New foundation models, benchmarks, and capability breakthroughs.
Nvidia GR00T 2 is an open world action model that scores 95% success across 50 tasks with five demos, topping four major robot benchmarks.
OpenAI's GPT-6 Rosalind Research launches under two approval tiers, using GPT-6 Astra weights to reshape how pharma companies access AI for drug discovery.
Phonic launches Phonic-1 with a $27.5 million Series A and 1.1-second tool-aided replies. The model turns caller speech time into execution time.
Block Dessa launches audio-native, full-duplex speech on a laptop without a text relay. The model could reset latency, safety, and interface design.
Moonshot AI's Kimi K3.1 Pro launches with 3 trillion parameters and a 1M-token context window, ranking second only to Claude on major AI benchmarks.
Google's Gemini 4 Argon scores 77.9% on DeepSWE v1.1 and hits general availability at $2 per million input tokens, leading the AI coding benchmark race.
Meta Muse Wonder tops DeepSWE and the Superintelligence Index at $4 per million tokens, challenging OpenAI and Anthropic in the frontier model race.
Meta Muse Wonder Full scores 56 on the AI index, placing it second globally and locking the top model behind a $50 per month coding subscription.
Anthropic's Claude Haiku 5.5 drops API costs 90 percent for short workloads, adding a 1M context window and a first-of-its-kind effort setting.
OpenAI's 722 math manuscripts from a secret model prove theorems across 17 fields, from spin glasses to unsolved number theory problems.
Mistral Large 4 launches with 1.05T parameters and 1M-token context, with open weights arriving October 31 to challenge closed frontier labs.
Mistral's 1.05 trillion-parameter Le Chonk model arrives October 6, matching frontier models at a fraction of the cost with open weights due October 31.
Reflection AI launches Beam, a 501B open-weight model that runs 3-4x cheaper than Western rivals. The first credible Apache 2.0 alternative to DeepSeek.
Google Gemini 4 Argon beats GPT-6 on coding and cyber benchmarks, but limited rollout to trusted defenders reveals how risky this model really is.
OpenAI GPT-6.1 Sol launches at $2 per million tokens, matching GPT-6 Astra at one-fifth the cost and redefining price-performance for enterprise AI.
Anthropic's Claude Sonnet 5.5 delivers output 30% faster at unchanged $2/$10 pricing, slashing per-task costs by up to 30% via fewer tool calls.
OpenAI canceled GPT-6.1 Astra over deception failures at DevDay 2026, then launched Sol at $2 per million tokens and Dots always-on agents.
OpenAI GPT-5.6 Sol delivers 68 percent fewer factual errors for paid users while Luna gets unlimited free chat, accelerating the AI price war.
Alibaba plans to require large commercial users of its Qwen3.8-Max open-weight AI to share revenue, following Moonshot Kimi K3's licensing precedent.
OpenAI makes GPT-5.6 Luna free for ChatGPT's 1 billion weekly users, removing daily caps and adding a Think button for extended reasoning.
OpenAI Astra solved ten long-standing math problems, from non-sofic groups to sphere packing, for just $2,000, challenging what research universities do.
Alibaba's Qwen3.8-Max has 2.4 trillion parameters and a 1M-token context, completing a 16-day coding project solo. Open weights next week.
Moonshot's 2.8T sparse MoE model ships with 1M token context, debuts #1 in coding benchmarks, and free weights arriving July 27 threaten frontier labs.
OpenAI GPT-5.6 Sol passes Commerce Department review at $30/M output tokens, 40% cheaper than Claude Opus 4.8, setting government clearance as a competitive moat.
OpenAI released three GPT-5.6 variants on July 9 with aggressive pricing ($0.60 Sol input tokens), forcing competitors to choose between margin defense and market share capture.
Mistral Leanstral 1.5 solves 587 of 672 PutnamBench proofs at $4 per problem, 75x cheaper than proprietary tools, finding five previously unknown bugs in open-source code.
