New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Google just integrated computer use directly into its mainstream Gemini 3.5 Flash model, doing away with the standalone Gemini 2.5 experimental setup. Developers can now build agents that observe and act across browser, mobile, and desktop environments out of the box, backed by new prompt injection safeguards and optional approval gates. It gives Google a fast, production-ready competitor in the automated workflow race.
Alibaba just revealed Wan Streamer, an end-to-end foundation model capable of real-time, full-duplex video conversations. Instead of chaining separate transcription and voice tools, a single Transformer handles text, audio, and video with roughly 200 ms model latency. Why it matters: It shifts AI interactions from turn-taking voice bots to sub-second face-to-face video agents.
ByteDance just launched the Doubao Seed 2.1 model family, introducing Pro and Turbo versions built specifically as productivity agents. Alongside strong coding and multimodal chart analysis, it can accurately overhaul PPT templates and generate websites from rough sketches. It marks ByteDance's clearest push yet to shift AI from a simple chatbot into a reliable digital coworker.
Baidu just introduced Unlimited OCR, a 3B MoE document parser that stops memory bloat across 40+ pages. By using an R-SWA mechanism to cap the text KV cache to the last 128 tokens, the model processes long documents at a completely constant speed.
Sakana AI just launched Fugu and Fugu Ultra, a learned orchestration system that acts like a single foundation model while coordinating a pool of expert agents under the hood. It routes complex tasks across swappable LLMs through one OpenAI-compatible endpoint. Why it matters: developers can tap into frontier collective intelligence for coding and research without wrestling with multi-agent orchestration or single-vendor lock-in.
OpenAI is reportedly preparing to roll out GPT-5.6 Pro, bringing a jump in reasoning budget from 768 to 960 and integrating Playwright for browser automation. Leaks show it can cook full websites from just a prompt and image references. Why it matters: rapid front-end prototyping is getting significantly faster, though early testers note overall UI styling still trails competitors.
OpenAI just rolled out a major health intelligence upgrade to ChatGPT powered by GPT-5.5 Instant, available to all free users. Guided by a global network of over 260 physicians, the model matches frontier Thinking models on tough medical benchmarks, while real-world factual errors in health queries dropped 71% over the last two months.
xAI just made its flagship Grok 4.3 model available on Amazon Bedrock, opening the doors for AWS developers to integrate it directly into their workflows. xAI touts the model as an industry leader in both tool-calling speed and low hallucination rates. It gives cloud builders direct access to Grok inside Amazon's enterprise-grade security stack.
Z.ai just introduced GLM-5.2, an open-source flagship model built for long-horizon coding tasks with a reliable 1M-token context. Powered by its new IndexShare architecture, it cuts per-token FLOPs by 2.9x at full context, edging past GPT-5.5 on FrontierSWE while remaining fully accessible under an MIT license.
Cartesia just launched Sonic-3.5 for text-to-speech and Ink-2 for live transcription, both topping Artificial Analysis's streaming leaderboards. Built on state space models rather than standard transformers, the duo pushes latency below 300ms with built-in turn detection. That finally eliminates the awkward pauses in real-time voice agents.
Moonshot AI just rolled out a HighSpeed mode for its K2.7-Code model, hitting up to 260 tokens per second — a 6x speed jump. It reportedly cuts reasoning tokens by 30% through post-training tweaks, though independent coding benchmarks have yet to confirm the claims.
Google just launched Gemini 3.5 Live Translate, a new audio model delivering near real-time voice translation across 70+ languages. Instead of waiting for you to stop talking, it streams continuous audio just seconds behind the speaker while preserving natural pitch, pacing, and intonation. It is rolling out to the Google Translate app, Google Meet, and developers via the Gemini Live API.
Anthropic just unveiled Claude Fable 5, its powerhouse Mythos-class model designed for general availability. It leads frontier benchmarks across software engineering, vision, and scientific research. To mitigate severe cyber risks, conservative classifiers automatically route sensitive queries—less than 5% of total sessions—back to Opus 4.8, while an unrestricted twin called Mythos 5 rolls out to vetted cyberdefenders.
Alibaba just released Qwen 3.7 Plus, a multimodal model built to navigate graphical interfaces, write code from templates, and run terminal commands within a single loop. In testing, it autonomously built a complete vocabulary app over eleven hours, leading Western rivals on GUI benchmarks like AndroidWorld. It also supports Anthropic's API protocol, making it a dirt-cheap drop-in backend for existing agent tools.
Google just introduced Agentic RAG for its Gemini Enterprise Agent Platform, ditching the single-step retrieval approach. A specialized Sufficient Context Agent evaluates draft answers, spots missing details, and runs targeted follow-up queries, driving a 34% jump in factual accuracy with almost zero added latency.
Nvidia just launched Nemotron 3 Ultra, an open 550B MoE model with 55B active parameters tailored for complex multi-turn agent workflows. It combines Mamba and Transformer layers to slash agentic task costs by up to 30% while pushing 5x higher throughput than comparable open models. Nvidia also rolled out a 4B safety guardrail and a sub-100ms multilingual speech model.
OpenAI just refreshed its official prompt engineering guide, laying out best practices for message roles and prompt structuring in the Responses API. The guide urges developers to version prompts directly inside application code, noting that hosted prompt objects will be phased out by November 2026.
Microsoft just rolled out seven in-house MAI models covering reasoning, coding, vision, and speech, all built from scratch on clean data without third-party distillation. The drop includes MAI-Thinking-1, a lightweight coding model for GitHub Copilot, and an enterprise tuning workflow called Frontier Tuning. Why it matters: Microsoft is aggressively engineering full-stack independence across its own silicon and software.
Anthropic’s new cybersecurity model Mythos is catching long-hidden software flaws in record time, but it comes with a massive cost. The model reportedly runs about six times more per token than Opus, burning through $1 million in tokens within weeks for Palo Alto Networks. Security executives still argue the price tag beats the nine-figure cleanup of a major breach.
MiniMax just released M3, an open-weight model packing a 1M token window, native multimodality, and desktop computer use into a single package. Its custom sparse attention architecture cuts decoding compute dramatically, yielding a 15x speedup over its predecessor and top marks on SWE-Bench Pro. It brings closed-source agentic capabilities straight to open-weight developers.
