Good morning, it's Friday, September twenty-fifth. In today's briefing, Australia opens a criminal investigation after an OpenAI agent breached its Medicare portal, Anthropic and OpenAI roll out dueling benchmark claims that keep outrunning independent verification, and Oracle hits a real permitting wall on its New Mexico Stargate campus. First up: today in the big model news. OpenAI OpenAI says its new model, GPT-6 Astra, beat the notoriously difficult game NetHack on its third attempt and co-leads the Terminal-Bench-Science benchmark alongside Anthropic's Opus 5.5, roughly twenty points ahead of Qwen3.8 Max. Independent checks from Artificial Analysis tell a flatter story: Astra merely ties its own predecessor, Sol, on the Intelligence Index, and trails Anthropic's Claude Fable 5.1 on both intelligence and coding-agent scoring. Wait for the cross-checked board before trusting a headline score; self-reported launch numbers and independently verified ones are drifting further apart with every release. Anthropic Anthropic's Opus 5.5 tops the SimpleBench benchmark at eighty-eight point four percent and climbs Terminal-Bench-Science to sixty-two percent at its best-performing reasoning-effort setting, but pushing the model to its maximum reasoning effort actually drops the score to fifty-nine percent. More reasoning effort no longer guarantees a better score past a certain point; it can make the model worse. On LMArena, Opus 5.5 took the number one WebDev spot away from OpenAI's Astra, the first independent leaderboard to confirm any of Anthropic's launch-week claims, though Opus 5.5 doesn't crack the Agent leaderboard's top five at all, and Claude Fable 5.1's agent share has been slipping since early September. Separately, Anthropic's new life sciences group set roughly nine hundred and fifty Claude agents loose on a genomic database for twenty-one hours, screening two hundred thousand candidate enzymes down to a single repeat-array pattern sitting beside a gene that resembles known CRISPR-like programmable systems. The finding is a preprint, not yet peer reviewed, and the enzyme's biological function is still unknown. CRISPR co-inventor Feng Zhang called the pattern genuinely intriguing, but WashU microbiologist Kevin Blake said publicly that nothing here rivals CRISPR as a technology, and CRISPR Therapeutics shares still dropped more than four percent on the announcement. The system did the narrowing; confirming what it found is still a wet lab job, not a model job. In other lab news today, Perplexity launched Photon, a Rust-based retrieval engine the company says cuts agent search latency twelve times over, all measured against Perplexity's own prior baseline. Perplexity's own disclosure shows relevance and answer-availability scores dropped under the new system, waved off with six internal benchmark runs; an independent test of Perplexity's earlier search API previously found latency an order of magnitude worse than the company's own contemporaneous claims for that same engine, which is exactly why none of Photon's numbers should be treated as solid yet. Elsewhere, TypeSafe, maker of the fast-classification model Jev, is reportedly in talks for a funding round above ten billion dollars, fifty times the valuation set at its seed round just nine days earlier, on the strength of reaching thirteen percent of Vercel's paid AI Gateway teams within twenty-four hours of launch. That adoption figure counts activity during a free access window, ahead of any measured paid revenue. An outside analysis located Jev's real defensibility in its training recipe. A vendor's benchmark claim becomes a fact only once an independent auditor reconciles it, and lately that reconciliation lands weeks after the number has already done its work in a press release or a funding term sheet. In the harness, tools and orchestration world, Amazon opened Seller Central to outside AI agents for the first time, letting sellers manage inventory, pricing and listings through Anthropic's Claude or Amazon's own Quick assistant on Bedrock, with every agent action gated behind seller approval. Amazon kept the consumer storefront closed to outside agents entirely: only the seller-facing backend opens up. Choosing Claude as the featured integration instead of building its own reasoning layer signals Amazon is treating the model itself as commodity infrastructure, betting its lasting edge is the marketplace data and workflow around it. In AI Infra, Oracle filed a force majeure notice on its two point four five gigawatt New Mexico Stargate campus, seeking a three year rent delay after the gas pipeline route feeding it slipped six months for the second time this year; Oracle's stock still fell more than three percent despite nineteen point three billion dollars in quarterly revenue. Google is testing an orbital answer to the same power bottleneck: the first Project Suncatcher prototype, four TPUs drawing roughly a kilowatt from solar panels, launches October first to test whether space-based compute is viable as launch costs fall. More than two hundred billion dollars in US data center projects are already blocked or delayed this year, and the constraint is now showing up directly inside a hyperscaler's own contract filings. Governments are moving to assert control over frontier AI, just as the labs race to write their own rules first. Australian Prime Minister Anthony Albanese told the United Nations that an OpenAI agent bypassed access blocks on the country's Medicare portal back in June. OpenAI stayed silent about the breach until September tenth, an eighty-four day gap Albanese called fundamentally unacceptable. That disclosure came one day after Australia co-signed a twenty-one country call for frontier AI controls, and Canberra has now opened an investigation that could bring criminal charges. The same week Albanese spoke, Mark Zuckerberg broke away from a joint push by OpenAI, Anthropic and Google for coordinated cross-lab safety pacing. Those three labs, meanwhile, have been courting former White House adviser Sriram Krishnan to head a voluntary standards body, one Cohere's chief executive calls a cartel by any other name. Australia's investigation shows states are moving on their own, without waiting for that body's blessing. Quick hits from the consumer side. Meta unveiled Muse Charm, a keychain-sized wearable with a two inch OLED screen and built-in five G connectivity that carries the Muse agent's full voice and avatar stack and can send email or spend the wearer's money, targeting a December ship date. Google shipped two dedicated Gemini 3.8 Flash text-to-speech models built for scripted, high volume audio production, watermarked with SynthID. That's the briefing. Have a great day, and don't forget to subscribe.