Good morning. It’s Wednesday, July 1st.

You read. We listen. Let us know what you think by replying to this email.

OpenAI engineers just cut inference costs for logged-out ChatGPT traffic by more than half, dropping the number of active Nvidia GPUs needed to power guest users down to just a few hundred. The precise optimization techniques remain a secret, but the hardware savings give the company crucial breathing room while data center expansions move slowly.

OpenAI engineers also fixed long-standing crashes within the data pipelines powering ChatGPT search. A rare bug affecting Rockset, the core data system behind search and plugins, was traced back to faulty Azure hardware and an 18-year-old race condition in GNU libunwind. By analyzing thousands of core dumps, engineers found that Linux signals were corrupting memory during C++ exception handling. OpenAI swapped in libgcc's unwinder and pushed the fix upstream.

On the AI capabilities front, OpenAI launched GeneBench-Pro, a new benchmark with 129 synthetic problems designed to test if AI can handle complex computational biology. It forces models to navigate ambiguous data in genomics and cancer research rather than follow static scripts.

OpenAI is launching physical developer hardware. Teaming up with custom keyboard maker Work Louder, the company teased a compact, square macro pad tied to its Codex coding assistant, dropping July 15.

A teaser video on 𝕏 showed a device based on Work Louder’s Creator Micro 2, featuring programmable keys, dials, and joystick controls. Unlike OpenAI’s separate consumer hardware project with Jony Ive, this device is built purely to accelerate engineering workflows by mapping complex AI coding shortcuts to dedicated physical buttons.

Anthropic just dropped Claude Sonnet 5 as its new default mid-tier model, and while it brings massive upgrades for heavy coding, early testers are pointing out a pretty wild financial catch.

The headline specs look great: a 1M token context window, real-time cyber safeguards, and five distinct reasoning effort settings designed for complex agentic workflows. However, independent benchmarking shows Sonnet 5 can actually end up costing more per completed task than Anthropic’s premium flagship, Opus 4.8. Even with promotional launch pricing slashed to $2 per million input tokens and $10 per million output tokens through August, the sheer volume of generated text means power users might see higher bills.

At the same time, Anthropic is branching out into specialized research and infrastructure. The company launched Claude Science in beta for macOS and Linux, an environment that hooks Claude models directly into high-performance computing clusters and 60 distinct scientific databases to automate biomedical data pipelines.

For engineers looking to deploy autonomous setups, Anthropic published a new development framework for building multi-agent, proactive loops inside Claude Code. To tie it all together, enterprise teams can manage these loops via the new Claude apps gateway, a self-hosted stateless Linux container that handles identity auth and budget tracking across AWS and Google Cloud. Furthermore, Claude Opus 4.8 and Claude Haiku 4.5 reached general availability within Microsoft Foundry.

Meanwhile, the US Department of Commerce officially lifted export controls on Anthropic’s top-tier systems, paving the way for the return of Claude Fable 5 and Mythos 5 after a multi-week ban over jailbreak concerns. To appease regulators and monitor who is using the systems, Anthropic may introduce strict KYC identity verification requirements for Fable 5 access.

Google Cloud is adding specialized "large quantitative models" from SandboxAQ to its marketplace, letting researchers combine Gemini’s workflow skills with math models trained directly on physics equations rather than internet text.

Data teams are also getting TabFM, a foundation model built to execute zero-shot predictions on databases without traditional training, which Google plans to include directly into BigQuery SQL.

For regular users, Gemini Spark is morphing into a full-fledged macOS desktop agent, letting users automate local file management, connect third-party apps like Canva, and eventually trigger multi-step workflows remotely from a phone.

The creative stack got a double dose of media tools, too: a lightning-fast Nano Banana 2 Lite image model launched on OpenRouter to spit out 1K images in four seconds, while a new Gemini Omni Flash video model dropped on a stateful Interactions API to allow iterative chat-style video editing for ten cents a second. To cap it off, NotebookLM is getting mobile-focused, turning dense research papers into 60-second, vertical, TikTok-style explainer clips for Pro subscribers.

Over on the flagship side, Gemini 3.5 Pro is officially clear for a July launch without government interference. While high cybersecurity scores triggered federal restrictions for Anthropic's Fable 5 and OpenAI's GPT-5.6, Google's model escaped review by staying below those unwritten hacking capability thresholds.

Cursor for iOS runs cloud or desktop coding agents from your phone to build software anywhere.

Foresight by Lightning Rod delivers calibrated, benchmark-verified AI forecasts for developers building agents and decision tools.

Oxlo.ai provides unified API access to over 35 frontier AI models with predictable monthly subscriptions.

v0 Design Systems 2.0 imports your existing code, components, and patterns to build and iterate apps in chat.

AgentPeek displays live Claude Code and Codex agent sessions in your Mac notch for easy local monitoring.

Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter.

Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X!

Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months.