Skip to content

Models and agents ​

Models and providers change often: run orcacompass status for what your own machine and subscriptions offer right now. Available models vary by account and agent version.

Agents ​

An agent is a coding CLI that Orca can launch as a worker. OrcaCompass only considers agents whose CLI is installed, that are enabled in Orca, that Orca does not report as unavailable, and that you have not switched off with orcacompass agents disable or agents only, so the options differ per user. orcacompass agents shows each agent's verdict and the reason.

AgentOrca agent IDExecutableModel list comes fromEffort levelsUsage data from OrcaStatus
Claude CodeclaudeclaudeOrca's model catalog, then its cached account catalog, then Anthropic families in the evidence snapshotfrom the same source, else low to maxsession, weekly, Fable weeklySupported
CodexcodexcodexOrca's model catalog, then ~/.codex/models_cache.json, then the evidence snapshot (Codex ladder families only)per model, from those filessession, weeklySupported
Cursorcursorcursor-agent (also %LOCALAPPDATA%\cursor-agent)cursor-agent modelsencoded in the model ID (…-high)monthly, plus the Cursor Models and Other Models bucketsSupported
Grok Build (xAI)grokgrok (also $GROK_HOME/bin, ~/.grok/bin)~/.grok/models_cache.json, then grok modelsper model from that cache; else low to xhigh (Grok 4.5 up to high)weekly credits or monthly budgetSupported; verified on Windows
Antigravity (Google)antigravityagy (also %LOCALAPPDATA%\agy\bin, ~/.local/bin)agy modelslow to highweekly, and 5-hour when reported, per quota groupSupported; verified on Windows
OpenCodeopencodeopencode (also ~/.opencode/bin)opencode modelsnone (Orca passes no effort to OpenCode)none: pay per use or localSupported; last resort by default
Muse (Meta)musemusebuilt-in list: muse-wonder-1.0, muse-spark-1.1, muse-ember-1.0, muse-spark-1.3 (neither Muse nor Orca lists Muse models)minimal to max, default high (Orca's Muse menu)none (Orca reports no Muse usage)Supported; model ids not yet checked against a Muse account

OpenCode lists every model it can reach: providers with an API key (OpenRouter and others, pay per token), OpenCode's own models, and local servers such as Ollama (this machine's resources). It has no usage window, so its routes are low confidence and say how they are paid for (runs on this machine (ollama), free model through openrouter, pay per use through openrouter). Because it costs money or local resources, it is in lastResortAgents by default: routing first tries every other agent, including a lower tier, and uses OpenCode only when nothing else can take the job (or instead of waiting under --no-downgrade), with a warning. --agent opencode, a preference for the tier (orcacompass prefer daily opencode [model]) or "lastResortAgents": [] makes it an ordinary agent; without usage data it still ranks after agents that have it. OpenRouter's ~vendor/x-latest aliases are skipped, a model listed twice is kept from its cheapest source, and a -free copy is rated as the model it copies. Local models usually have no benchmark evidence, so they are unranked and skipped; to use one, name it in a preference (orcacompass prefer light opencode gemma4:12b). tierOverrides sets only a model's cost tier and cannot make it qualify. Orca accepts --model for OpenCode only in an existing worktree (--worktree current or a selector), not a new one.

Model lists are read when the placement table is built: every placementsTtlHours (5 hours by default), on orcacompass refresh, or when your tier settings change (see Tiers and routing). Claude Code's and Codex's lists come first from Orca's model catalog (agent-model-catalog.db in Orca's data folder, %APPDATA%\orca on Windows): the models Orca's model picker offers, with the efforts Orca allows for each. Orca saves only these two agents' lists there; it lists the others live, as OrcaCompass does. The catalog is read read-only and needs node:sqlite (Node.js 22.13+ without flags); without it, or when Orca lists nothing for the agent, the agent's own source below is used and its note says why. Cursor, Antigravity and OpenCode use live commands; Grok first reads its native model cache. For live listings, the last good list is used only when the command fails; a sign-in prompt skips the agent as not signed in. After Orca's catalog, Claude's list is the catalog Claude Code caches under ~/.claude/cache/model-catalog for your account, then the evidence snapshot. Model makers reachable today: Anthropic and OpenAI (directly and through Cursor); Google, xAI, Meta, Moonshot, Zhipu, and DeepSeek through Cursor; xAI through Grok Build; Google, Anthropic and OpenAI through Antigravity; and Cursor's own Auto. Gemini models route through Antigravity or Cursor; the Gemini CLI is not an agent here.

