Appearance
How it works
OrcaCompass answers one question for an orchestrator: which agent, model and reasoning effort should this worker get? It narrows the field in four moves: what the job needs, which models can do it, which of those costs least, and which subscription has room. This page follows one job through every step.
The scores, prices and usage below come from a real run in November 2026. Your models, data and usage will differ; the steps stay the same.
The job
Add a CSV export button to the reports page, with tests.
A few files in one area, tests to run and fix until they pass, and no hard design decision.
1. The orchestrator rates the job
The orchestrator (the agent handing out work) rates what the job needs on three axes, and how much the worker should think:
sh
orcacompass route --need coding=mid,agentic=mid,reasoning=low --demand moderate --jsonEach level is a minimum score on that axis. min has no minimum.
| Axis | low | mid | high | top |
|---|---|---|---|---|
| coding | 60 | 75 | 82 | 87 |
| agentic | 35 | 70 | 80 | 82 |
| reasoning | 70 | 80 | 86 | 90 |
low sits under the good small models, so a simple job can go to one. mid sits under the current daily workers and high under the frontier models. top is for the rare job high is not enough for; today only Opus 5.5 and Fable 5.2 reach it on every axis. A model that meets a level also meets every level below it.
So this job asks for coding 75 or more, agentic 70 or more and reasoning 70 or more. The cuts travel with the evidence snapshot, so a data refresh can move them. Skill workflow explains how to rate a job; --tier is a shorthand for the same level on every axis.
2. OrcaCompass finds what you can launch
It looks for each agent's CLI, asks Orca which agents are switched on, and reads each agent's current model list. In the example, Claude Code, Codex and Cursor are available; Grok and Antigravity are switched off in Orca, and OpenCode is not installed. Banned models and models an agent refused at an earlier launch are left out here. See Models and agents.
3. Evidence decides who can do the job
Every model has a coding, agentic and reasoning score from the LLMBenchmark evidence snapshot. A model qualifies only when it meets the minimum on all three axes. Price plays no part in this step.
| Model | coding ≥ 75 | agentic ≥ 70 | reasoning ≥ 70 | Result |
|---|---|---|---|---|
| Haiku 5.5 | 68.4 | 53.5 | 74.6 | out |
| Composer 2.5 | 67.1 | 35.7 (assumed) | 75.0 | out |
| GPT-6.1 Terra | 76.4 | 69.1 | 82.3 | out |
| Grok 4.7 | 64.8 | 43.7 | 75.1 | out |
| GPT-6.1 Sol | 79.8 | 75.5 | 88.7 | qualifies |
| Sonnet 5.5 | 84.3 | 78.2 | 87.3 | qualifies |
| Opus 5.5 | 87.3 | 82.0 | 90.6 | qualifies |
| GPT-6 Astra | 83.8 | 80.5 | 89.7 | qualifies |
| Fable 5.2 | 87.8 | 85.4 | 94.7 | qualifies |
Each model that falls short is listed under skipped with the reason, for example below need on agentic (69.1 < 70). A model with no evidence at all is unranked and skipped. A new release of a known family inherits the evidence of the release before it until its own data arrives. Composer 2.5 has no agentic benchmark, so Kimi K2.5's score stands in for it and is labelled assumed. See Model data.
4. Price decides how much model to spend
Each model also has a cost tier, set by its API price in the snapshot, blended 3:1 input to output, in USD per million tokens. The price only sorts models from cheap to expensive; it is not what your subscription charges.
| Cost tier | Blended $/M | Qualifying models in the example |
|---|---|---|
| Light | below 1 | none |
| Daily | 1 to below 5 | Sol ($4), Sonnet 5.5 ($4), Gemini 4 Carbon ($4) |
| Advanced | 5 to below 12 | Opus 5.5 ($8) |
| Frontier | 12 or above | Fable 5.2 ($16), Astra ($20) |
The cheapest tier with at least one qualifying model is the fit group: Daily. Opus, Fable and Astra could do the job, but they are not used for it. A pricier tier is used when no cheaper tier has a qualifying model, or when every qualifying route in the cheaper tiers is blocked or close to its usage limit (see step 5). Change the thresholds with costBands, or move a model with tierOverrides; see Tiers and routing.
5. Subscription headroom picks the winner
Inside the fit group, OrcaCompass reads your live usage from Orca and projects it: at the current pace, how much of each usage window will be left when it resets? The route with the most room wins. Price no longer matters at this point.
| Route | Usage | Headroom |
|---|---|---|
| Codex, Sol | weekly 8% used, 31% of the week gone | 62% |
| Claude Code, Sonnet 5.5 | weekly reset less than a day ago | 55% |
| Cursor, Sonnet 5.5 or Gemini 4 Carbon | 76% used, 78% of the window gone | 3% |
A route at or over a limit is blocked and skipped. A route expected to end its usage window above usageRiskLimit (90%) is strained: it ranks after the others. When every route in the fit group is blocked or strained, the cheapest pricier tier with a route in good standing takes over, and the route carries a warning that names the tier it passed over. When every covering route is strained, the job keeps its level on the cheapest tier's route with the most room, with a warning; it is downgraded only when every covering route is blocked. In the example the Cursor routes are strained, but Codex and Claude Code are not, so the fit group stays Daily. Routes handed out in the last handoutMinutes (10) count as handoutPercent (5%) more usage, so workers launched back to back spread across subscriptions. A favorite set with orcacompass prefer can take the place of this pick; see Configuration.
6. Effort follows the demand
The demand moderate maps to medium, an effort Sol supports, passed as --effort medium. Agents that encode effort in the model id get it there instead. See Reasoning effort.
7. The answer
The route comes back with launch arguments and the reasons behind it:
text
route: codex gpt-6.1-sol, effort medium, confidence high
reasons:
codex gpt-6.1-sol: weekly 8% used, 31% of window elapsed (headroom 62%) → on pace for 26% before reset in 4d
needs coding=mid agentic=mid reasoning=low; fit group: daily cost tier (4 models)
cost: daily ($4.00/M blended, maker price)
coding 79.8 measured, agentic 75.5 measured, reasoning 88.7 measured
effort: medium from moderate reasoning demand (balanced policy), passed as --effort
next best: claude claude-sonnet-5-5 (weekly headroom 55%, resets in 6d)The orchestrator appends launchArgs to orca orchestration worker-start, then runs orcacompass verify --dispatch <id> to confirm the agent accepted the model. If the agent refused it, OrcaCompass remembers the refusal and the next route picks another model. See Verify and rejections and Commands for the full JSON.
What changes the answer
- Codex is at its limit. Same fit group, next most room: Sonnet 5.5 through Claude Code.
- The job needs more coding. At
coding=high(82 or more), Sol (79.8) drops out and Sonnet 5.5 (84.3) still qualifies in the Daily tier. - No Daily model qualifies. The fit group moves up to Advanced (Opus 5.5), then Frontier.
- Every Daily route is blocked or close to its limit. The fit group moves up to the cheapest tier with a route below
usageRiskLimit: Opus 5.5 through Claude Code, with a warning. - Every qualifying route is close to its limit. The job keeps its level: the cheapest tier with an unblocked qualifying route (Daily here) is used, and its route with the most room, with a warning.
- Every qualifying route is blocked. OrcaCompass picks the closest model that falls short and marks the route
downgraded, or returnswaitwith--no-downgrade. See Fallbacks. - Nothing can be routed. Cursor
autois the last fallback before waiting; Cursor then picks the model itself.