July 2026

Claude Sonnet 5, Grok 4.5, GPT-5.6 Sol/Terra/Luna, and Kimi K3 all shipped within three weeks. Here's the full timeline — what launched, what it costs, how it benchmarks, and what the wave means.

Moonshot AI's Kimi K3 is the largest open-weight model released yet — 2.8T parameters, a 1M-token context, and $3/$15 pricing. Here's how it benchmarks against Opus 4.8, GPT-5.6, and Grok 4.5, plus how to use it and when open weights land.

The Google DeepMind blog index features recent posts covering new Gemini and Gemma model releases, AI safety and responsibility initiatives, scientific research applications, and partnerships with A24 and Singapore.

Anthropic detailed the cybersecurity safety classifiers for its globally redeployed Claude Fable 5 model and released an early draft of an AI jailbreak severity framework developed with Glasswing partners. The company also launched a HackerOne program for researchers to submit potential cyber jailbreaks.

TestingCatalog: Meta AI web gets scheduled tasks support

Meta is developing scheduled task functionality for its Meta AI web interface to close feature gaps.

TestingCatalog: Claude Pro and Max Users Gain Artifacts in Claude Code

Anthropic expands Artifacts feature to Claude Code for Pro and Max subscribers, enabling interactive page sharing via private links in beta for Team and Enterprise plans.

TestingCatalog: Claude Code rolls out Artifacts for Pro and Max users

Anthropic expands Artifacts to Claude Code for Pro and Max users, enabling shareable interactive pages in beta for teams.

TestingCatalog: Which Team Are You Rooting For?

TestingCatalog asks followers which team they support without providing context on the specific competition or topic.

Sam Altman: Sam Altman celebrates USA World Cup win on Fourth of July week

Altman shares his excitement about watching the USA win a World Cup match during America's birthday week.

TestingCatalog: Codexonical 👌

TestingCatalog posts a brief, cryptic message saying "codexonical" with an okay hand emoji.

Claude Devs: Artifacts in Claude Code expand to Pro and Max plans

Artifacts in Claude Code are now available on Pro and Max plans, letting users generate live, private code pages that update in real time.

TestingCatalog: Vellum Launches Cooperative AI Assistants for Slack

Vellum introduces AI assistants with memory and permissions that can collaborate in Slack channels, available for testing via a new feature launch.

TestingCatalog: Vellum AI assistants gain memory, permissions, and Slack cooperation

Vellum launched AI assistants with granular memory and permissions that can cooperate autonomously in Slack channels.

TestingCatalog: Vellum Debuts Agent-to-Agent Communication in Slack

Vellum launches multi-agent collaboration in Slack enabling autonomous AI coworkers with isolated permissions to negotiate tasks and gather team input.

TestingCatalog: Vellum launches autonomous AI coworkers in Slack

Vellum introduces agent-to-agent communication in Slack, letting personalized assistants collaborate and negotiate tasks autonomously.

TestingCatalog: Vellum launches agent-to-agent AI communication in Slack

Vellum introduced agent-to-agent communication in Slack, allowing AI assistants to carry user context and split tasks.

TestingCatalog: Source: Anthropic in talks with Samsung to manufacture custom AI chip

TestingCatalog shared a source article from The Information about Anthropic discussing custom AI chip manufacturing with Samsung.

TestingCatalog: Anthropic reportedly teaming with Samsung on AI chip

Anthropic is reportedly exploring an AI chip partnership with Samsung, following OpenAI's similar move with Broadcom.

TestingCatalog: Anthropic may develop custom AI chip with Samsung

Anthropic is reportedly in talks with Samsung to develop its own AI chip, following OpenAI's recent partnership with Broadcom.

Claude: Claude Fable 5 arrives in Claude Tag

A conversation with Boris Cherny and Cat Wu explores Claude Tag's evolution across Anthropic, alongside the launch of Claude Fable 5 in Claude Tag.

Min Choi: Restaurants charging same price for DoorDash as dine-in should be studied

Min Choi questions why restaurants don't mark up delivery prices, suggesting it's a phenomenon worth studying.

Min Choi: Restaurants charging same for DoorDash as dine-in need study

Min Choi suggests restaurants that charge identical prices for delivery and in-person dining deserve academic scrutiny.

Claude: Anthropic announces $100K Built with Claude life sciences hackathon

Anthropic is hosting a global virtual life sciences hackathon with Gladstone Institutes, inviting participants to build with Claude Science and Claude Code for a $100k credit prize pool.

