This week’s OpenAI and Anthropic recap, reset week edition (Week 27, 2026)

OpenAI was on reset week but still showed off a Science plan and GeneBench-Pro, Anthropic launched Claude Sonnet 5, brought back Fable 5 and Mythos 5 after export controls lifted, and shipped Claude Science, and more

OpenAI was on a company-wide reset week with most of the team on time off, but a few things still landed and got spotted: HP scaled its OpenAI Frontier partnership, OpenAI introduced GeneBench-Pro (a research-level computational biology benchmark), published an engineering post on fixing an 18-year-old libunwind bug, shared new Signals data on how ChatGPT adoption has widened globally and a report mapping Europe’s AI jobs transition, Clarative.ai joined the company, and a ChatGPT “Sophia” Science plan (chatgptsciplan), a “Calculators” medical calculator catalog, a Codex mechanical keyboard with Work Louder (Project 2077) tied to a July 15 Codex shortcuts teaser, and a “dotball” soccer Easter egg all surfaced, plus reports that OpenAI found a way to cut inference costs in half and proposed handing the US government a 5% stake

Anthropic launched Claude Sonnet 5 (its most agentic Sonnet yet), redeployed Fable 5 and Mythos 5 after the US lifted export controls (with new cyber classifiers and a proposed jailbreak severity framework), released Claude Science (an AI workbench for scientists), made Claude in Microsoft Foundry generally available, introduced the Claude apps gateway for Amazon Bedrock and Google Cloud, brought Claude Desktop to Linux in beta, added Artifacts to Claude Code on Pro and Max, shipped admin spend visibility and controls for Enterprise plus a batch of Managed Agents updates, removed fast mode for Opus 4.6, extended a 2x Cowork usage promotion through August 5, and signed a first-of-its-kind deal giving California state agencies Claude at half price, and more

HP Inc. announced it will scale its OpenAI Frontier strategic partnership after a series of successful pilots, expanding AI across customer and partner experiences, security, employee productivity, and software development, using Frontier as the unified platform to govern access, context, deployment, and evaluation

[embed]

Spotted a new ChatGPT “Sophia” plan for Science internally referenced as “chatgptsciplan”, offered as one seat at $200 a month with the first 12 months free, requiring a promo code and US-only billing

[embed]

Spotted a new ChatGPT “Calculators” feature, a catalog of medical and clinical calculators

[embed]

Spotted OpenAI and Work Louder building a Codex-integrated mechanical keyboard codenamed Project 2077 (a Creator Micro v2 variant with a reasoning-effort rotary encoder and fast, approve, reject, and fork keys), tied to an official OpenAIDevs teaser promising a Codex shortcuts upgrade on July 15

[embed]

[embed]

Anthropic introduced the Claude apps gateway for Amazon Bedrock and Google Cloud, a self-hosted control plane that adds corporate SSO login, centrally enforced policy, role-based access, spend caps, and per-user cost attribution for Claude Code

Anthropic made Claude generally available in Microsoft Foundry hosted on Azure, starting with Opus 4.8 and Haiku 4.5 in the Messages API, running in your Azure environment with existing authentication, billing, and governance controls

[embed]

OpenAI Economic Research extended its AI Jobs Transition Framework to the EU, mapping four transition archetypes across member states and finding the EU has a smaller share of jobs with higher near-term automation potential than the US

[embed]

Governor Newsom announced a first-of-its-kind partnership giving California state agencies and local governments access to Claude at half price, with free workforce training and technical help, making Claude the first AI productivity tool available to all state agencies through the state’s SITeS portal

[embed]

Anthropic removed fast mode for Claude Opus 4.6, so requests with speed “fast” now run at standard speed and standard pricing instead of returning an error, with fast mode continuing on Opus 4.8

Anthropic launched Claude Sonnet 5, its most agentic Sonnet model yet, with performance close to Opus 4.8 at lower prices, a 1M token context window, adaptive thinking on by default, and introductory pricing of $2 and $10 per million tokens through August 31

[embed]

Anthropic redeployed Fable 5 globally and restored Mythos 5 for a set of US organizations after the US lifted export controls, adding new classifiers that block more cybersecurity tasks (falling back to Opus 4.8), drafting a jailbreak severity framework with Glasswing partners, and scaling up US government collaboration on testing and safeguards

[embed]

Anthropic released Claude Science in beta, an AI workbench that brings scientific tools into one environment with 60+ curated skills, reproducible auditable artifacts, on-demand compute scaling, and a reviewer agent, already used for genomics, protein structure prediction, and long-form review writing

[embed]

OpenAI introduced GeneBench-Pro, a research-level benchmark of 129 synthetic problems testing whether AI agents can work through messy biological data and make real research judgment calls, where its strongest model GPT-5.6 Sol passes 28.7% at the highest reasoning level

[embed]

OpenAI’s engineering team detailed how they fixed seemingly impossible crashes in their Rockset data infrastructure by taking a population-level approach, tracing them to silent CPU corruption on one Azure host and an 18-year-old race condition in GNU libunwind

[embed]

OpenAI Signals shared new data showing ChatGPT adoption widening and deepening worldwide, with users sending 50% more daily messages and doubling their tasks six months in, fastest relative growth in Africa and Asia, and non-English and typically-female-named usage now the majority

Reported that OpenAI developed a software-only optimization that cuts the cost of running its models by more than half using existing servers, shrinking the GPUs serving logged-out ChatGPT traffic to a couple hundred, as it targets a 52% API gross margin by year-end

[embed]

Anthropic brought Claude Desktop to Linux (Ubuntu and Debian) in beta, giving Claude Code, Claude Cowork, and chat a first-class desktop experience on all paid plans, though Computer Use isn’t in this release

[embed]

Anthropic shipped a batch of Managed Agents updates, including event deltas for streaming, backward pagination, per-session config overrides, credential injection location settings, and lifecycle webhooks for agents, deployments, and runs

[embed]

Clarative.ai is joining OpenAI

Anthropic laid out more detail on Fable 5’s cyber safeguards, sorting cybersecurity uses into prohibited, high-risk dual use, low-risk dual use, and benign categories, and published an early draft of its proposed Cyber Jailbreak Severity framework scoring jailbreaks across capability gain, breadth, ease of weaponization, and discoverability

Reported that OpenAI discussed handing the US government a 5% stake to secure political buy-in, with Sam Altman suggesting other US AI labs allot a similar share to a sovereign-fund-style vehicle, though any deal might require an act of Congress

[embed]

Anthropic brought Artifacts to Claude Code on Pro and Max plans, so you can ask for an artifact and Claude writes the code, publishes it live to claude.ai at a private link, and updates it in real time as it keeps working

[embed]

Anthropic introduced richer admin analytics, model-level entitlements, and spend alerts for Claude Enterprise, showing usage and cost by group and user, adding value and usage tabs for Claude Code, and letting admins set model defaults and spend-threshold notifications

Spotted a new ChatGPT “dotball” Easter egg that appears when you use trigger words like “Messi”, “soccer”, “FIFA”, or “World Cup”

[embed]

Anthropic extended its 2x Cowork usage promotion, keeping the higher 5-hour session limit in place through August 5

[embed]

That’s a wrap for this week. A lot of the news came back to the same idea, people getting more out of AI the more they use it. AIPRM is how you speed that up in ChatGPT and Claude, with thousands of ready-to-go prompts, your own writing styles and tones, and easy sharing across your team, so you spend less time starting over and more time on the answer. Check it out at AIPRM.com.

Follow me on X/LinkedIn/Threads/Bluesky for more