(../_next/static/css/06308a86deed49fb.css@dpl=dpl_CB5DWiSBAaR8o57bBpCvALr7h6Xu.css)  (../_next/static/css/2dfabeb5d8d508d9.css@dpl=dpl_CB5DWiSBAaR8o57bBpCvALr7h6Xu.css)  (../_next/static/chunks/webpack-0d986104732a17f5.js@dpl=dpl_CB5DWiSBAaR8o57bBpCvALr7h6Xu)  (https://www.googletagmanager.com/gtag/js?id=G-MLWHF2QR9E)   NanoGPT Speedrun Frontier               (../favicon.ico@e708dcf1b77b3a3b)  (../../cdn.prod.website-files.com/66239f0441b09824acb92c75/671b6f3bbf6572d40beb8cf2_Positive.jpg)  (https://cdn.prod.website-files.com/66239f0441b09824acb92c75/671b6f6054736f752438334f_Positive.jpg)      (https://www.primeintellect.ai/) (PRIME Intellect)   (https://www.primeintellect.ai/#lab) TRAINING 01  (https://www.primeintellect.ai/#inference) INFERENCE 02  (https://www.primeintellect.ai/#compute) COMPUTE 03  (https://www.primeintellect.ai/#research) RESEARCH 04   (https://docs.primeintellect.ai/introduction) DOCS  (https://www.primeintellect.ai/blog) BLOG  (https://jobs.ashbyhq.com/PrimeIntellect) CAREERS 24  (https://www.primeintellect.ai/contact) Book a call  (https://app.primeintellect.ai) Login (https://app.primeintellect.ai/dashboard/home/quickstart) Start training                        (LIGHT mode)            (DARK mode)         NanoGPT Speedrun Frontier We ran 153 autonomous runs across 18 frontier models on the nanoGPT optimizer speedrun. (https://www.primeintellect.ai/blog/measuring-autonomous-research) Read Blog   Record over time Record over tokens    Share of the human record gap closed human record100%  0%   20%   40%   60%   80%   100%   0d   1d   2d   4d   7d   9d                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           Fable 5 Opus 5 Kimi K3 · prime-agent Kimi K3 Opus 4.8 GPT-5.6 Sol GPT-5.6 Sol Pro Sonnet 5 GPT-5.6 Luna Grok 4.5 Qwen3.8 Max GLM 5.2 DeepSeek V4 Pro GPT-5.6 Terra Grok 4.6 Muse Spark 1.2 Muse Spark 1.1 GPT-5.5 Kimi K2.7  Agent time (days)    Share of the human record gap closed  0%   20%   40%   60%   80%   100%   0d   1d   2d   4d   7d   9d                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                           Fable 5 Opus 5 Kimi K3 · prime-agent Kimi K3 Opus 4.8 GPT-5.6 Sol GPT-5.6 Sol Pro Sonnet 5 GPT-5.6 Luna Grok 4.5 Qwen3.8 Max GLM 5.2 DeepSeek V4 Pro GPT-5.6 Terra Grok 4.6 Muse Spark 1.2 Muse Spark 1.1 GPT-5.5 Kimi K2.7  Agent time (days)      Collapseall models  All models Best validated result for each model  All trajectories     Log scale         Filter by model         1  Fable 5 note  2,726 81.7% closed   claude-code· high @24H3,010 8.7d Open   2  Opus 5 serial era  2,920 53.6% closed   claude-code· max @24H3,045 2.9d Open   3  Kimi K3 serial era  2,930 52.2% closed   prime-agent· max @24H3,125 3.6d Open   4  Kimi K3 note  2,974 45.8% closed   kimi-code· max @24H3,135 5.1d Open   5  Opus 4.8 3,018 39.4% closed   claude-code· max @24H3,180 3.0d Open   6  GPT-5.6 Sol note  3,042 35.9% closed   codex· xhigh @24H3,160 