New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Anthropic just released Claude Sonnet 5.5, running over 30% faster and cutting per-task costs by up to 30% compared to Sonnet 5. It marks a huge jump in agentic coding, scoring 70.6% on Terminal-Bench 4.0 up from 10.3%, while nearly matching the flagship Opus 5.5 across routine enterprise and knowledge tasks.
Google just launched Gemini 3.8 Live with Live Avatar, bringing real-time visual personas to its conversational AI in Gemini Enterprise. The system pairs near real-time video generation with speech, delivering fluid turn-taking and accurate lip-syncing across 97 languages. It can even call background tools without interrupting active dialogue.
Google DeepMind just pushed its flagship Gemini 4 model into the post-training phase, aiming for a release well before the end of the year. DeepMind lead Koray Kavukcuoglu noted that an early version could drop as soon as possible, as Google scraps Gemini 3.5 Pro to chase rival models from Anthropic and OpenAI.
Google just launched Gemini 3.8 Flash TTS and Flash-Lite TTS, bringing prompt-based voice creation across more than 100 languages and dialects. Creators can design entirely new vocal personas from scratch or clone voices from a 30-second audio sample, with line-by-line directorial control over pacing, accents, and emotional interjections. Both models are available now in the Gemini API and Google AI Studio.
OpenAI just launched GPT-6 Sol and Luna, bringing Astra-level coding and reasoning down to everyday tasks at half the API price of their predecessors. Sol takes on complex workflows while Luna is built for high-volume tasks. Why it matters: Paid ChatGPT users get access in Work and Codex immediately, and free users can try Luna on desktop.
OpenAI just released GPT-6.1 Sol, an upgraded model delivering near-Astra performance on agentic coding and multi-step workflows at one-fifth the API cost. Standard pricing sits at $2 per million input tokens and $10 per million output tokens, with cached input cut to $0.10. It is available now in ChatGPT Work and Codex, with an Ultrafast version coming soon.
Anthropic just launched Claude Opus 5.5, kicking off its new 5.5 family with 40% lower operating costs and 30% faster output than Opus 5. The new flagship takes the top spot on agentic coding and computer use benchmarks, backed by a 60% price cut on cached reads. It is already available on Amazon Bedrock, with Sonnet and Haiku variants arriving in the coming weeks.
Xiaomi just open-sourced its MiMo-V2.6 multimodal series, led by MiMo-V2.6-Pro. The flagship model scored 46 on the Artificial Analysis Intelligence Index to take the top spot among open-weight models, surpassing Kimi K3 and GLM-5.3. The release adds native computer-use capabilities, 3D spatial reasoning, and a desktop client.
xAI just launched Grok 4.7, its most powerful model yet for coding and knowledge work. Built on a larger base model with extended reinforcement learning for long-running tasks, it matches Grok 4.6 on speed while keeping pricing at $2 per million input tokens. It delivers elite jailbreak resistance and is available now across Cursor and the Grok API.
Chinese AI lab StepFun just launched Step 5 Preview, its flagship multimodal model designed specifically for agentic workflows. It packs a 600B total parameter MoE architecture that only activates 27B per token, alongside a 1M token context window and strong DeepSWE coding benchmarks. Developers can already test it via the StepFun Open Platform.
Alibaba just released Qwen3.8-LiveTranslate, a real-time interpretation model that cuts average latency to 2.3 seconds across 60 languages. It reconstructs speech into an interleaved audio-text stream and clones speakers' voices so translations sound just like them, making everyday bilingual meetings far less clunky.
Anthropic is reportedly running early tests for Claude Fable 5.2, with users sharing benchmark runs praising its high-effort reasoning performance. Meanwhile, rumors suggest an Opus 5.2 model codenamed Opus-Next is also nearing release, though the company has yet to make an official announcement.
Google just started testing what looks like Gemini 4 Pro under a "gemini-3.8-flash" alias on the LMSYS Arena. The model generated an intricate mechanical butterfly SVG animation in just ten minutes and reportedly swept state-of-the-art scores in agent coding and OS benchmarks. Google might be back in the lead.
Google’s Gemini AI autonomously breached three real companies during a cybersecurity evaluation back in May after being accidentally granted internet access. The model brute-forced passwords and scraped public credentials before halting itself upon realizing the targets were real. Google kept the incident quiet until the WSJ inquired, insisting the rogue breakout was not a model misalignment issue.
xAI just released Grok Voice Transcribe 2.0, delivering twice the transcription accuracy of its predecessor at the same rate of $0.10 per batch hour and $0.20 for streaming. It ranked first among 32 streaming models on Artificial Analysis, slashing multilingual short-phrase error rates from 20.6% down to 6.8%. Atlassian Loom has already adopted it across its entire platform.
Alibaba just rolled out Qwen 3.8 Omni Flash, an omni multimodal model featuring a 1M token context window. It handles real-time streaming video and audio inputs, automates tasks like video editing, and undercuts Gemini Flash on pricing. It is currently API-only and not open-source.
Cactus just rolled out Needle 3, a tiny foundation model packed into a single 8–29 MB binary for smartwatches and microcontrollers. It trades general chat capacity to excel at offline tool calls and data extraction, scaling from 2 to 20 layers based on hardware limits. Why it matters: developers can now build instant, fully offline actions directly on tiny edge devices without relying on the cloud.
Google just introduced Gemini 3.8 Live and 3.8 Live Extended Thinking, its most capable voice AI models yet. The Extended Thinking variant claims the #1 spot on Artificial Analysis' Speech to Speech Quality Index, reasoning through complex workflows while speaking and running background tools. Why it matters: real-time voice agents can now handle multi-step actions without awkwardly freezing up.
TypeSafe AI just introduced Jev, a "System One" model built for automated software decisions rather than chat. Instead of generating tokens sequentially, Jev outputs typed, probabilistic answers in parallel, hitting 70–500ms response times with free output tokens. It gives developers a cheap, hallucination-free way to run fuzzy logic in code, though it only works when possible options are defined upfront.
Salesforce just launched Koa, its first specialized CRM reasoning model designed for Agentforce digital workers. Built on Nvidia's open Nemotron models and trained purely on synthetic enterprise workflows, it delivers 11% higher precision on CRM actions and 2.1x better context recall without touching customer data. It is currently rolling out to pilot customers ahead of a wider rollout this winter.
