AI MODEL RELEASES

Successor to GLM 5.1. AA Intelligence Index 50.7, GPQA 89.5, Terminal-Bench 50.8 — open-weight reasoning gains over 5.1.

First public Mythos-class model — a tier above Opus. Highest AA Intelligence Index ever (64.9), Arena Elo 1932, SWE-bench Pro 80.3. Stripe compressed 2+ months of work into 1 day on a 50M-line codebase.

Frontier-tier successor to M2.7. AA Intelligence Index 54.7, GPQA 92.9, Arena 1668 — continues the open-weight cost-disruption pattern.

New frontier leader. SWE-bench Pro 69.2% tops GPT-5.5 and Gemini 3.1 Pro; ~4x less likely to let code flaws pass unremarked. Same pricing as 4.7, fast mode 2.5x speed.

Flash-tier model with near-frontier reasoning: GPQA 92.2, ARC-AGI-1 96%. AA Intelligence Index 55.3 — top of the cost-efficient tier.

Frontier-tier Qwen flagship. AA Intelligence Index 56.6, GPQA 92.3, Terminal-Bench 50.8 — closes the gap to Western frontier at lower cost.

Flagship successor to V3 (Max Effort reasoning). Open-weight, competitive with closed frontier at a fraction of the cost. Awaiting Multimodal eval coverage.

Frontier reasoning model. #1 on AA Intelligence Index (60.24), tops Arena Elo at 1781, crosses 90% on ARC-AGI-2, and leads Terminal-Bench 2.0 at 82.7%.

Successor to K2.5 with stronger reasoning and Arena Elo; trades some agentic score for benchmark gains

Top-tier Qwen variant above Plus, highest Arena Elo in the family at 1511

SuperGrok Heavy beta with native video understanding. AA now tracking with full per-benchmark coverage; #2 globally on initial AA pass.

1M context, stronger coding and vision. SWE-bench Pro 64.3% beats GPT-5.4 and Gemini 3.1 Pro.

Meta's first proprietary model from Superintelligence Labs, non-open-weight

Apache 2.0 dense 31B from the Gemma 4 family. Native vision + audio, 140+ languages. Best open-weight from Google Q2 2026.

Agentic coding focus, 1M context, strong multimodal performance

Two-version jump from M2.1 (skipping intermediate M2.5). AA Intelligence Index 49.6 — top-tier Chinese open-weight reasoning at half the cost of Western frontier.

First mainline model incorporating GPT-5.3-Codex coding capabilities. Native computer use, 1M context, surpasses human baseline on OSWorld. Codex branch ends here.

Highest GPQA Diamond score ever at 94.3%, doubled ARC-AGI-2 to 77.1%

Opus-level performance at one-fifth the cost, default model on claude.ai

Four specialized agents debate before answering, 65% hallucination reduction

Open-weight 397B MoE with visual agentic capabilities, 201 languages

Final specialized Codex release before coding was folded into mainline GPT-5.4. Peak on LiveCodeBench and Terminal-Bench.

1M context with parallel agent teams, #1 on Arena Elo at 1504

Top-performing Chinese open model, strong coding + reasoning

1T MoE with 100-agent swarm coordination, 99% HumanEval

Reasoning-focused flagship, competitive with frontier Western models

Flash-tier model outperforming previous-gen Pro on most benchmarks

First model to score 100% on AIME 2025 and 80% on SWE-bench

Open-weight coding model, 72.2% SWE-bench Verified at 7x lower cost than Sonnet

Major leap for Mistral, closed the gap with US frontier labs

Matched GPT-5 class at a fraction of the cost, open-weight MoE

First model to break 80% on SWE-bench Verified

First model to break 1500 Arena Elo, 100% on AIME with code execution

#1 on Arena thinking mode, 1M context, strong agentic coding

Incremental upgrade with improved reliability and instruction following

Haiku-tier model matching Sonnet 4 on coding at one-third the cost

First Sonnet to score 100% on AIME, closed the Opus gap entirely

Unified reasoning and chat, 400K context, 94.6% on AIME 2025

Improved agentic reliability with better tool use and planning

Open-weight MoE that matched frontier closed models on coding

Deep reasoning model, strongest on abstract math at launch

First Flash model with thinking, near-Pro performance at one-eighth the price

Strongest agentic model at launch, sustained multi-hour autonomous coding

Matched Opus 4 on coding at one-fifth the price, 1M context window

Hybrid reasoning MoE trained on 36T tokens across 119 languages

Full reasoning model with tool use, first to break 80% on GPQA Diamond

First 10M token context window, 17B active params from 109B MoE

First open MoE from Meta, natively multimodal with 1M context

Built-in thinking mode, native audio and video understanding

First hybrid reasoning model, extended thinking mode for complex problems

Matched frontier labs on reasoning, trained on 200K H100 cluster

Reasoning at 75% lower cost than o1, made chain-of-thought economically viable

Open-weight reasoning model that triggered the DeepSeek market shock

Trained for $5.5M, proved frontier performance was possible at low cost

First reasoning model, uses chain-of-thought at inference time to solve hard problems

Near-frontier performance at flash pricing, native tool use and code execution

405B-class performance distilled into 70B parameters

Best coding model of 2024, dominated SWE-bench for months

Trained on Colossus, xAI's 100K GPU cluster, first competitive Grok

European frontier model with strong multilingual and code performance

Largest open-weight model at 405B parameters, GPT-4 class performance

Natively multimodal with voice, 2x faster and 50% cheaper than GPT-4 Turbo

Open-weight model competitive with GPT-4 class, massive fine-tuning ecosystem

First Claude to match GPT-4, introduced the Opus/Sonnet/Haiku tier system

First 1M token context window, processed entire codebases in one pass

Google's first natively multimodal model, launched the Gemini brand

Open-weight model that kickstarted the open-source LLM ecosystem

First Claude with 100K context, established Anthropic as a frontier lab

Powered Bard and Google Workspace AI, strong multilingual performance

First multimodal GPT, passed the bar exam, defined the frontier for a year

Launched ChatGPT, fastest consumer product to 100M users in history

Largest dense model at launch, first to show chain-of-thought reasoning at scale

Proved most LLMs were undertrained, reshaped scaling strategy industry-wide

First large language model to demonstrate emergent few-shot learning