AI reasoning models leaderboard

The 90% threshold

By mid-2026, five models in this table clear 90% on provider-published GPQA Diamond results: Gemini 3.1 Pro, Claude Opus 4.7, Claude Opus 4.8, Kimi K2.6, and DeepSeek V4-Pro. Claude Sonnet 4.6 sits just below at 89.9%.

Scores clustered within a few points are difficult to turn into a product decision. First decide whether GPQA resembles your task, then test the candidates near your required threshold on held-out work and compare cost, latency, and failure type.

What GPQA can and cannot indicate

GPQA Diamond contains difficult questions in biology, chemistry, and physics. It is relevant when a workload requires similar scientific reasoning, but it does not directly measure citation accuracy, tool use, financial modeling, medical safety, or performance on a team's private documents. Treat it as one screening signal, not a production guarantee.

For purely creative or conversational tasks, the benchmark is less relevant. A customer support bot doesn't need a 90% GPQA score. A clinical decision support tool probably does.

Methodology

The ranking uses provider-published GPQA Diamond scores only; where no provider figure exists, the table shows —. It covers the curated comparison set in the table rather than every record in models.json. Input prices come from provider rate cards.

SWE-bench cells are provider-published unless explicitly marked benchr editorial estimate. Those estimates are planning aids, not lab measurements; see each record's benchmarks_estimated field. Providers can select favorable benchmark results, so evaluate the models on your own task distribution before making an infrastructure decision from rank alone.

Using the threshold in real products

The useful cutoff is the lowest-cost model that clears a threshold defined on your work. A technical research assistant may benefit from testing the frontier cluster; a routing classifier needs a task-specific classification set, not a GPQA target.

Run your own threshold test by grouping failures: wrong final answer, missed constraint, fabricated citation, or incomplete chain of reasoning. If two models fail in the same way on your workload, choose the cheaper or faster one. If the higher-scoring model avoids a failure mode that creates legal, medical, financial, or engineering risk, the premium may be justified.

Failure cost matters more than rank

Reasoning benchmarks are most useful when the cost of a wrong answer is high. If the task is low-risk and easy to verify, a lower-ranked model may be the right choice. If the task involves expert review, regulatory exposure, or expensive downstream actions, the premium for stronger reasoning can be rational even when the benchmark gap looks small.

For evaluation, ask for explicit assumptions and a concise rationale, then compare both with the source material. The failure mode is not only a wrong final answer; it is a confident explanation built on a false premise.

Keep a separate holdout set for reasoning evaluations. If the same examples are reused in prompt tuning, the score stops measuring general reasoning and starts measuring adaptation to your test. A small unseen set of difficult internal cases is often more useful than another public benchmark.

For teams using reasoning models in sensitive contexts, keep a human-review policy tied to confidence and evidence. A leaderboard can identify candidates, but it cannot decide when an answer must be escalated. The escalation rule is part of the product, not a property of the model.