Reasoning models spend hidden tokens before answering. A low headline price can become expensive if the model spends thousands of reasoning tokens per request.
| # | Model | Provider | Input / 1M | Output / 1M | Context | Score | Profile |
|---|---|---|---|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | $2.000 | $12.000 | 1.0M | 0.61 | View | |
| 2 | Gemini 4 Argon | $4.000 | $20.000 | 1.0M | 0.52 | View | |
| 3 | GPT-6 Astra | OpenAI | $10.000 | $50.000 | 1.1M | 0.44 | View |
| 4 | Claude Fable 5.1 | Anthropic | $10.000 | $50.000 | 1.0M | 0.42 | View |
Composite score = 0.2×cost + 0.15×context + 0.45×intelligence + 0.2×speed, where each factor is normalized 0–1 across the active catalog. Cost is inverted (lower price = higher score). Weights reflect the priorities of this use case. Scores are a starting point — always test on your own workload.
They can be — hidden reasoning tokens are often billed as output tokens.
Compare intelligence per dollar on your own problem set, not just headline benchmarks.