Model & API Radar
Verified 2026-08-04. Prices are USD per million standard text tokens unless noted. Provider pages, previews, caching, batch modes, tool calls, subscriptions, and regional channels can change the effective cost. Recheck official sources before committing money or production traffic.
Personal Default
For interactive coding where mistakes are expensive, the current default is GPT-5.6 Sol with high reasoning. This records a working preference, not a universal benchmark winner. Do not downgrade a high-consequence task merely because a token table looks cheaper.
Route only after the quality bar is explicit:
- Quality-first coding and review: GPT-5.6 Sol / high.
- Routine implementation with good checks: trial GPT-5.6 Terra against the same tasks.
- Bulk classification, translation, and extraction: trial Luna or a low-cost Flash model with deterministic validation.
- Important disagreement or provider-specific work: cross-check with a strong Claude or Gemini model.
Representative API Snapshot
| Provider | Model | Input | Output | Positioning |
|---|---|---|---|---|
| OpenAI | gpt-5.6-sol | $5.00 | $30.00 | flagship complex reasoning and coding |
| OpenAI | gpt-5.6-terra | $2.00 | $12.00 | balance of intelligence and cost |
| OpenAI | gpt-5.6-luna | $0.20 | $1.20 | cost-sensitive volume |
| Anthropic | claude-fable-5 | $10.00 | $50.00 | most capable, long-running agents |
| Anthropic | claude-opus-5 | $5.00 | $25.00 | complex agentic coding |
| Anthropic | claude-sonnet-5 | $3.00 | $15.00 | speed/intelligence balance |
gemini-3.6-flash | $1.50 | $7.50 | fast frontier model with grounding | |
gemini-3.5-flash-lite | $0.30 | $2.50 | high-volume agentic and simple processing |
Anthropic lists introductory Sonnet 5 pricing of 10 through 2026-08-31. That temporary price is not used as the durable comparison baseline.
How to Compare Price-Performance
Do not rank models from price and public benchmarks alone. Replay a small personal evaluation set and record:
- task success and severity of mistakes;
- amount of human correction;
- wall-clock latency;
- input, cached input, output, and reasoning tokens;
- tool calls, retries, and failed runs;
- total cost per accepted result.
The cheapest successful run is useful. The cheapest attempted run is not.
Update Rule
On review, replace the table and routing decision in place. Add a changelog entry only when the change affects the personal default, a major workflow, or expected cost by a meaningful amount.