How a model gets its profile ​

Each model has two separate things. Its capability is the coding, agentic and reasoning scores from the benchmark evidence; a new release with no data yet inherits the evidence of the release before it. Its cost tier comes from its API price, or from a family hint when there is no price. The job's needs pick the models that qualify, the cheapest cost tier among them is the fit group, and usage headroom picks inside it. orcacompass ranking shows the evidence, orcacompass tiers the cost tiers. Cursor auto is only a fallback and has no profile. See How it works for a worked example and Tiers and routing for the full order.

Effort follows the job's effort demand (--demand); without it the requested tier's default applies.

Inspect available models ​

Run orcacompass status to see the models discovered for your installed agents, their tiers, and any exclusions. Availability depends on your account, agent version, and current provider catalog. Model IDs in this guide are examples, not a guarantee of access.

Cursor maps Cursor-trained models (cursor- IDs), Composer and Auto to the Cursor Models bucket, and other models to Other Models. This mapping is inferred because Orca does not report it directly.

Antigravity keeps provider launch IDs while normalizing them for classification: for example, claude-opus-4-6-thinking is classified as claude-opus-4-6. Its Gemini and Claude/GPT quota groups are evaluated separately.

Your preferences and bans ​

orcacompass prefer <tier> <agent> [model] stores a preferred route per tier in pins. It must survive every exclusion (your agent choice, request flags, Orca's switches, availability, bans) and be unblocked. When projected usage exceeds preferRiskLimit (100% by default), the best unblocked alternative with known usage within that limit is chosen. If every alternative is risky, unavailable, blocked, or has unknown usage, the favorite is retained. Projection is used% / elapsed fraction, with no extrapolation under 10% into a window. The route explains a switch, for example preferred codex gpt-6.1-terra skipped: weekly on pace for 150% (limit 100%), resets in 2d. It also explains when an over-limit favorite is retained.

orcacompass ban <model-or-family> stores entries in banModels. An entry matches a model's base ID, so claude-opus-5-5 also covers Cursor's claude-opus-5-5-high, or its catalog family, so grok covers Grok 4.5 to 4.7 but not Grok Build (grok-build). Banned models are never chosen or shown as alternatives and are listed as banned by you.

Models an agent refuses at launch ​

A listed model can still be one the account cannot use, and Orca starts the worker anyway, reporting the requested model as effective. orcacompass verify --dispatch <id> reads the new worker's terminal and recognizes each agent's refusal (seen November 2026, with an unknown or unavailable model):

AgentIn an Orca workerRefusal
Codexidles at its prompt, no worker_doneThe 'claude-opus-5-5' model is not supported when using Codex with a ChatGPT account.
Claude Codeidles at its prompt, no worker_doneThere's an issue with the selected model (gpt-6-terra). It may not exist or you may not have access to it.
Cursorcursor-agent exits 1; Orca fails the start after its 60 s readiness waitCannot use this model: claude-opus-9-9. Available models: …
Grok Buildexits 1Couldn't set model 'grok-9-fake': Invalid params: "unknown model id".
Antigravityexits 1invalid model selection (--model "gemini-9-fake" …): model gemini-9-fake is not recognized as a known model …

A refusal counts only on a line that names the model. Soft warnings that still run the model (Codex's Model metadata for `x` not found, Claude Code's doesn't recognize the model "x") do not count. A refused agent and model is skipped for rejectionTtlHours (a week by default) and listed as refused by <agent> …; orcacompass rejected --clear forgets them.

Agents you enabled in Orca (experimental) ​

Orca keeps its agent switches (Settings → Agents) in settings.disabledTuiAgents, a list of disabled agent IDs (empty means every detected agent is enabled). No Orca command exposes it, so OrcaCompass reads it read-only from the active profile's profile-state.db, or orca-data.json for a profile without a database. A disabled agent is skipped with the reason not enabled in Orca. If the setting cannot be read, or the field is missing or not a list, it is reported as unknown and no agent is excluded for it.