Model guide

This is the full list of every model stella seeds into its catalog, with the numbers you need to decide: context window, list price, and what the model is good at. The pricing figures come from stella's seed catalog; prices are list $/Mtok (input / output / cached-input). The Released and Best for columns are added by these docs for context — the catalog itself doesn't store or sync them. Verify with your provider before budgeting, since prices and lineups change.

The catalog

Model by role

A run distributes work across separate roles, and each one wants different things from a model:

• Worker. Writes the code. It wants the strongest coding model you can afford — this is where quality is worth paying for. Worker tokens dominate a run, but most of them are cached-input tokens (see below), so the real price gap between tiers is smaller than the list price suggests.
• Verifier. Checks the work at high effort. It wants a different family than the worker: cross-family judging avoids same-model bias, where a model forgives the mistakes it would have made itself. A Claude worker with a GPT verifier (or the other way around) catches more than either family judging its own work. This is goal mode's standing verifier.

An installed wrapper plugin can add further roles of its own. What it runs, and how it routes them, is up to its manifest, not stella's.

Ready-made lineups you can copy:

Each one links to a worked settings.json profile in Examples. Wire any of them up per agent with agent_engine_config.

The live model catalog (stella models refresh)

The seed table above can't keep up with weekly model launches. Without a refresh, an invalid slug is only caught for seeded providers — OpenRouter and custom gateways accept anything and only fail later, mid-run, on the wire. The live catalog fixes both:

stella models refresh                      # sync the master list (models.dev, no API key)
stella models list --provider anthropic    # browse valid slugs + live pricing

• Master list, locally. One public document, needing no login, covers every provider stella can select (plus about 160 more). It lands in a user-tier SQLite catalog (catalog.db, next to usage.db): model cards keyed per API provider, and model card versions holding each pricing configuration. A refresh adds a new version only when pricing actually changed, so the version history is a real change history, and the latest version is the pricing that displays everywhere (cost metering, stella models list, the deck's model picker). On top of that master list, refresh also overlays each configured provider's own live /models listing (OpenRouter, Anthropic, Gemini, and any OpenAI-compatible endpoint), so newly-shipped slugs and per-model pricing/capability come straight from the serving provider — this overlay resolves through each provider's configured credential (only the models.dev master list needs no key).
• Incremental. The HTTP ETag is saved and reused, so an unchanged list is one conditional request and zero writes. A stale list (older than 24 hours) re-syncs automatically at startup, and STELLA_CATALOG_AUTO_REFRESH=0 turns that re-sync off — though when automatic syncing can first fire differs by source (see below).
• Strict slug validation, everywhere. Once a provider's master-list rows are synced, an invalid --model slug shows a clear, immediate error with did-you-mean suggestions — before any wire call, for every provider including OpenRouter and settings.json-defined gateways. The seed table always passes, and a custom endpoint the master list doesn't know about is trusted as-is, exactly as before.
• Exact telemetry. A model_aliases table connects every string form a model travels under — bare slug, provider/slug, date-stamped snapshots, region-prefixed Bedrock profiles, and any id a provider echoes on the wire (learned automatically from telemetry) — to a single (api provider, model provider, model version, model card) record, so per-model cost analytics never split one model across its different spellings.

When sources sync

Because the native overlay resolves through each provider's configured credential, the slugs and prices you get back are exactly what that key can see. This is the same credential chain that authorizes inference — resolved once, used for both:

Only the models.dev master list is keyless.

Prompt caching changes the math

The cached-input column is what most worker input actually costs partway through a run. stella keeps the prompt prefix stable on purpose — the context engine adds to it rather than rewriting it — so after the first turn, most of every request is a cache hit billed at the cached-input rate: a 10x–20x discount depending on provider. This is why real per-run cost lands far below a simple input-tokens-times-list-price estimate, and why a model's cached-input price matters more than its headline input price for long agentic runs.

The cache-write rate

Reads are only half the story. Writing a prefix into the cache is itself billed, at a premium over the input rate, and the seed catalog carries that figure per model alongside the other three. It doesn't appear in any column above, because it is not a price you shop on — it's an upfront cost you earn back over time:

The Claude family is the only one in the seed catalog carrying a write premium — and, not coincidentally, the only one where caching is opt-in rather than automatic: Anthropic, Bedrock, and Claude routed through OpenRouter each need an explicit cache marker, while the rest of the field caches for you automatically. With reads at roughly a tenth of the input rate and writes at 1.25x, a written prefix breaks even after two requests — and an agent loop replays its prefix every turn, so the premium is paid once and earned back right away.

There's one failure case worth knowing: a prefix that goes cold before it is read back is a write you paid for and never used. stella stats reports this directly — a negative SAVED ($) means the write premium cost more than the reads it bought, and TTL REWRITES counts the prefixes that expired between turns. See stella stats.