6.1d Open   7  GPT-5.6 Sol Pro serial era  3,058 33.6% closed   codex· xhigh @24H3,100 3.4d Open   8  Sonnet 5 note  3,105 26.8% closed   claude-code· max @24H3,120 2.0d Open   9  GPT-5.6 Luna note  3,110 26.1% closed   codex· xhigh @24H3,170 1.9d Open   10  Grok 4.5 note  3,120 24.6% closed   grok-cli· xhigh @24H3,160 2.7d Open   11  Qwen3.8 Max running  3,120 24.6% closed   qwen-code· max @24H3,225 1.9d Open   12  GLM 5.2 3,150 20.3% closed   pi· high @24H3,200 1.8d Open   13  DeepSeek V4 Pro running  3,205 12.3% closed   claude-code· max @24H3,205 1.1d Open   14  GPT-5.6 Terra serial era  3,214 11.0% closed   codex· xhigh @24H3,214 1.1d Open   15  Grok 4.6 running  3,220 10.1% closed   grok-cli· xhigh 0.6d Open   16  Muse Spark 1.2 running  3,230 8.7% closed   muse-code· xhigh 0.6d   17  Muse Spark 1.1 3,232 8.4% closed   pi· max @24H3,240 3.7d Open   18  GPT-5.5 serial era  3,234 8.1% closed   codex· xhigh @24H3,234 1.1d Open   19  Kimi K2.7 note  3,240 7.2% closed   kimi-code· max @24H3,240 1.6d Open   20  GLM 5.3 running  — no record   claude-code· xhigh                Model Harness ↑ Record   Gap Closed   @24H   Total Tok   Output Tok   Exps   Calls   Days   Traces    Fable 5  claude-code· high note   2,726 81.7% 3,010 800M 1.1M 811 3k 8.7 Open    Opus 5  claude-code· max serial era   2,920 53.6% 3,045 183M 690k 292 401 2.9 Open    Kimi K3  prime-agent· max serial era   2,930 52.2% 3,125 112M 2.2M — 488 3.6 Open    Kimi K3  kimi-code· max note   2,974 45.8% 3,135 682M 1.4M 713 4k 5.1 Open    Opus 4.8  claude-code· max  3,018 39.4% 3,180 318M 2.3M 427 2k 3.0 Open    GPT-5.6 Sol  codex· xhigh note   3,042 35.9% 3,160 2.9B 2.2M 963 28k 6.1 Open    GPT-5.6 Sol Pro  codex· xhigh serial era   3,058 33.6% 3,100 1.2B 4.6M 509 7k 3.4 Open    Sonnet 5  claude-code· max note   3,105 26.8% 3,120 998M 2.1M 213 2k 2.0 Open    GPT-5.6 Luna  codex· xhigh note   3,110 26.1% 3,170 894M 888k 362 12k 1.9 Open    Grok 4.5  grok-cli· xhigh note   3,120 24.6% 3,160 46M 385k 399 4k 2.7 Open    Qwen3.8 Max  qwen-code· max running   3,120 24.6% 3,225 216M 629k 312 866 1.9 Open    GLM 5.2  pi· high  3,150 20.3% 3,200 57M 1.7M 194 1k 1.8 Open    DeepSeek V4 Pro  claude-code· max running   3,205 12.3% 3,205 26M 319k 189 309 1.1 Open    GPT-5.6 Terra  codex· xhigh serial era   3,214 11.0% 3,214 417M 298k 154 3k 1.1 Open    Grok 4.6  grok-cli· xhigh running   3,220 10.1% — 27M 346k 97 691 0.6 Open    Muse Spark 1.2  muse-code· xhigh running   3,230 8.7% — 41M 910k 56 724 0.6 —    Muse Spark 1.1  pi· max  3,232 8.4% 3,240 122M 1.6M 489 2k 3.7 Open    GPT-5.5  codex· xhigh serial era   3,234 8.1% 3,234 70M 77k 185 614 1.1 Open    Kimi K2.7  kimi-code· max note   3,240 7.2% 3,240 160M 763k 187 3k 1.6 Open    GLM 5.3  claude-code· xhigh running   — — — — — — — — —          Compare Results Traces  Equal-budget comparison Give each model's best final run the same resource budget and compare the best validated record it reached