This open-source model tops GPT-5, Claude Sonnet 4.5

Moonshot AI just shook up the AI landscape with Kimi K2 Considering. This open-weight model now beats GPT-5, Claude Sonnet 4.5, and MiniMax-M2 on reasoning, coding, and agentic benchmarks. It’s a trillion-parameter Mixture-of-Experts system, activating 32 billion parameters per inference, running as much as 200–300 sequential tool calls, and handling 256k-token context windows with INT4 quantization, meaning it will probably sustain long multi-step workflows like autonomous coding loops or structured evaluation without breaking a sweat.

Benchmark highlights: 60.2% on BrowseComp, 71.3% on SWE-Bench Verified, 83.1% on LiveCodeBench v6, and 56.3% on Seal-0. K2 Considering outputs reasoning traces, giving full transparency for multi-step, agentic reasoning. Runtime costs are tiny with $0.15 per million tokens for cache hits, far cheaper than GPT-5. Released under a modified MIT license, it’s fully open for research and business use, with a light-touch attribution for high-scale deployments. Read more.

OpenAI just cleared a serious milestone with over 1 million paying business customers, making it the fastest-growing enterprise AI platform in history. Adoption is fueled by the 800M weekly ChatGPT users, shortening pilot cycles and speeding org-wide rollouts. ChatGPT for Work now sits at 7M seats (+40% in two months), and ChatGPT Enterprise is up 9× year-over-year.

Users can now interrupt long-running queries and inject recent context mid-stream, a subtle but big upgrade for deep research and GPT-5 Pro workflows. The corporate continues to expand globally with a 50/50 three way partnership with SoftBank will deploy a localized enterprise stack, “Crystal intelligence,” with SoftBank because the launch customer.

Altman projects $20B+ in annual revenue for 2025, climbing toward a whole bunch of billions by 2030. OpenAI is committing $1.4T over eight years to expand AI Cloud, enterprise tools, consumer devices, and robotics. No government bailouts will likely be requested. Any public support, Altman says, should fund broad AI infrastructure or chip production for national security. Current capability limits are already delaying recent features, particularly in scientific and medical workflows.

On consumer traction, Sora launched on Android with ~470K installs on day one across seven markets, 4 times the adjusted iOS debut. Plus, recent ChatGPT apps, including Peloton and Tripadvisorexpand the platform’s ecosystem. Read more.

Google rolls out DeepMind’s AlphaEvolve, an AI system attacking classical and unsolved math problems. It blends evolutionary coding with LLMs, iteratively generating, testing, and refining solutions. Gemini Deep Think verifies outputs, while AlphaProof formalizes proofs in Lean. Across 67 combinatorics, geometry, and number theory challenges, AlphaEvolve rediscovered known results and sometimes produced formulas that outperformed human solutions. Mathematicians including Terence Tao validated outputs, proving hybrid human-AI discovery isn’t just hype.

Google’s seventh-generation Ironwood TPU goes live for large-scale AI training and inference. Expect 4× gains for training and inference, with superpods scaling to 9,216 chips, 9.6 Tbps interconnects, and 1.77 PB high-bandwidth memory. Complementing Ironwood, Arm-based Axion VMs handle preprocessing and pipelines with as much as 2× higher price-performance versus x86. Early adopters like Anthropic committed billions for as much as 1,000,000 TPUs, ensuring low-latency, enterprise-scale inference.

Opal, Google’s no-code AI mini-app builder, is now in 160+ countries, letting users automate workflows, generate content, and prototype apps in minutes. File Search throughout the Gemini API simplifies RAG pipelines, robotically managing storage, chunking, embeddings, and context injection. Early adopters like Phaser Studio cut multi-hour research to seconds using semantic vector search with built-in citations.

Maps becomes Gemini’s “all-knowing copilot,” turning navigation into conversational, task-driven guidance with Street View evaluation across 250 million places. Multi-step voice workflows cover stops, parking, and Calendar events, while Lens adds on-device scene understanding.

Google Finance integrates Gemini Deep Research with prediction market data from Kalshi and Polymarket, plus AI-powered earnings insights. Read more.

OpenEnv also links to major RL ecosystems including TorchForge, verl, TRL, and SkyRL, supporting composable, scalable agent development. Meta and Hugging Face are inviting RFC feedback and contributions, positioning OpenEnv as an ordinary framework for protected, production-ready autonomous AI workflows. Read more.

• Foxconn is sending humanoid robots to Texas to construct AI servers in the approaching months

• Helios is Quantinuum’s big leap toward practical, error-corrected quantum computing

• As OpenAI flirts with “backstop” talk, Trump’s AI czar draws a tough line: no bailouts coming

• A UK court just ruled that Stable Diffusion isn’t an infringing copy, dealing a serious blow to Getty’s copyright case

• Germany just dropped its largest fully licensed text dataset

• MIT’s recent AI mapping system builds accurate 3D environments in seconds

• China tests ‘wolf robot’ to destroy Taiwan’s defenses in 10 seconds

• Ten robot dogs leading the charge in military, research, and industrial applications

• A bug-inspired computer model helps AI sync sight and sound the best way human brains do

• Edtech corporations are keeping schools at the hours of darkness about how their AI actually works

• This AI smart ring enables you to record voice notes with a whisper

• If AI keeps doubling its task horizon every six months, the economy could hit breakneck growth

• Nvidia’s Jensen Huang: ‘China goes to win the AI race,’ FT reports

• Bank of America is doubling down on AI, pouring $4 billion a yr

• Pinterest CEO touts open source AI: ‘tremendous performance’ with reduced costs

• China accelerates AI self-sufficiency with sweeping ban on foreign data centre chips

• Your thoughts, captioned. Recent AI turns non-invasive brain scans into detailed scene descriptions

• Sandbar’s Stream Ring turns your voice into text and AI responses, promising a “mouse for speech” experience

• Edison Scientific debuts Kosmos, an AI scientist that makes multi-day experiments and verified discoveries

• Microsoft AI says it’ll make superintelligent AI that won’t be terrible for humanity

• Foursquare co-founder launches BeeBot, an AI “radio DJ” that talks you thru your neighborhood

• AI adoption is triggering a brand new wave of layoffs, pushing October job cuts to a 22-year high

• Laude Institute launches Slingshots grants to fuel 15 ambitious AI research projects

• Meet Polaris Alpha. A brand new under-the-radar model designed to outperform on coding, tools, and real-world use

• Ai2 launches OlmoEarth, an open platform turning satellite data into real-time planetary insights for everybody

• Greg Brockman is quietly becoming probably the most powerful builder within the AI boom

• Watch: Xpeng debuts IRON humanoid robot, calling it its most anthropomorphic yet with solid-state power

• Tinder to make use of AI to get to know users, tap into their Camera Roll photos