Claude: Built with Claude life sciences hackathon set for July 7-13

The Built with Claude: Life Sciences virtual hackathon will run July 7-13, and interested participants must apply by Sunday via the provided link.

Min Choi: OpenAI reportedly discussed giving US government 5% stake

Min Choi reacts to news that OpenAI explored giving the US government a 5% equity stake, calling AI national infrastructure.

Min Choi: OpenAI reportedly considered giving US government 5% stake

Min Choi reacts to news that OpenAI discussed offering the US government a 5% equity stake, calling AI national infrastructure.

Min Choi: OpenAI Reportedly Discussed Giving US Gov 5% Stake

OpenAI reportedly discussed giving the US government a 5% equity stake, signaling that AI is evolving into national infrastructure.

Min Choi: Min Choi drops another 'It's over' take

Min Choi posts a cryptic 'It's over...' message, likely reacting to a recent disruptive AI product launch or update.

TestingCatalog: Xbox Controller SVG Prompt Awaiting Generation

An AI prompt requesting an optimized SVG of an Xbox controller appears to be queued or pending processing.

TestingCatalog: OpenAI proposes 5% US government stake worth $42.6B

OpenAI reportedly proposed giving the U.S. government a 5% stake worth roughly $42.6 billion to share AI upside with the public.

TestingCatalog: OpenAI proposes 5% US government stake worth $42.6B

OpenAI has proposed giving the U.S. government a 5% stake worth roughly $42.6 billion, with Sam Altman arguing it would share AI upside with the public.

TestingCatalog: OpenAI proposes 5% US government stake worth $42.6B

OpenAI reportedly offered the US government a 5% stake valued at about $42.6 billion to share AI upside with the public.

Min Choi: Claude Fable 5 returns from timeout, users go wild

Min Choi announces the return of Claude Fable 5 and teases a thread of 10 wild user examples.

Min Choi: Claude Fable 5 returns from timeout with wild user examples

Anthropic's Claude Fable 5 is back online and users are already sharing impressive use cases and examples.

Vellum has released agent-to-agent communication in Slack, where each user’s AI assistant retains its own memory and permissions while cooperating on tasks.

Anthropic has expanded its Artifacts feature to Claude Code users on Pro and Max plans, letting them generate live, self-contained web pages from a chat session.

Anthropic is reportedly in discussions with Samsung to manufacture a custom AI chip, following a broader industry trend of AI labs pursuing dedicated silicon to reduce reliance on third-party GPU providers.

Vellum has introduced agent-to-agent communication in Slack, allowing each team member’s AI assistant to retain individual memory and permissions while collaborating on tasks like scheduling and gathering input.

In a viral post on X, commentator Min Choi observed that restaurants charging identical prices for DoorDash delivery and in-person dining defy conventional wisdom about delivery markups and 'should be studied.'

Anthropic has announced a global virtual hackathon focused on life sciences in partnership with the Gladstone Institute. The 'Built with Claude: Life Sciences' event will run July 7-13 and offer $100,000 in credits to participants building with Claude Science and Claude Code.

Mistral's news page lists recent announcements including the Mistral OCR 4 document intelligence model, Vibe agent updates with Work and Code modes, a Search Toolkit, and physics AI foundations for engineering. Additional updates cover Mistral Small 4, Medium 3.5, Forge for enterprises, and a partnership with NVIDIA.

OpenAI has reportedly discussed offering the US government a 5% equity stake, signaling a potential shift toward treating AI as national infrastructure.

OpenAI has reportedly proposed granting the U.S. government a 5% equity stake in the company, a holding valued at roughly $42.6 billion, according to the Financial Times.

Explore the essential factors in choosing an LLM API, ensuring portability and avoiding vendor lock-in while maintaining performance and cost-effectiveness.

Anthropic is restoring global access to Claude Fable 5 and U.S. access to Mythos 5 after the U.S. government lifted export controls imposed on June 12, 2026. The company also deployed an improved safety classifier to address a reported jailbreak technique.

Anthropic launched Claude Sonnet 5, an agentic AI model that plans, uses tools, and operates autonomously at performance levels near Opus 4.8 for a lower price. The company said it improves on Sonnet 4.6 in reasoning, coding, and safety, with substantially weaker cybersecurity capabilities than Opus models.

