What's New ✨

The race at the frontier has accelerated dramatically this week with major releases from Meta Superintelligence Labs and Mistral AI, alongside expanded general availability for Google:

• Meta Muse Wonder & Muse Wonder Full (Oct 12): Meta's flagship reasoning release featuring Active Memory and Active Focus for long-horizon agent workflows, leading SWE-Marathon benchmarks with a 57.5 score and an Intelligence Index of 56.
• Mistral Large 4 “le Chonk” (Oct 6 Preview): Mistral AI's 1.05T-parameter European MoE flagship (52B active per token) with native 1.6B multimodal vision, 1M context, and top-tier cybersecurity/terminal performance (93% on Cybench).
• Google Gemini 4 Argon Broad Availability (Oct 13): Formally expanded beyond cyber defenders to Google AI Pro/Ultra subscribers and developers on the standard Gemini API.
• High-Speed Agent Lineup (Oct 7): Anthropic's Claude Haiku 5.5 (160 tk/s) and OpenAI's GPT-6 Instant (150 tk/s) are now tracked for ultra-fast, budget-conscious production agent deployments.

Explore full token speeds, benchmark breakdowns, and side-by-side pricing on the Models page and Compare page.

We've just added several major new models to the leaderboard from late September and early October 2026. This has been a huge month for reasoning models and lightweight agents!

• Gemini 4 Argon: Added Google's new frontier reasoning model optimized for cybersecurity with a 1-million-token context window.
• Claude Sonnet 5.5: Anthropic's perfectly balanced model for speed and intelligence is now fully scored.
• GPT-6.1 Sol: OpenAI's new high-performance coding model, delivering incredible benchmark scores at a lower cost per task.
• Kolibri 78B: We've welcomed Aleph Alpha to the platform as a provider, tracking their brand new 78B Mixture-of-Experts transformer.

Head over to the Compare page to see how they stack up!

Welcome to the new and improved AI Leaderboard! We have completely overhauled our data schema and frontend design to offer a much cleaner, responsive, and robust platform.

The new table supports triple-tap sorting, mobile responsiveness, new input/output modality tracking, and distinct effort-level tracking for reasoning models like GPT-6 Astra and Claude Opus 5.5.