New releases, benchmarks, pricing and API changes from GPT, Claude, Gemini, DeepSeek, Qwen and other large language models, updated daily.
Subquadratic just introduced SubQ, a frontier model featuring a 12 million token context window powered by sub-quadratic sparse attention. At 1MM tokens, it runs 52x faster than FlashAttention at under 5% of Opus's cost, cutting compute by nearly 1,000x.
OpenAI just rolled out GPT-5.5 Instant as the new default model for all ChatGPT users. The update cuts hallucinated claims by 52.5% on high-stakes topics like medicine and finance, trims conversational fluff, and handles personal context from past chats much faster. A new memory sources tool also lets users see and delete the exact history used to shape answers.
xAI just rolled out Grok 4.3 alongside a new web-based voice cloning suite. The flagship model features permanently active reasoning, a 1-million-token context window, and agentic tool use like web browsing and code execution. API pricing has dropped roughly 40% to 60% compared to Grok 4.2, costing $1.25 per million input tokens and $2.50 per million output tokens.
OpenAI's GPT-5.5 just became the second AI model to fully crack UK AISI's enterprise cyberattack simulation, trailing only Anthropic's Claude Mythos. It completed the 32-step penetration test—typically a 20-hour job for human experts—in 2 out of 10 attempts while hitting a 71.4% pass rate on expert-level CTF challenges, though industrial control systems remain unbreached.
OpenAI just explained why its models started obsessing over goblin and gremlin metaphors starting with GPT-5.1. A post-mortem traced the verbal tic to reward training for ChatGPT's "Nerdy" personality, which favored creature words and accidentally leaked into the broader model via RL feedback loops. OpenAI retired the persona in March and cleaned its training data.
NVIDIA just unveiled Nemotron 3 Nano Omni, an open multimodal model unifying vision, audio, and language into a single 30B-A3B architecture. It delivers up to 9x higher throughput than comparable open omni models by eliminating the need to pass data across separate perception models. Why it matters: Agentic workflows like GUI navigation and document processing can now run in real time with far lower latency.
SpAItial just unveiled Echo-2, a frontier world model that turns text prompts or a single photo into persistent, interactive 3D environments rendered via 3D Gaussian Splatting. It supports object-level editing and claims to beat World Labs' Marble 1.1 across all WorldScore benchmark metrics.
DeepSeek just launched and open-sourced DeepSeek-V4 Preview, making a 1M token context standard across its services. It delivers two MoE models—the flagship V4-Pro and the lightweight V4-Flash—built around hybrid compressed attention that slashes KV cache memory to just 2% of standard architectures. It basically tackles the persistent context blowout in long-running coding agents, and the new models are already live in chat and via API.
OpenAI just launched GPT-5.5 and GPT-5.5 Pro, heavily optimized for autonomous coding, knowledge work, and complex computer use. The model matches GPT-5.4's per-token latency while requiring fewer tokens and setting a new high score of 82.7% on Terminal-Bench 2.0. Why it matters: AI assistance is shifting from step-by-step chatting to letting agents independently plan, test, and finish messy engineering projects.
xAI just rolled out Grok Voice 1.0 via API, designed for complex multi-step enterprise workflows like customer support and phone sales. It reasons in the background with zero added latency, handles noisy calls and strong accents across 25+ languages, and is already handling Starlink's support lines.
Google just made Gemini Embedding 2 generally available via the Gemini API. The natively multimodal embedding model can search and reason across text, images, video, and audio in a single pipeline, helping developers move complex multimodal search and analysis projects straight into production.
Odyssey just introduced Odyssey-2 Max, a 3x-larger world model capable of continuous, real-time physical simulation. Unlike traditional video generators, it predicts open-ended futures autoregressively to internalize real-world mechanics, topping VBench 2 physics scores. It is currently available in private beta for robotics and gaming partners.
Google just upgraded its autonomous research agents with Gemini 3.1 Pro, introducing Deep Research for low-latency tasks and Deep Research Max for heavy-duty background diligence. Both now connect to proprietary datasets like FactSet via MCP and generate native visual charts. It basically turns Gemini into an enterprise-ready research analyst.
TwelveLabs just unveiled Pegasus 1.5 and Rodeo at NAB Show 2026, alongside an integration with Autodesk Flow Capture. Pegasus 1.5 extracts time-based structured metadata in a single API call, outperforming Gemini 2.5 Pro by 30% on video segmentation benchmarks. Meanwhile, Rodeo acts as an AI co-pilot that lets editors find, cut, and assemble footage using plain language.
Alibaba just released an early preview of Qwen3.6-Max-Preview, offering a first look at its next flagship model. Compared to Qwen3.6-Plus, it brings major leaps in agentic coding, instruction following, and real-world agent reliability. It is Alibaba's latest push to make autonomous agents dependable for complex tasks.
xAI just rolled out Grok 4.3 Beta to its $300/month SuperGrok Heavy subscribers, adding direct generation of formatted PowerPoint decks, populated spreadsheets with working formulas, and PDFs. Musk clarified the model sits at roughly 0.5 trillion parameters, with a 1T version close to finishing training. Why it matters: Grok is finally expanding from chat and reasoning into hands-on office productivity deliverables.
OpenAI just unveiled GPT-Rosalind, a frontier reasoning model purpose-built for biology, genomics, and drug discovery. It beats GPT-5.4 on six of 11 LABBench2 tasks and landed in the 95th percentile against human experts on RNA sequence prediction. While model access requires trusted vetting, OpenAI also dropped a free research plugin on GitHub.
Anthropic just released Claude Opus 4.7 across its web apps and API, keeping pricing steady at $5 per million input tokens and $25 per million output tokens. The model brings notable improvements to complex software engineering, higher-resolution vision, and long-running autonomous workflows. It also debuts new safeguards to detect and block prohibited cybersecurity risks.
Google just released Gemini 3.1 Flash TTS, an expressive speech model featuring inline audio tags that let developers steer tone, pace, and delivery using natural language commands. It supports native multi-speaker dialogue across 70+ languages, notched a 1,211 Elo on Artificial Analysis, and watermarks all outputs with SynthID.
OpenAI just purchased access to Epoch AI’s FrontierMath verifiers to automatically check whether solutions generated by its models actually hold up. FrontierMath tests AI against extremely difficult mathematics problems, giving OpenAI automated, rigorous ground truth to evaluate advanced reasoning capabilities.
