New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Google just launched Gemini 3.7 Flash, its latest workhorse model targeting software engineering and multi-step agents. It delivers major benchmark jumps—like DeepSWE climbing from 49% to 65.3%—at half the token cost ($0.75 per million input tokens). Google AI Pro and Ultra subscribers also get the new model in their Gemini Spark agent starting today.
DeepSeek just launched DeepSeek-V4-Pro into general availability, introducing major agent upgrades alongside flexible reasoning levels (low, high, and max) across V4-Pro and V4-Flash. It natively supports OpenAI's Responses API for Codex and rolls out off-peak API pricing that cuts token costs by 50%.
OpenAI just rolled out Ultrafast mode for GPT-5.6 Sol, a new API tier powered by Cerebras chips that outputs up to 750 tokens per second—making it roughly 14 times faster than standard processing. Instead of sacrificing intelligence for speed by dropping to tiny models, teams can now run frontier-class AI on real-time incident response and financial workflows. It is currently available in a limited preview.
Google DeepMind just introduced SL2T, a sign-language-to-text model rolling out on the Pixel 11's Gboard and Live Transcribe. Trained on over 100,000 hours across 50+ sign languages, it tracks body landmarks on-device to translate ASL directly into English text for searches, chats, and Gemini prompts. It basically gives signers the exact same hands-free dictation hearing users have had for years.
xAI just rolled out Grok 4.6, a major upgrade tailored for long-running AI agents and visual development. The model matches GPT-5.6 Sol on the Artificial Analysis composite benchmark, excelling at sustained multi-step coding and self-verification. It is live now on Cursor, Grok Build, and via API starting at $2 per million input tokens.
Microsoft just rolled out MAI-Code-1.1-Flash inside GitHub Copilot, slashing inference costs by 75% while boosting token speed by 25%. The model specifically targets CLI and .NET workflows, showing a 22% jump on Terminal-Bench. A newly added 3-bit quantized version also lets developers run the 256K-context model entirely on local hardware.
Nvidia is developing Nemotron 4, an open-source model double the size of its current flagship, to take on the world’s best open weights. Led by VP Bryan Catanzaro, the initiative aims to broaden GPU demand well beyond a handful of frontier labs like OpenAI, backed by $28 billion in cloud compute commitments through 2031. Why it matters: Nvidia wants to ensure it doesn't just sell shovels, but shapes the ecosystem to keep everyone buying them.
Anthropic just committed to embedding invisible watermarks in Claude-generated text and attaching C2PA credentials to files to comply with the EU AI Act. The system bakes subtle patterns into token selection, allowing tags to survive copy-pasting without degrading writing quality. It marks a big step toward tracking machine-made content, though detection tools remain limited to regulators and select partners for now.
MATS Program researchers just exposed a vulnerability across OpenAI, Anthropic, and Google APIs that lets users extract hidden thinking tokens from frontier models. The encrypted reasoning blobs were fully portable across models and leaked dozens of API keys. All three AI labs have begun rolling out patches.
OpenAI just launched its Daybreak cybersecurity models on AWS through Amazon Bedrock. Approved enterprise teams can deploy Daybreak Blue for defense (powered by GPT-5.6 Sol) and Daybreak Red for vulnerability research and exploit testing. Why it matters: security teams can now run frontier AI directly within their existing AWS governance workflows without moving sensitive data.
OpenAI just expanded its Daybreak program with GPT-5.6-Cyber, a specialized model engineered for vetted cyber defenders. While standard models reject sensitive security prompts, this new tier completes 95% of advanced exploit-related requests—up from just 1.5% on GPT-5.6 Sol. It has already uncovered high-severity Chrome V8 flaws and over 400 kernel vulnerabilities.
OpenAI just warned that its upcoming Astra model has shown such rapid advancements in coding and cybersecurity that it may reach the "Critical" risk threshold under its Preparedness Framework. That means the model could potentially identify and exploit real-world zero-day vulnerabilities without human guidance. In response, OpenAI is locking down internal testing environments, strengthening weight protections, and pausing Astra activities that don't meet new safety controls.
Anthropic just updated its biology safety classifiers for Claude Fable 5, cutting false-alarm fallbacks to Opus 5 by about 85%. Users can now ask everyday medical and educational questions—like analyzing lab tests or symptoms—without triggering defensive blocks. Virology and molecular design remain locked down, with trusted access planned for researchers.
OpenAI just overhauled ChatGPT's everyday models. Free users now get unlimited text chats powered by GPT-5.6 Luna along with a dedicated "Think" button for complex queries. For Plus and Pro subscribers, an updated GPT-5.6 Sol cuts factual mistakes by 68% and introduces a slider to dial reasoning effort up or down on the fly.
Alibaba just launched Qwen3.8-Max, its most powerful flagship multimodal model featuring 2.4 trillion total parameters while activating just 95 billion via sparse MoE. It supports a 1M token context window and ranks second in Vision Arena. The model is accessible via API on Alibaba Cloud Model Studio, with weights releasing next week.
MiniMax just unveiled H3, an omni-modal generation model that natively handles text, images, video, and stereo audio in a single context. It generates up to 15-second clips at 2K resolution at a fraction of competitors' costs. MiniMax also plans to open-source the model weights in the coming days, taking a rare open-weights bet against closed video giants.
Thinking Machines just launched Inkling-Small, an open-weights MoE model with 276B total parameters and only 12B active. Operating at roughly a quarter of the compute cost, it actually edges out the flagship Inkling across reasoning and agentic coding benchmarks like Humanity's Last Exam. Why it matters: You get multimodal reasoning and a 1M context window without the heavyweight inference tax.
OpenAI just dropped API pricing across the GPT-5.6 lineup, slashing Luna by 80% down to $0.20 per million input tokens and Terra by 20%, alongside a new 2.5x faster Fast mode for Sol. The massive price drop makes multi-step agentic workflows and high-volume data processing drastically cheaper to run at enterprise scale.
Microsoft just unveiled its in-house MAI model lineup, rolling out over a dozen task-specific models that slash GPU costs by 50% to 89% across tools like GitHub Copilot and PowerPoint. Its new MAI-Cyber-1-Flash topped the CyberGym benchmark at half the cost of prior options. Why it matters: Big tech is shifting away from running mega-models on everyday tasks, proving specialized smaller models can match frontier quality at a fraction of the budget.
xAI just released Grok Voice 2.0, a speech-to-speech model that reasons through queries in parallel while speaking, cutting time-to-first-audio down to 0.7 seconds. It beat GPT-Realtime-2.1 on Artificial Analysis benchmarks and costs $0.08 per minute, making high-speed, human-like voice agents practical for real-world phone calls.
