AI Reasoning Model Comparison

This is mostly a curated data-lookup tool — the comparison table, the “pick by use case” helper and the filters all read from one hand-verified dataset of reasoning models. The only computation is the optional per-task cost estimate, which has two parts.

Visible cost (exact). Every provider publishes a price per million input and output tokens. The visible cost of a request is:

cost = input ÷ 1,000,000 × price_in + output ÷ 1,000,000 × price_out

Hidden reasoning cost (estimate).Reasoning models also generate hidden “thinking” tokens. OpenAI, Anthropic and Google all bill these at the outputtoken rate (cited in each vendor's reasoning guide below), so they can quietly multiply the bill. Hidden tokens can't be predicted exactly without running the model, so we estimate them with a transparent heuristic: a base budget of 2,000 tokens scaled by an effort multiplier — low ×0.25, medium ×1.0, high ×2.5. Models that always reason and can't be metered (DeepSeek-R1, Grok 4) use a documented typical budget instead. The estimated total is:

total = input ÷ 1e6 × price_in + (output + reasoning) ÷ 1e6 × price_out

Sorting and filtering are pure operations on the stored rows — a stable sort on the chosen column (price, context, AIME, GPQA or SWE-bench) and boolean predicate matching for the provider, controllable-effort and open-weight filters. Benchmark columns are verbatim vendor-reported figures with a source link per model; we never recompute them, and missing figures sink to the bottom rather than counting as zero. No request leaves your browser — the whole dataset ships with the page, last verified 2026-06-27.

• OpenAI — Reasoning guide (reasoning_effort, reasoning-token billing)
• OpenAI — API pricing
• Anthropic — Extended thinking docs
• Anthropic — API pricing
• Google — Gemini thinking docs
• Google — Gemini API pricing
• DeepSeek — API pricing & R1 model card
• xAI — Grok models & pricing
• Alibaba Cloud — Qwen (Model Studio) models & pricing

Prices and benchmark scores were last cross-checked against these sources on 2026-06-27 (June 2026 vendor snapshot). Benchmark figures are vendor-reported and not independently re-run. The page is reviewed on each major reasoning-model release.