OpenAI has launched GPT-Live, a full-duplex voice model capable of simultaneous listening and speaking. For complex tasks requiring web search or deep reasoning, GPT-Live hands off the workload to GPT-5.5 in the background while maintaining active conversation. It is rolling out globally to ChatGPT users, with early API access opening for developers.
Sourceopenai.com↗Cognition, the creator of Devin, has introduced SWE-1.7. Post-trained from Kimi K2.7 inside its agent harness, the model matches GPT-5.5 within a single point on Cognition's FrontierCode benchmark at a cost of approximately $2 per task. It is available inside Devin and supports coding sessions lasting up to six hours.
Sourcecognition.com↗OpenAI has officially released its GPT-5.6 model family, featuring the flagship Sol model alongside Terra and Luna tiers. Sol approaches Anthropic's Fable on intelligence benchmarks while excelling in agentic coding, computer use, design, and cybersecurity. Alongside the models, OpenAI introduced ChatGPT Work, a platform similar to Claude Cowork, and merged Codex into a revamped desktop app featuring computer control.
Sourcex.com↗Elon Musk's rebranded AI lab, SpaceXAI, has released Grok 4.5. Co-trained with Cursor for coding and agentic workflows, the model reportedly outperforms Claude Opus 4.8 on several coding benchmarks at a lower cost and with fewer steps. It is now the default model for Grok Build and is available to Cursor users.
Sourcex.ai↗Meta has launched Muse Spark 1.1, an agentic AI model featuring a 1M-token context window, computer use capabilities, and parallel subagent execution. Priced at $1.25/$4.25 per million input/output tokens, it aims to undercut rivals while leading on agent reasoning and tool use benchmarks.
Sourceai.meta.com↗Meta has launched Muse Image, an AI-powered model for generating high-quality visuals. It is integrated with Instagram and WhatsApp, allowing users to create images, use AI effects, and redesign rooms via Facebook Marketplace. It is free for general use with advanced options in subscription plans.
Mistral has launched Leanstral 1.5, an open-source model designed for formal verification within the Lean 4 programming language. The model scored 100% on the miniF2F math benchmark, solved 587 of 672 PutnamBench problems, and successfully identified 11 real bugs (including 5 new vulnerabilities) across 57 Rust repositories.
Sourcegithub.com↗ByteDance has released Seedream 5.0 Pro, an image model designed to understand design with advanced editing features like precision editing, layer separation, and multilingual support across over 10 languages.
Sourceseed.bytedance.com↗OpenAI has released GPT-5.6, featuring three models: Sol, Terra, and Luna. Sol leads in intelligence and efficiency for coding, cybersecurity, and science, outperforming competitors like Claude Fable 5.
Sourceopenai.com↗OpenAI released gpt-realtime-2.1-mini and gpt-realtime-2.1, single-model architecture voice models that process audio and text over a live connection, reducing latency and costs.
Sourcedevelopers.openai.com↗OpenAI is rumored to be preparing a rapid release of GPT-5.6, potentially as early as July 7, to compete with Anthropic's Fable 5 models. The model reportedly shows strong benchmark performance, and Sam Altman compared its math breakthroughs to a child forming words.
Sourcex.com↗ByteDance is rumored to be launching its Dreamina Seedance 2.5 model on July 9, capable of outputting 180-second-long videos across Dreamina, CapCut, and other partner platforms.
Sourcetestingcatalog.com↗Meta teased its upcoming Muse Video model, which currently ranks No. 3 on the Arena leaderboard in its preview stage.
Google introduces Gemma 4, a new generation of open-weight multimodal language models ranging from 2.3B to 31B parameters, featuring a thinking mode, long-context efficiency, and a unified encoder-free architecture.
Sourcearxiv.org↗Kyutai and General Intuition, in collaboration with Epic Games, released MIRA, an open-source world model that runs a live 2v2 Rocket League game entirely inside a neural network without a game engine.
Sourcemira-wm.com↗Meta's superintelligence chief Alexandr Wang reportedly shared with employees that the company's upcoming model, code-named 'Watermelon,' has matched OpenAI's GPT-5.5 in internal benchmarks. Currently in training, Watermelon utilizes roughly 10x the compute of its predecessor, Muse Spark. Wang also teased an upcoming Muse Spark update featuring significant coding and agentic improvements, alongside a separate Opus-level coding model expected to release soon.
Sourcebusinessinsider.com↗Meta has launched Muse Image, its first in-house image model developed by Alexandr Wang's Superintelligence Labs (MSL). The model features agentic capabilities, web search, tool use, and self-editing, debuting at No. 2 on Arena's text-to-image and editing leaderboards.
Sourceai.meta.com↗Microsoft has developed MAI-Thinking-1, a reasoning model trained on 30 trillion pretraining tokens of primarily human-generated data. The model uses reinforcement learning to generate original chains of thought across STEM, coding, and safety specialists, which were then distilled into a single student model. It achieves 97.0% on the AIME 2025 math benchmark, outperforming Claude Sonnet 4.6 and DeepSeek V3.2, though it trails on graduate-level science and agentic coding.
