Models
The right model for every task, picked for you
You describe the evaluation. Evals.sh looks at what the task needs (tool use, reasoning depth, context size, vision) and routes it to the model best suited to the job. The runtime, tools and scoring stay identical whichever model does the reasoning.
- Model families
- 3
- Models in the router
- 6
- Evaluations completed
- 110,000+
- Browser actions executed
- 2.8M
Families
The model families we build on
OpenAI
GPT
Strong tool use and long-horizon reasoning. Our default for agentic browser sessions, where the model has to plan many steps ahead and recover from unexpected UI states.
Anthropic
Claude
Careful, well-calibrated code reasoning. Particularly good at reading large repositories and explaining why a pattern is a problem, not just that it is one.
Moonshot AI
Kimi
Open-weight frontier models built for long-running agents, with native vision and million-token context. A strong fit for extended browser sessions and repository-wide review.
Catalogue
Models in the router
Every model the router can choose from. Relative speed and reasoning depth are our own qualitative ratings from running evaluation workloads, not vendor benchmarks.
| Model | Family | Tier | Context | Speed | Depth | Best for |
|---|---|---|---|---|---|---|
GPT-6 LunaDefault gpt-6-luna | GPT | Balanced | 1.05M | Project evaluation, UI audits, API testing | ||
GPT-6 Sol gpt-6-sol | GPT | Frontier | 1.05M | Complex multi-step flows, Code review | ||
GPT-6 Astra gpt-6-astra | GPT | Frontier | 1.05M | Hardest code reviews, Ambiguous rubrics | ||
Claude Opus 5.5 claude-opus-5-5 | Claude | Frontier | 1M | Code review, Assignment grading | ||
Claude Fable 5.1 claude-fable-5-1 | Claude | Frontier | 1M | Large repositories, Architecture review | ||
Kimi K3 kimi-k3 | Kimi | Frontier | 1M | Long browser sessions, Screenshot analysis |
Routing
You don't pick a model. Evals.sh does.
Every module stresses models differently. The router reads the shape of each job and sends it to the model that handles that shape best, so a quick audit and a repository-wide review don't pay for the same model.
- Project evaluation
- Reliable tool use over long, multi-step browser sessions
- Code review
- Long context and careful reasoning across many files
- Frontend audit
- Fast interpretation of metrics and screenshots
- API testing
- Precise structured output and edge-case generation
- Assignment grading
- Consistent rubric application across many submissions
Consistency
Same guarantees on every model
One runtime
Identical tools, sandbox and step logging regardless of the model doing the reasoning.
One scoring engine
Evidence is scored against the same rubric, so results stay comparable.
No training
Inference runs under enterprise terms that exclude training on your data.
Same security
Encryption, isolation and retention policies apply to every request.
Get started
Run your first evaluation in minutes.