This week in AI: GPT-5.6, chips, and agents

Good morning,

Here’s your AI recap for the week.

OpenAI is rolling out GPT-5.6 more carefully after government feedback, while Anthropic is partially restoring access to its strongest model after a government-related suspension. Elsewhere, Claude moved into Slack, Gemini gained computer use, and Mistral shipped practical document AI.

Let's dive in!

OpenAI is starting GPT-5.6 with a limited preview

OpenAI announced GPT-5.6 Sol, its next flagship model, alongside two smaller models: Terra and Luna. The rollout starts with a limited group of trusted partners before wider availability.

• Sol is the new flagship model for agentic work.
• Terra is a performance model at half the cost.
• Luna is the most cost-efficient model

The interesting part is not just the models. OpenAI says it previewed the plans and capabilities with the U.S. government before launch, and is taking this phased approach after government feedback.

Frontier models are becoming powerful enough that launch strategy, safety testing, and government coordination are now part of the product story.

Anthropic partially restores Mythos 5 access after government suspension

After suspending access to both Claude Fable 5 and Claude Mythos 5 on June 12 following a U.S. government export control directive, Anthropic announced that it has received approval to redeploy Mythos 5 to a set of U.S. organizations that operate and defend critical infrastructure.

The company said it has been working closely with the government since the suspension and is restoring access for these organizations quickly. Broader access to Mythos 5 and general availability of Fable 5 are still pending further discussions.

This marks a partial reversal of one of the most aggressive government interventions on a frontier model release so far. While select U.S. organizations focused on critical infrastructure are regaining access, broader availability for developers and general users is still pending.

OpenAI is building its own AI chip

OpenAI and Broadcom announced Jalapeño, OpenAI’s first custom inference chip for running large language models. Engineering samples are already being tested, with first deployment planned for the end of 2026.

OpenAI is moving further down the infrastructure stack. As AI usage keeps growing, the winners will not only need better models. They will also need cheaper and faster ways to run them at scale.

Claude is moving into Slack

Anthropic launched Claude Tag, a Slack-based agent for Team and Enterprise users. Teams can bring Claude into selected channels, give it approved context and tools, and let it work asynchronously.

This is a clearer version of workplace AI. Instead of opening a separate chatbot, teams can bring Claude into the same channels where projects are already discussed.

Google added computer use to Gemini 3.5 Flash

Google added computer use as a built-in tool for Gemini 3.5 Flash. Developers can now build agents that see interfaces, click, type, and complete tasks across browser, mobile, and desktop environments.

Agents are getting closer to using software the way humans do. That is powerful, but it also raises the bar for safeguards around sensitive actions, permissions, and prompt injection.

Sakana launched Fugu

Sakana AI launched Fugu and Fugu Ultra. They look like one model through an OpenAI-compatible API, but behind the scenes they coordinate multiple expert models and agents.

This is an interesting approach to model performance. Instead of relying on one large model for everything, Fugu routes work across several systems. Promising idea, but the benchmark claims should still be read carefully.

Mistral shipped OCR 4

Mistral launched OCR 4, a document intelligence model for extracting structured content from PDFs and enterprise files. It supports 170 languages and includes layout information, confidence scores, markdown output, and block classification.

This is not the flashiest launch, but it is very practical. Better document extraction means better search, RAG, compliance workflows, and internal knowledge systems.

Quick hits 🗞️

• Meta keeps pushing AI glasses. The company launched new Meta Glasses styles starting at $299. The bigger signal is that AI assistants are still searching for a better interface than the chat box.
• OpenAI is moving AI security toward actual patches. It expanded Daybreak and launched “Patch the Planet” with Trail of Bits to help open-source maintainers find, validate, and fix vulnerabilities.
• GLM-5.2 is showing up across more providers. GLM-5.2 appears to be gaining distribution across model providers, often with long context and competitive pricing. Worth watching, especially for developers who care about cost, routing, and open-weight options.
• Claude is gaining paid consumer traction. TechCrunch reports Claude’s paid consumer revenue is up ~75% since January. It’s still far behind ChatGPT, but it’s becoming a serious option for heavy professional users.
• Anthropic accused Alibaba-linked operators of large-scale distillation. The company claims fraudulent accounts were used to extract nearly 29 million Claude interactions. The allegation highlights a growing reality: model outputs are now valuable assets worth stealing.
• OpenAI voice mode rumors. TestingCatalog spotted signs of a more natural bidirectional voice experience in ChatGPT, though nothing has been officially announced yet.

See you next week ツ

Got feedback on the newsletter? Just hit reply, I read every response.

If this helped you catch up quickly with AI, forward it to someone who’d find it useful. They can subscribe here.