Tokyo-based Sakana AI released Fugu and Fugu-Ultra, two models designed to delegate tasks to other models and agents under a single API. Fugu handles discrete tasks like basic coding and chat, while Fugu-Ultra is built for long-running tasks like extensive coding and research.
Tencent has developed Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters. Hy3 outperforms similar-sized models and is available on Hugging Face, as well as for free on OpenRouter until July 21.
Sourcesimonwillison.net↗Tencent has released Hy3, a highly efficient open-source model under a permissive Apache 2.0 license, avoiding regional bans and offering cheap, reliable agent deployment.
Sourcehy.tencent.com↗The U.S. government granted Anthropic permission to offer its Claude Mythos 5 model to roughly 100 companies and federal agencies, two weeks after forcing the company to suspend both Claude Mythos 5 and Claude Fable 5.
Claude Cowork is now available across web and mobile, allowing users to run scheduled tasks even when their computer is off. The update is rolling out to Max subscribers. Additionally, Anthropic has extended access to Fable 5 on paid plans through July 12.
Sourceclaude.com↗Anthropic has extended access to Fable 5 via Claude subscriptions until July 12, allowing users to use up to 50% of their weekly limits before transitioning to prepaid usage credits.
Sourcesupport.claude.com↗Anthropic has rolled out a beta 'Reflection' dashboard for users with Memory enabled. The tool tracks chat habits over the past year, mapping behavior against Anthropic's 4D AI Fluency Framework and offering quiet hours and break reminders.
Sourceanthropic.com↗Anthropic has appointed former Federal Reserve Chair Ben Bernanke to its independent Long-Term Benefit Trust to monitor how AI economics impact the labor market.
Sourcecnbc.com↗Anthropic launched 'Claude for Open Source', offering qualifying open-source maintainers six months of free Claude Max 20x (valued at $1,200) to run agentic refactoring loops and triage issues.
Sourceclaude.com↗Google Photos has introduced a Gemini Omni-powered tool that allows users to relight video clips, replace backgrounds, and apply artistic styles like watercolor and oil painting.
Sourceblog.google↗Microsoft is replacing third-party models from OpenAI and Anthropic with its own internally built MAI models in applications like Excel, Outlook, GitHub Copilot, and Teams to reduce dependency and costs.
Sourcebloomberg.com↗Oasis has launched Oasis 1, a $289 smart ring that enables users to dictate to AI using Whispr Flow and control connected devices via an integrated trackpad.
OpenAI has acquired applied-AI firm Northslope, adding hundreds of forward-deployed engineers to its team to assist with enterprise AI deployment.
Sourceaxios.com↗OpenAI has introduced ChatGPT Work, a new mode powered by GPT-5.6 (available in Sol, Terra, and Luna tiers) that can run complex multi-hour workflows, pull data from apps like Google Drive, Slack, and Salesforce, and output finished documents, spreadsheets, and web apps.
Sourceopenai.com↗Sotheby's is auctioning the 'Jensen Jacket,' an autographed leather jacket previously worn by Nvidia CEO Jensen Huang, with bids estimated to reach $40,000 to $60,000.
Sourcesothebys.com↗Anthropic has discovered "J-space," an emergent internal scratchpad inside Claude where the model performs silent reasoning before responding. This discovery allows researchers to audit a model's silent planning, such as detecting words like "fake" and "fraud" in a sabotage-trained model before any suspicious output is generated. Anthropic has open-sourced the "J-lens" tool for auditing, and Neuronpedia has released an interactive demo.
Sourceanthropic.com↗An OpenAI analysis of SWE-Bench Pro's construction, model failures, and task metadata concluded that approximately 30% of its public tasks were broken, highlighting how flawed evaluations can distort assessments of coding ability and progress.
Sourceopenai.com↗Anthropic published research showing that Claude utilizes a hidden workspace to process and answer difficult prompts, a setup they describe as strikingly similar to human brain function.
Sourceanthropic.com↗DoorDash detailed DashBench, an internal benchmark testing AI code reviewers against 105 past code changes. A pairing of Kimi K2.6 and Claude Fable 5 performed best, catching about two-thirds of the problems.
Sourcecareersatdoordash.com↗A new proxy called pxpipe claims to cut Claude Code bills by up to 70% by rendering bulky context as cheaper image tokens. While some skeptics warn that lossy compression can misread exact strings like hashes, researchers note that pixel-based reading has been validated by AI labs for years.
Sourcegithub.com↗Toya is an AI executive assistant that operates via iMessage, makes phone calls, and integrates with user workflows.
Alibaba is reportedly banning its staff from using Anthropic's Claude Code starting July 10, labeling it as high-risk software. This follows a viral Reddit post claiming the tool contained hidden code targeting Chinese users, which Anthropic explained was part of an anti-abuse experiment to prevent model distillation and reselling.
Sourcetechcrunch.com↗ByteDance's Doubao and Alibaba's Qwen are shutting down their customizable AI agent features to comply with China's upcoming Interim Measures for AI-Based Anthropomorphic Interactive Services, which regulate AI simulating human personality traits and emotional interaction.
Sourcebloomberg.com↗A video demonstration showing how the new Claude Cowork update works on mobile and web.
Sourceyoutube.com↗Anything about this issue — a missed story, off categories, clustering, tone. Free-form, like notes you'd send the team.