Models
Every model id you can call, its context window, and what it costs per million tokens.
One key reaches all of these. Use the API id column verbatim in the model field of a request — display names are not accepted.
| Model | API id | Context | $/M in | $/M out | Saving |
|---|---|---|---|---|---|
| Claude Opus 4.8 | claude-opus-4-8 | 200K | $3 | $15 | 80% |
| Claude Fable 5 | claude-fable-5 | 200K / 1M | $3 | $15 | 80% |
| Claude Sonnet 4.6 | claude-sonnet-4-6 | 200K / 1M | $0.9 | $4.5 | 70% |
| Claude Haiku 4.5 | claude-haiku-4-5 | 200K | $0.4 | $2 | 60% |
| GPT-5.5 | gpt-5.5 | 256K | $3 | $12 | 70% |
| GPT-5 | gpt-5 | 256K | $1.5 | $6 | 70% |
| Gemini 3 Pro | gemini-3-pro | 1M | $1 | $6 | 60% |
| Gemini 3 Flash | gemini-3-flash | 1M | $0.2 | $1.2 | 60% |
| MiniMax M3 | minimax-m3 | 1M | $0.5 | $2.5 | 58% |
Prices are per million tokens in credits. Live ids are always available from the models endpoint:
curl https://aiprimetech.io/v1/models -H "Authorization: Bearer $CLAUDEAPIKEY"Choosing a model
| Task | Recommended | Why |
|---|---|---|
| Agentic coding, refactors, architecture | claude-opus-4-8 | Strongest multi-step reasoning; worth the output price on hard work |
| Everyday coding, chat, general work | claude-sonnet-4-6 | The default — roughly a fifth of Opus's rate at close quality on most tasks |
| Classification, extraction, routing | claude-haiku-4-5 | Cheapest Claude; these tasks do not need a frontier model |
| Very large documents | gemini-3-pro | 1M context at a low input rate |
| High-volume batch | gemini-3-flash or minimax-m3 | Lowest per-token cost in the catalog |
Long context
claude-sonnet-4-6 also serves a 1M-token variant, billed at a higher rate than the 200K version because long context is more expensive upstream. Gemini models take 1M natively. Sending 500K tokens of context on every turn is rarely the cheapest way to solve a problem — see Context management.
Model aliases
Ids are pinned, not floating. claude-sonnet-4-6 always means that version — it will not silently become a different model under you. When we add a successor it gets a new id, and the old one keeps working until formally retired via the changelog.