Anthropic released Claude Sonnet 5 with Opus-level benchmarks at 20 percent lower cost, starting July 1, forcing competitors to respond.
Google's Gemini 3.5 Pro launches with 2M token context window, Deep Think reasoning mode, and $15/$60 pricing—the largest frontier model memory to date.
Z.ai's GLM-5.2 ships a 1M-token open-weight model with an Anthropic-compatible API one day after the U.S. government banned Fable 5.
Moonshot AI's Kimi K2.7 Code delivers 30% fewer reasoning tokens and 21.8% higher coding benchmark scores, open-sourced under MIT license on June 12.
Anthropic's Fable 5 tops SWE-Bench Pro at 80.3%, beating GPT-5.5 by 22 points, opening the Mythos-class tier at $10 per million tokens.
Apple unveiled 5 AFM foundation models at WWDC 2026, distilling Google Gemini and running the cloud tier on Nvidia GPUs in Google Cloud infrastructure.
OpenAI's GPT-5.6 appears in Codex routing logs under codename iris-alpha, with 1.5M tokens of context and 80-89% Polymarket odds of a June launch.
OpenAI will retire GPT-4.5 from ChatGPT on June 27 and o3 on August 26, completing the first full clearout of the pre-GPT-5 model generation.
Alibaba's Qwen 3.7 Plus adds vision and video to its Qwen backbone at $0.40 per million tokens, 6x cheaper than Qwen 3.7 Max, shipping June 1.
Anthropic's leaked 512,000-line Claude Code source map exposes Sonnet 4.8 and a skipped 4.7, plus hidden API tiers and routing thresholds rivals can study.
Gemini 3.5 Pro ships a 2 million token context window, double Flash, as Google launches its cheaper model first to pressure rivals on price.
Microsoft MAI-Code-1-Flash, a 5B in-house coding model, beats Claude Haiku 4.5 on SWE-Bench Pro and uses up to 60% fewer tokens in GitHub Copilot.
Google Gemma 4 12B is an encoder-free multimodal model that runs text, image, audio, and video on a 16GB laptop under Apache 2.0 with a 256K context.
Microsoft launched 7 MAI models spanning code, reasoning, image, voice, and transcription at Build 2026, a full-stack bid to cut its OpenAI reliance.
Claude Opus 4.8 scores 61.4 on the Artificial Analysis Intelligence Index and 69.2% on SWE-Bench Pro, beating GPT-5.5 just 41 days after Opus 4.7.
OpenAI's GPT-Realtime-2 cuts voice agent latency below one second with GPT-5-class reasoning and a 128K context, resetting voice AI economics.
DeepSeek V4-Pro made its 75% price cut permanent at $0.87 per million tokens, a frontier-grade coding model that undercuts US labs with open weights.
MiniMax M3, an open-weight model from Shanghai, beats GPT-5.5 on SWE-Bench Pro at 1/12 the cost, putting the AI coding frontier up for free download.
Ideogram 4 is a 9.3B open-weight text-to-image model with native 2K resolution that tops every open rival on text rendering and design quality.
xAI Composer 2.5 launched June 1, 2026, scoring 69.3 on agentic coding versus Grok 4.20 at 47.1, built on Moonshot Kimi K2.5 open weights.
Qwen3.7 Max scored 56.6 on the Artificial Analysis Index on May 20, 2026, the top Chinese model, beating DeepSeek V4 Pro on agentic coding.
OpenAI GPT-Rosalind beats GPT-5.5 across drug discovery and genomics while using 31% fewer tokens, and opens its science model preview worldwide.
Nvidia Cosmos 3 launches as the first fully open physical AI omnimodel, generating video and robot actions to cut robotics training from months to days.
Google Gemma 4 ships under Apache 2.0 and its 31B model beats 400B open rivals on agentic and coding tests while running on a single GPU.