AI Models Leaderboard
Updated 2026-08-16 · 78 models · 69 sourced benchmark figures
Compare 78 large language models on the benchmarks vendors actually report in 2026, with pricing, context window, and throughput. Every figure links to its source. Estimates and inherited numbers are greyed out and never ranked.
There is no single "best model" score here, on purpose. Labs publish only the benchmarks that flatter them — Anthropic reports HLE (frontier average ≈54) and skips GPQA Diamond (≈91); Google and Alibaba do the reverse. Averaging across whatever each lab chose to publish ranks publication strategy, not capability: when we tested it, models moved by up to six places depending on which benchmarks they happened to report. Rank on the benchmark that matches your workload, or see Artificial Analysis and LiveBench for independently-run composites.
| Model | Vendor | Src | Ctx | In $/M | Out $/M | GPQA | SWE-V | SWE-Pro | TB 2.1 | HLE |
|---|
Benchmark Leaders
Models within 1 point share a rank — sub-point gaps are below measurement noise. Only sourced figures appear.
Price vs Performance
X: blended $/M (log). Y: score on the benchmark selected at left. Pareto-optimal models are highlighted.
Cost Calculator
Track real spend per provider with Helicone or Langfuse. Need access to every model under one bill? OpenRouter routes to all of them.
How to read this leaderboard
Every number links to where it came from. A figure is only ranked if it comes from a vendor model card, a system card, or an open leaderboard. Numbers marked est are estimates, numbers marked ? were inherited without recorded provenance, and neither ever contributes to a ranking or a chart. The Src column counts how many currently-reported benchmarks a model has a citable source for — a low number means the vendor published little, not that the model is weak.
Ranks are assigned in tiers. Models within one point share a rank. SWE-bench reports to one decimal on a 500-problem set where a single problem is worth 0.2 points, so a 0.1 gap is noise. Ordering two models on that difference would imply a precision the benchmark does not have.
Never compare SWE-bench Verified against SWE-bench Pro. They are different benchmarks and Pro runs roughly 20 points lower on the same model — Claude Opus 4.8 scores 69.2 on Pro and 88.6 on Verified. They get separate columns here for that reason. The same applies to Terminal-Bench 2.0 versus 2.1.
The blended price column is a 3:1 input-to-output weighted average, which approximates the cost shape of most production agent workloads (heavy input from context and history, smaller output per call). Prompt caching changes this calculus substantially — Anthropic and OpenAI both ship aggressive caching that can drop effective price by 60-80% on repeated-context workloads. If your application is conversational or RAG-heavy, the cached-read price is the one you should optimise against, not the headline input price.
Which benchmark actually matters
SWE-bench Pro measures end-to-end issue resolution on real repositories, built to resist the memorisation that has begun to inflate SWE-bench Verified. Currently the widest-reported coding benchmark and the most predictive single number for AI coding tools.
Terminal-Bench 2.1 measures agentic work in a real terminal — multi-step tasks with tool use and recovery from failure. Now the headline agentic eval for most labs, and it replaced Terminal-Bench Hard and 2.0, which are not comparable to it.
GPQA Diamond measures graduate-level science reasoning on questions verified to be hard for non-experts. Still the best indicator of novel reasoning over retrieved facts, though frontier models now cluster above 90, so it is beginning to saturate.
SWE-bench Verified measures issue resolution on real GitHub repositories. Still widely quoted, but frontier scores now sit in the high 70s to 90s and the spread between good models has narrowed considerably.
HLE (Humanity's Last Exam) is the hardest benchmark in general use — frontier models score in the 40s and 50s. Check whether a quoted figure was run with tools, since that shifts results by roughly 5-15 points.
MMMU-Pro measures multimodal reasoning across images and text, and is a harder variant of the original MMMU. Only matters if your application uses vision.
Benchmarks we no longer rank
Several benchmarks that used to appear here have been retired or superseded, and are kept as greyed-out columns for historical reference only. MATH-500 and AIME were dropped from active reporting. τ²-Bench was replaced by τ³-Banking. Terminal-Bench Hard and 2.0 were replaced by 2.1. HumanEval is saturated — every frontier model scores 92-96, so it separates nothing. The Aider Polyglot leaderboard is unmaintained; its top entry is still GPT-5 and it tracks no model released in 2026. MMLU-Pro and MMMU remain in use but have been dropped from the major composite indices.
This churn is the main reason benchmark tables go stale. A table that keeps demanding numbers the industry stopped publishing will either sit empty or get filled with guesses.
Picking a model for production
Three rules of thumb after spending the past 12 months testing most models on this list in production-shaped workloads:
1. Start cheaper, upgrade only on observed regression. Most agent flows do not need a frontier model. We default to Claude Sonnet 4.6 or DeepSeek V3.1 and only move to Opus / GPT-5 / o3 when we see a quality cliff. The cost difference compounds fast at production volume.
2. Optimise for cached-read price, not headline price. If your workload has any repeated context (system prompts, conversation history, RAG corpora), the cached-read price column dominates monthly spend. Anthropic and OpenAI both offer caching; Google's flash variants have implicit caching as well.
3. Benchmark your own task before committing. Run the cheapest plausible candidate on 30 examples of your real task. The leaderboard tells you which models to consider; only your eval set tells you which model to ship.
Methodology and sources
Benchmark figures come from vendor model cards and system cards, and from SWE-bench. Each cell stores its own source URL and the date it was read; scores are frozen at that date rather than silently recomputed later. A build-time validator rejects any figure that lacks an attributable source, so an unsourced number cannot reach this page — and it rejects figures sourced to secondary summary sites, several of which publish benchmark scores that appear in no vendor release.
Vendor-reported figures are self-run and are labelled as such on hover. We cross-check our orderings against Artificial Analysis and LiveBench, both of which run their own independent evaluations and publish composite indices. Their numbers are not reproduced here — for an independently-measured overall ranking, go to the source.
Pricing is taken from each provider's first-party API on 2026-08-16; open-weights pricing reflects the typical OpenRouter / Together / Fireworks rate. Models are marked superseded when a newer model exists in the same series and unreleased when announced without public availability; neither is ranked by default.
For the long-form analysis behind these numbers — which models we use day-to-day, which we ruled out, and how the landscape shifted in the last six months — read the companion post LLM Benchmark Comparison 2026.