Anthropic published results this week from a project it calls Claude Science, in which Claude Opus 4.8 and the experimental Mythos Preview model were asked to design "minibinders" — small proteins engineered to latch onto a specific biological target, an early and usually months-long step in drug development. Given a single expert-written prompt, Claude worked largely on its own: researching each target, choosing binding sites, running open-source structure and sequence tools, and screening its own candidates for solubility and novelty.
The models generated 1,320 candidate designs across 15 targets, including cancer and Alzheimer's-linked proteins. Two outside labs, Adaptyv Bio and Twist Bioscience, physically built and tested them without Anthropic's involvement. The result: 354 confirmed binders spanning 14 of the 15 targets, with hit rates of 22 to 35 percent against a typical field rate of 10 to 15 percent. On one target, Mythos Preview hit 40 percent versus 3.7 percent for human entrants in a public design contest.
Anthropic is careful to note a binder is not a drug, and critics including former pharma executive Martin Shkreli called the affinities modest. The company says its most capable models remain restricted for life-science use pending an access program for vetted scientists — a nod to the same tools' dual-use risk.
Facing more than 100 rumored data center proposals and rising local backlash, Pennsylvania Governor Josh Shapiro signed an executive order on August 18 that fundamentally changes the state's posture toward AI infrastructure. Under the new Governor's Responsible Infrastructure Development, or GRID, requirements, the Department of Environmental Protection will only review a data center permit if the developer has made a legally binding commitment to the standards and secured local approval first.
The order removes all AI data center projects from Pennsylvania's Fast Track permitting program, bars state agencies from signing nondisclosure agreements with developers, and requires companies to cover the full infrastructure costs their facilities create rather than passing them on to ratepayers. "If the local community doesn't approve a project, the state won't approve it either," Shapiro said at the signing.
The reversal is notable: Shapiro had previously championed data center growth, including Amazon's $20 billion commitment to two Pennsylvania sites. Those projects, already permitted, are grandfathered in. Advocacy groups called the order a step forward but said it stops short of the moratorium some communities wanted.
Starting August 19, U.S. college students can claim a free year of Google AI Pro — normally $19.99 a month — while students in more than 140 other countries get a year of Google AI Plus. Both plans fold into a new Student Hub inside the Gemini app at gemini.google.com/students, which centers on study notebooks built from a student's own course material, AI-generated flashcards, practice quizzes, and calendar-linked deadlines.
The U.S. tier throws in four times higher usage limits, the Gemini Spark agent platform, Gemini integration in Gmail and Docs, 5TB of storage, and a Google Health Premium subscription. Redemption runs through December 31, and the plans auto-renew as paid subscriptions unless canceled, with a valid payment method required up front.
The timing lands three months after Gemini crossed one billion monthly users, and follows the same back-to-school playbook OpenAI ran with a free ChatGPT tier for students. With Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 Sol both currently ranked ahead of Gemini on major benchmarks, the push doubles as a bid to lock in habits before the frontier race is settled.
Project-management platform Linear released usage telemetry this week showing just how fast AI has moved from novelty to default inside software teams. Two years ago, fewer than one issue in a thousand created in Linear was written by AI. Today, AI agents and integrations author close to half of everything logged on the platform, and at the current growth rate they will soon outpace humans and integrations combined.
The catch is what the data does not show: faster delivery. Pull requests have roughly tripled across the period the report covers, but overall development time has still crept upward rather than down, echoing separate industry findings that AI-generated code often needs more review, not less, before it ships. Stack Overflow's own developer survey found "almost right, but not quite" code as the top frustration cited by AI users.
Together, the numbers point to a widening gap between how much AI output teams are producing and how much of it is actually landing in production without extra human cleanup — a gap Linear's report frames as the next real bottleneck for AI-native software delivery.
A newly disclosed research finding describes a flaw in how three major labs handle reasoning continuity in stateless API sessions. OpenAI returns encrypted reasoning items for developers to replay manually, Anthropic carries full reasoning in an encrypted signature, and Google uses encrypted "thought signatures" — all designed to let a model's internal reasoning persist across calls without exposing it to the application layer or other models.
Researchers demonstrated that weaker models could, under certain conditions, decode reasoning content meant to stay opaque, effectively reading a stronger model's private chain-of-thought. The team disclosed the findings privately to the affected providers, along with Microsoft and Hugging Face, before publishing; all three labs shipped mitigations, and the researchers say the main extraction technique no longer works as of this month. No evidence of the flaw being exploited in the wild has surfaced.
The practical guidance for developers is straightforward: strip reasoning blocks and opaque fields before sharing API traces publicly, and avoid committing raw transcripts even after visible text has been sanitized — the reasoning field itself can still carry information the visible output does not.