Models
The right model for each job
Forge picks a model for every job from what it measured on real projects: the strongest where depth matters most, a lower-cost one where it does as well. You can pick your own in the app, for one job or all of them.
Which model each job uses
When a model isn’t available, Forge uses the next one in the list. If a model declines a request, Forge tries once more on another provider’s model. The strongest models are on Pro and Team; on Personal, Forge uses the first model in the list that the plan includes.
| Job | Model | Thinking | Why |
|---|---|---|---|
| Audits and re-audits | GPT-6 Sol then Claude Opus 5.5 (Pro), Claude Sonnet 5 | Medium | Evidence-backed findings at about a third of the strongest model's cost, with strong core findings in tests; on Pro, pick Claude Opus 5.5 for the most thorough scan. |
| Plans: Light | GPT-6 Luna then Claude Haiku 4.5, Claude Sonnet 5 | Low | One call with no lookups, written from the context Forge hands it: a small, fast model wrote equivalent plans in tests. |
| Plans: Standard | GPT-6 Sol then Claude Opus 5.5 (Pro), Claude Sonnet 5 | Medium | A few lookups and a careful plan: in tests it took the same approach as the strongest model, more briefly, at a fraction of the cost. |
| Plans: Deep investigation | Claude Opus 5.5 (Pro) then GPT-6 Astra (Pro) | High | Explores the repository like a scan before planning: the strongest model, thinking more. |
| Manager questions | GPT-6 Luna then Claude Sonnet 5 | Low | Answers from Forge's own findings and plans with a few lookups: a small, fast model answered correctly in tests, a little less detailed. |
| Analyze (progress) | Claude Opus 5.5 (Pro) then GPT-6 Astra (Pro), Claude Sonnet 5, GPT-6 Sol | Medium | Records what's built across the whole repository: smaller models claimed things were built without reading the files, so Pro runs it on the strongest model and Personal on Claude Sonnet 5. |
| Ask by voice | GPT-6 Luna then Claude Sonnet 5 | Low | Short spoken answers where speed matters, sharing the manager's conversations: the manager's model, few lookups. |
| Scan cost estimates | Claude Haiku 4.5 | Low | Reads the file list and manifests (no source code) to say what each scan would look at: one call on a fast, low-cost model, the same for everyone. |
Every model
Cost shows how each model compares with the others. Before a job starts, Forge shows its estimate in credits and the most it can cost; see Pricing.
| Model | Best for | Thinking | Context | Cost |
|---|---|---|---|---|
Claude Opus 5.5 AnthropicpremiumPro | The most thorough scans, deep plans and analysis: reading a whole codebase and backing every finding with evidence. Not ideal: Quick questions, where a mid-size model answers as well for half the price. | Low–Max | 1M | High |
Claude Opus 5 AnthropicpremiumPro | The previous strongest Claude: deep work when Opus 5.5 isn't available. Not ideal: New work: Opus 5.5 costs less and thinks more at the same level. | Low–Max | 1M | High |
Claude Sonnet 5 Anthropicstandard | Questions, voice and Light plans on Claude: good answers from context at half the strongest model's price. Not ideal: Scans and deep plans, where reading more of the code before concluding matters. | Low–Max | 1M | Medium |
Claude Haiku 4.5 Anthropicfast | Fast, low-cost answers to simple questions. Not ideal: Jobs that need careful reasoning: it has no thinking setting. | None | 200k | Low |
GPT-6 Astra OpenAIpremiumPro | OpenAI's strongest: scans and deep plans when Claude isn't available. Not ideal: Everyday questions: the most expensive model Forge offers. | Low–Max | 1.05M | High |
GPT-6 Sol OpenAIstandard | Forge's pick for scans and Standard plans: evidence-backed findings at about a third of the strongest model's cost. Not ideal: The deepest scans, where a premium model reads more before concluding. | None–Max | 1.05M | Medium |
GPT-6 Luna OpenAIfast | Forge's pick for questions, voice and Light plans: the lowest-cost model, good at answers from the context Forge gives it. Not ideal: Scans and plans, which need a stronger model. | None–Max | 1.05M | Lowest |
Anthropic and OpenAI process requests under their API terms and don’t use API data to train models; requests to OpenAI are sent with storage switched off. See Privacy.