Anthropic has opened a Seoul office and signed an MOU with Korea's Ministry of Science and ICT to advance AI safety, while announcing new enterprise, startup, and research partnerships across South Korea.

Anthropic has released Claude Science, a beta AI workbench that integrates scientific tools, databases, and compute resources into a single environment for researchers, with pre-configured skills for genomics, proteomics, and other life sciences domains.

Anthropic introduced Claude Tag in beta for Enterprise and Team customers, a Slack integration that lets teams tag @Claude as a collaborative channel member to delegate tasks asynchronously. The tool features persistent memory, proactive ambient updates, and administrative controls over tools and spend, and Anthropic reports its internal version now generates 65% of its product team’s code.

Anthropic has published an Advanced AI Framework that would grant the U.S. government authority to block dangerous deployments of the most capable AI models and require transparency, independent evaluation, and robust security from frontier developers. The proposal applies to models trained above 10²⁵ FLOPs and recommends societal resilience measures while arguing against federal preemption of stronger state AI laws.

OpenAI engineers investigated mysterious crashes in their Rockset C++ data infrastructure by applying an epidemiological approach to a year’s worth of core dumps, uncovering two unrelated bugs: silent hardware corruption on a single Azure host and an 18-year-old race condition in GNU libunwind.

An OpenAI blog post previewed GPT-5.6 Sol, a next-generation model, via a series of ASCII art patterns.

OpenAI and Broadcom have unveiled Jalapeño, a custom AI accelerator designed specifically for large language model inference. Early testing indicates the chip delivers substantially better performance per watt than current state-of-the-art hardware, with deployment expected to begin at gigawatt scale by the end of 2026.

Mistral published multiple updates on its news blog, including the Mistral OCR 4 document intelligence model, new Work and Code modes for the Vibe agent, a Search Toolkit for production pipelines, and a new physics AI initiative for engineering systems.

Alibaba’s Happy Horse 1.1 video generation model is hosted on Replicate, offering text-to-video, image-to-video, and reference-to-video generation. According to Replicate, output costs $0.14 per second at 720p and $0.18 per second at 1080p, with clip lengths from 3 to 15 seconds.

The Google DeepMind blog features multiple announcements covering new Gemini and Gemma model releases, AI safety and responsibility programs, and scientific research tools. Updates span faster text generation, voice translation, computer-use capabilities, and international partnerships for housing, education, and robotics.

OpenAI has introduced GeneBench-Pro, a benchmark of 129 synthetic problems designed to test whether AI agents can make high-level judgment calls in computational biology. The company’s strongest evaluated model, GPT-5.6 Sol, achieved a 28.7 percent pass rate, a significant improvement from the sub-5 percent score of earlier frontier models on the original GeneBench.

HP Inc. announced it will scale activation of its OpenAI Frontier strategic partnership after pilots showed strong results, with plans to deploy AI across customer experiences, operations, software development, and cybersecurity. Frontier will serve as a unified governance and evaluation layer for HP’s enterprise-wide AI workflows and agent deployments.

OpenAI published an economic research paper finding that its Codex agent has become the primary AI tool across all company departments, with non-developer adoption surging and users delegating tasks estimated to exceed multiple hours of human work.

TestingCatalog: Early Look: Anthropic's Claude Science Research App

TestingCatalog published an exclusive early preview of Anthropic's upcoming Claude Science application designed specifically for researchers.

Claude Devs: Fable 5 rate limits reset for all users

The 5-hour and weekly rate limits have been reset for all users now that Fable 5 is ready to build again.

Claude Devs: Fable 5 rate limits reset for all users

Anthropic resets the 5-hour and weekly rate limits for all users now that Fable 5 is ready to build again.

Min Choi: Claude Fable 5 is back online

Min Choi confirms the return of Claude Fable 5 after its timeout.

TestingCatalog: Fable Adds New Dedicated Usage Section

AI tool Fable appears to have introduced a new dedicated usage section in its interface, hinting at updated features or billing controls.

TestingCatalog: Anthropic Says Behavior Unchanged Since Ban

Anthropic indicated that its systems or policies remain the same as they were prior to a previously imposed ban, offering no new changes.

TestingCatalog: Claude Fable 5 returns with 50% usage promo until July 7

Anthropic has brought back Claude Fable 5 for paid users, allowing up to 50% of weekly usage limits until July 7 before requiring credits.

