LLM Reasoning Benchmark

Frontier and open-source models head-to-head on a pinned subset of the public MMLU-Pro reasoning set, re-run on demand as prices and models change. Accuracy is exact-match on the correct option letter; the winner is the cheapest model clearing the accuracy bar (70%). Cost is each model's published list price applied to the actual tokens used, so a cheap-per-token model that reasons at length can still cost more per correct answer.

run 2026-06-22 · run #1 · subset v1 (n=98)
ModelLabAccuracy $ / correct answerProj. $/mo @ 1,000,000Savings vs incumbentn scored
Qwen3 235B WINNERQwen80.6%$0.0001$64.0199%98
DeepSeek V4 Flash DeepSeek83.7%$0.0002$201.9095%98
Llama 4 Maverick Meta83.7%$0.0004$306.1393%98
GPT-5.4 Mini OpenAI77.6%$0.0008$621.8386%98
DeepSeek V4 Pro DeepSeek87.8%$0.0010$892.0879%98
Grok 4.3 xAI87.8%$0.0022$1,940.2755%98
Claude Haiku 4.5 Anthropic77.6%$0.0025$1,958.5455%98
GLM 5.2 Zhipu76.5%$0.0040$3,086.6429%98
Claude Sonnet 4.6 Anthropic90.8%$0.0051$4,631.69-7%98
Qwen3.7 Max Qwen91.8%$0.0065$5,942.08-35%97 (1 errors)
GLM 5.1 Zhipu73.2%$0.0101$7,404.82-69%97 (1 errors)
Kimi K2.6 Moonshot75.3%$0.0099$7,426.33-69%97 (1 errors)
Gemini 3.5 Flash Google86.7%$0.0117$10,133.23-133%98
GPT-5.5 OpenAI92.9%$0.0127$11,801.58-172%98
Gemini 3.1 Pro Google85.7%$0.0165$14,171.82-226%98
GLM 4.7 Flash Zhipu56.1%$0.0016$891.9779%98
Claude Opus 4.8 incumbentAnthropic89.8%$0.0048$4,345.82incumbent98

Per-question receipts

Run history

RunDateWinnerWinner $/mo
#12026-06-22qwen3-235b$64.01receipts

Re-run this yourself

Reproduce this exact ranking with the open-source redcrown CLI:

pip install redcrown
redcrown build-dataset mmlu-pro --ids-file subset-v1.txt --out exp.json
redcrown eval exp.json --concurrency 10

Methodology

Every run executes the same pinned MMLU-Pro subset through each candidate with the open-source redcrown CLI, routed through a single OpenRouter gateway so every model runs under identical conditions. Answers are scored by deterministic exact-match on the option letter, with no LLM judge. Cost is each model's published native list price (not the gateway markup) applied to the measured token usage; the monthly projection assumes 1,000,000 questions. Errors and truncations are counted and shown, never dropped. Caveat: frontier labs may have trained on MMLU-derived data, so absolute accuracy can be optimistic; the durable value here is the cost-per-correct ranking and how it shifts as prices and models change, and the subset can be rotated. Built and run by RedCrown, which sells no models and routes no traffic: the proof has no thumb on the scale.