Models and cost
You bring your own model provider and pay them directly. AWOS records the cost of every model call and stops a job at its budget.
Cloud: Claude Haiku 5.5
The default cloud model is Claude Haiku 5.5, called directly through the Anthropic API with prompt caching on. Set ANTHROPIC_API_KEY and check it with awos keys.
| Measured run | Result | Cost |
|---|---|---|
| 73 real GitHub issues from 26 Python projects, one attempt each, judged by hidden tests | 72 / 73 solved | $0.0071 per issue on average |
| Finance tracker built from a README spec in an empty folder (8 features) | 58 / 58 hidden tests, 8.8 min | $0.16 |
| Six realistic multi-step tasks (migration, debugging, performance, features) | 6 / 6 solved | $0.46 for all six |
Runs of 2026-10-11, one run per task. These are public projects, so the model may have seen some fixes in training, and your results will vary with your codebase and tests. Costs are AWOS's own ledger at list prices.
Local: a free model on your machine
Run an open model on your own computer and point AWOS at it with AWOS_PROVIDER=local. AWOS talks to a local OpenAI-compatible server such as llama-server; awos keys reports whether it is reachable. A local job costs $0 per ticket and needs no network.
- Measured: Qwen3.5 9B on a 16 GB MacBook solved 8 of 10 easy real issues through AWOS (2026-10).
- It is slow: about 20 minutes or more per issue on that laptop, so it suits overnight queues of routine fixes.
- Combine it with
--privacy local_onlyto guarantee a job cannot reach a cloud model (see Security).
Budgets
awos submit --budget USDcaps a job's spend. The default is 1.0;0means no cap.- The cap covers the whole job, including any continuation: each attempt gets what is left.
- A job stops when it reaches its cap.
awos status --digestshows today's spend, counted from AWOS's own model calls.