TestingCatalog: Claude Fable 5 returns with expanded usage limits

Anthropic brings back Claude Fable 5, allowing paid users to use up to 50% of their weekly limit until July 7 before requiring usage credits.

TestingCatalog: Claude Fable 5 returns with 50% weekly usage allowance

Anthropic has brought back Claude Fable 5 until July 7, allowing paid users to use up to 50% of their weekly plan limit before needing usage credits.

Claude: Claude Fable 5 blog post goes live

Anthropic has published a blog post detailing Claude Fable 5, directing readers to their official news site for the full announcement.

Claude: How to report mistaken flags in Claude Code

Users can report mistakenly flagged requests in Claude Code using the /feedback command, or via thumbs buttons on the web and Cowork platforms to help reduce false positives.

Claude: Fable 5 is back

Claude's official account simply announces that Fable 5 is back.

Min Choi: Min Choi asks where Claude Fable 5 went

The user inquires about the sudden disappearance or downtime of Claude Fable 5.

TestingCatalog: xAI launches Voice Agent Builder on xAI Console

xAI announced a new conversational Voice Agent Builder on xAI Console that lets users create solutions powered by Grok voices and connectors.

TestingCatalog: xAI launches Voice Agent Builder on xAI Console

xAI announced a new conversational Voice Agent Builder on xAI Console that lets users create agents powered by Grok voices and connectors.

Min Choi: ChatGPT now connects to financial accounts for money tracking

OpenAI's ChatGPT gains the ability to link with financial accounts to analyze spending and offer useful budgeting insights.

TestingCatalog: Anthropic unveils Claude Science desktop research app

Anthropic is distributing Claude Science, a desktop app for scientific research that runs as a local web server with memory categories and custom specialists.

TestingCatalog: Anthropic unveils Claude Science desktop app for researchers

Anthropic is distributing Claude Science, a desktop app for scientific research that runs as a local web server and features custom agents called Specialists.

TestingCatalog: Google testing new Gemini Flash checkpoint on LM Arena

Google is testing a new Gemini Flash checkpoint on LM Arena that may release as Gemini 3.6 Flash or Gemini 4 Flash.

Min Choi: Claude Fable 5 returns after unexpected downtime

Min Choi reacts with surprise as Anthropic's Claude Fable 5 comes back online.

Anthropic: Anthropic to redeploy Claude Fable 5 globally with new safety classifiers

Anthropic announces Claude Fable 5's global return with enhanced cybersecurity classifiers, a new jailbreak assessment framework with tech partners, and expanded US government collaboration on model testing.

Anthropic: Anthropic restores Claude Fable 5 globally with new safety classifiers

Anthropic announces Claude Fable 5 will be available globally again after working with the US government to implement new classifiers blocking cybersecurity tasks, alongside a new industry framework for AI jailbreaks.

Anthropic: Anthropic to globally redeploy Claude Fable 5 with new safety classifiers

Anthropic announces Claude Fable 5 will return worldwide tomorrow with stricter classifiers for cybersecurity tasks, a new industry framework for AI jailbreaks, and expanded US government collaboration on testing and saf

Anthropic: Anthropic to globally redeploy Claude Fable 5 with new cybersecurity classifiers

Anthropic announces Claude Fable 5's global return with new government-collaborated classifiers to block cybersecurity misuse, plus a new industry framework for AI jailbreak response.

Explore the Model Context Protocol (MCP), its significance for LLMs, and practical guidance for adoption in your projects.

Explore AI safety and alignment fundamentals, including capabilities, threat models, and key distinctions to understand current challenges.

Explore the real implications of 1M-token models, examining long context tradeoffs in performance, recall, latency, and costs.

Explore prompt caching, its mechanisms, design strategies for efficiency, and why it’s crucial for optimizing LLM performance.

Explore crucial LLM cost optimization strategies, including prompt caching and model routing, to manage and reduce operational expenses effectively.

Unpack the importance of on-device LLMs for privacy, explore current capabilities, and understand the tech behind it.

Learn the essentials of self-hosting LLMs, focusing on serving engines, batching techniques, and cost considerations for open-weights production.

Explore how latency impacts real-time voice agents, from ASR to TTS, and learn about efficiency in speech-to-speech LLM systems.

Explore how LLMs integrate with browser agents for web automation, including DOM parsing and action space intricacies.

Explore code completion LLMs, examining context, predictions, and what distinguishes top-performing tools like GitHub Copilot.