NanoGPT Speedrun FrontierWe ran 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun.Read BlogShare of the human record gap closedhuman record 100%0%20%40%60%80%100%0d1d2d4d7d9dFable 5Opus 5Kimi K3 · prime-agentKimi K3Opus 4.8GPT-5.6 SolGPT-5.6 Sol ProSonnet 5GPT-5.6 LunaGrok 4.5Qwen3.8 MaxGLM 5.2DeepSeek V4 ProGPT-5.6 TerraGrok 4.6Muse Spark 1.2Muse Spark 1.1GPT-5.5Kimi K2.7Agent time (days)Share of the human record gap closed0%20%40%60%80%100%0d1d2d4d7d9dFable 5Opus 5Kimi K3 · prime-agentKimi K3Opus 4.8GPT-5.6 SolGPT-5.6 Sol ProSonnet 5GPT-5.6 LunaGrok 4.5Qwen3.8 MaxGLM 5.2DeepSeek V4 ProGPT-5.6 TerraGrok 4.6Muse Spark 1.2Muse Spark 1.1GPT-5.5Kimi K2.7Agent time (days)All modelsBest validated result for each model1Fable 52,72681.7% closedclaude-code · high@24H 3,0108.7d2Opus 52,92053.6% closedclaude-code · max@24H 3,0452.9d3Kimi K32,93052.2% closedprime-agent · max@24H 3,1253.6d4Kimi K32,97445.8% closedkimi-code · max@24H 3,1355.1d5Opus 4.83,01839.4% closedclaude-code · max@24H 3,1803.0d6GPT-5.6 Sol3,04235.9% closedcodex · xhigh@24H 3,1606.1d7GPT-5.6 Sol Pro3,05833.6% closedcodex · xhigh@24H 3,1003.4d8Sonnet 53,10526.8% closedclaude-code · max@24H 3,1202.0d9GPT-5.6 Luna3,11026.1% closedcodex · xhigh@24H 3,1701.9d10Grok 4.53,12024.6% closedgrok-cli · xhigh@24H 3,1602.7d11Qwen3.8 Max3,12024.6% closedqwen-code · max@24H 3,2251.9d12GLM 5.23,15020.3% closedpi · high@24H 3,2001.8d13DeepSeek V4 Pro3,20512.3% closedclaude-code · max@24H 3,2051.1d14GPT-5.6 Terra3,21411.0% closedcodex · xhigh@24H 3,2141.1d15Grok 4.63,22010.1% closedgrok-cli · xhigh0.6d16Muse Spark 1.23,2308.7% closedmuse-code · xhigh0.6d17Muse Spark 1.13,2328.4% closedpi · max@24H 3,2403.7d18GPT-5.53,2348.1% closedcodex · xhigh@24H 3,2341.1d19Kimi K2.73,2407.2% closedkimi-code · max@24H 3,2401.6d20GLM 5.3—no recordclaude-code · xhighModelHarnessTracesFable 5claude-code · high2,72681.7%3,010800M1.1M8113k8.7Opus 5claude-code · max2,92053.6%3,045183M690k2924012.9Kimi K3prime-agent · max2,93052.2%3,125112M2.2M—4883.6Kimi K3kimi-code · max2,97445.8%3,135682M1.4M7134k5.1Opus 4.8claude-code · max3,01839.4%3,180318M2.3M4272k3.0GPT-5.6 Solcodex · xhigh3,04235.9%3,1602.9B2.2M96328k6.1GPT-5.6 Sol Procodex · xhigh3,05833.6%3,1001.2B4.6M5097k3.4Sonnet 5claude-code · max3,10526.8%3,120998M2.1M2132k2.0GPT-5.6 Lunacodex · xhigh3,11026.1%3,170894M888k36212k1.9Grok 4.5grok-cli · xhigh3,12024.6%3,16046M385k3994k2.7Qwen3.8 Maxqwen-code · max3,12024.6%3,225216M629k3128661.9GLM 5.2pi · high3,15020.3%3,20057M1.7M1941k1.8DeepSeek V4 Proclaude-code · max3,20512.3%3,20526M319k1893091.1GPT-5.6 Terracodex · xhigh3,21411.0%3,214417M298k1543k1.1Grok 4.6grok-cli · xhigh3,22010.1%—27M346k976910.6Muse Spark 1.2muse-code · xhigh3,2308.7%—41M910k567240.6—Muse Spark 1.1pi · max3,2328.4%3,240122M1.6M4892k3.7GPT-5.5codex · xhigh3,2348.1%3,23470M77k1856141.1Kimi K2.7kimi-code · max3,2407.2%3,240160M763k1873k1.6GLM 5.3claude-code · xhigh—————————Equal-budget comparisonGive each model's best final run the same resource budget and compare the best validated record it reached within that budget.6hBudget of 24 hours9dModelRecordHuman 2,600Baseline 3,290Gray runs ended before the selected budgetOpen Traces to explore 41 curated full agent trajectories, including tool calls, subagents, and scratchpads.