbox resolves its environment's model when it is created (get_sandbox → model), and a session or job's session when it is created (get_session → model, model_policy). Its turns use that id until someone switches on purpose (send_to_session(model=…), with the usual cold-cache warning). A newer model in the list never changes a running conversation (spec 0012: each model has its own prompt cache).
• Cached. Lists are fetched in the background, at startup for the providers the environments' policies need and then every HANGAR_MODEL_LIST_TTL_MINUTES (360). A create never waits on a provider. A stale list is still used, and refreshed.
• Fallback, said out loud. With no list (no key, the provider unreachable, nothing qualifying), hangar uses the last id it resolved for that harness and policy. Failing that, it uses a built-in one, dated 2026-10-14: balanced: claude-opus-5-5, Codex gpt-6.1-sol; frontier: claude-fable-5-1, Codex gpt-6-astra; pi openrouter/anthropic/claude-opus-5.5 and …/claude-fable-5.1 (and one per vendor for frontier@openai and the like); gemini-3.8-flash-high. Every built-in policy has one, so a session never starts on the harness's own default and switches model when the list arrives. The session's activity says which, and why. list_environments shows what each harness's model resolves to now (models, with source: live, cached, last, seed or pinned).
• Keys. OpenRouter's list is public. The others need an API key in the vault, named in HANGAR_MODEL_LIST_CREDENTIALS:

hangar cred add anthropic-models --harness webhook --kind generic   # paste the key; never a sandbox's
hangar config set HANGAR_MODEL_LIST_CREDENTIALS '{"anthropic": "anthropic-models", "openai": "openai-models", "google": "gemini-models"}'

The key only lists models. Lists are per account, so use a key from the account the agents run on: an Anthropic key without Mythos access never resolves to Mythos. Codex with a ChatGPT login has no API key to list with. Give it an OpenAI key from the same organization, or pin codex: gpt-6-astra if you have it. A ChatGPT login is also interactive-only: jobs, schedules and experiments run Codex on an OpenAI API key, or a ChatGPT Enterprise Codex access token stored with --kind api_key (team.md).