within that budget. Agent-hours Experiments Output tokens  6h Budget of 24 hours 9d    (24)     Model Record Human 2,600 Baseline 3,290    Fable 5    3,010 (Fable 5: 3,010 steps within 24 hours)      Opus 5    3,045 (Opus 5: 3,045 steps within 24 hours)      GPT-5.6 Sol Pro    3,100 (GPT-5.6 Sol Pro: 3,100 steps within 24 hours)      Sonnet 5    3,120 (Sonnet 5: 3,120 steps within 24 hours)      Kimi K3 · prime-agent    3,125 (Kimi K3 · prime-agent: 3,125 steps within 24 hours)      Kimi K3    3,135 (Kimi K3: 3,135 steps within 24 hours)      GPT-5.6 Sol    3,160 (GPT-5.6 Sol: 3,160 steps within 24 hours)      Grok 4.5    3,160 (Grok 4.5: 3,160 steps within 24 hours)      GPT-5.6 Luna    3,170 (GPT-5.6 Luna: 3,170 steps within 24 hours)      Opus 4.8    3,180 (Opus 4.8: 3,180 steps within 24 hours)      GLM 5.2    3,200 (GLM 5.2: 3,200 steps within 24 hours)      DeepSeek V4 Pro    3,205 (DeepSeek V4 Pro: 3,205 steps within 24 hours)      GPT-5.6 Terra    3,214 (GPT-5.6 Terra: 3,214 steps within 24 hours)      Grok 4.6    3,220 (Grok 4.6: 3,220 steps; run ended at 13.7 hours before the selected 24 hours budget)      Qwen3.8 Max    3,225 (Qwen3.8 Max: 3,225 steps within 24 hours)      Muse Spark 1.2    3,230 (Muse Spark 1.2: 3,230 steps; run ended at 13.8 hours before the selected 24 hours budget)      GPT-5.5    3,234 (GPT-5.5: 3,234 steps within 24 hours)      Muse Spark 1.1    3,240 (Muse Spark 1.1: 3,240 steps within 24 hours)      Kimi K2.7    3,240 (Kimi K2.7: 3,240 steps within 24 hours)      GLM 5.3    — (GLM 5.3: no validated record within 24 hours)      Gray runs ended before the selected budget     Loading available trajectories…    Open Traces to explore 41 curated full agent trajectories, including tool calls, subagents, and scratchpads.       (https://www.primeintellect.ai/) (Prime Intellect)    Be your own Lab.  (https://app.primeintellect.ai/dashboard/home/quickstart) Start training       (https://www.primeintellect.ai/contact) Book a call    Platform (https://www.primeintellect.ai/#lab) Lab (https://www.primeintellect.ai/#compute) Compute (https://www.primeintellect.ai/#research) Research   Company (https://jobs.ashbyhq.com/PrimeIntellect) Careers24   (https://primeintellect.supply) Merch (https://www.primeintellect.ai/contact) Contact   Community (https://x.com/PrimeIntellect) X (https://www.linkedin.com/company/primeintellect-ai) LinkedIn (https://discord.gg/primeintellect) Discord (https://luma.com/primeintellect) Luma    Resources (https://docs.primeintellect.ai/introduction) Docs (https://www.primeintellect.ai/blog) Writings (https://luma.com/primeintellect) Events (https://primeintellect.supply) Merch (https://status.primeintellect.ai) Platform Status   Terms (https://www.primeintellect.ai/terms-of-service) Terms of Service (https://www.primeintellect.ai/privacy-policy) Privacy Policy (https://www.primeintellect.ai/security) Security Policy      © 2026Prime Intellect, Inc.