The key race: OpenAI actually ships GPT-6 Astra as its first Critical-cyber model, while Google shelves Gemini 3.5 Pro and Grok 5 keeps slipping
OpenAI's frontier reinforcement-learning freeze resolved into more than a gated pause: on September 3 the company began rolling out GPT-6 Astra, first to vetted testers in its Daybreak cybersecurity program, with wider access to ChatGPT Plus, Pro, Business and Enterprise users plus the API and AWS following within days. Sam Altman called it a new capability level and said it had already changed his own workflows, while President Greg Brockman described it as a generational leap some may eventually read as the arrival of AGI. Astra is also the first model OpenAI has designated Critical under its Preparedness Framework for cybersecurity, meaning it can find and exploit unknown security flaws without step-by-step human guidance, so the company is running a split release: unrestricted for most capabilities, hard-gated through Daybreak and the new Daybreak Blue defensive program for the one capability regulators worry about most. That split is the sector's first live test of whether a documented safety framework can hold once the model behind it is actually shipping rather than still paused.
Google DeepMind gave up on shipping its own flagship rather than fixing it: after missing four separate windows since its May 19 unveiling, research firm SemiAnalysis reported in early September that Google has shelved Gemini 3.5 Pro outright and redirected engineering resources into Gemini 4, whose pretraining run on Ironwood TPUs began July 21 and, based on Google's own past release cadence, is not expected to ship before November or December; Google has not confirmed the shelving itself. In the meantime Google kept filling the gap with narrower models, including a coding-focused Gemini 3.8 Flash and, on September 2, a specialized Gemini 3.8 Flash Cyber variant for autonomous vulnerability discovery, gated through a new Fairwind Program for vetted governments, healthcare providers, telecoms and critical-infrastructure operators. xAI remains furthest behind of the four largest US labs: Grok 5 has missed every window Musk has named for it since late 2025, remains in training on Colossus 2 with no committed date, and the nearest real proof point is the incremental Grok 4.7, targeted for mid-September 2026 rather than the generational leap xAI has promised.
Who's ahead: Anthropic narrows its IPO to a real window as OpenAI wrestles with its own trillion-dollar ultimatum, and Moonshot formally opens its Hong Kong listing
Anthropic converted a month of momentum into the sector's most concrete IPO timeline yet: having confidentially filed an S-1 on June 1 off a $65 billion Series H that set a $965 billion valuation, reporting in the first days of September narrowed the target further, a prospectus unveiling after Labor Day (September 7), an investor day in mid-September, and a listing as soon as late September or early October, at a valuation some investors are modeling between $1.5 trillion and $2 trillion that could raise more than $60 billion, which would make it the largest IPO on record. Anthropic used the same stretch to widen its legal standing: a federal court ruled the Pentagon's 'supply chain risk' designation of the company unlawful First Amendment retaliation, and by September 2 Commerce Secretary Howard Lutnick was telling reporters the administration now trusts Anthropic after months of clashes, even as a separate designation remains before the D.C. Circuit and the government is expected to appeal.
OpenAI is working through a harder version of the same decision: it confidentially filed for its own IPO on June 8, but advisers have reportedly presented Sam Altman with a choice between waiting until 2027 for a $1 trillion valuation or going public before the end of 2026 at a lower one, and Altman has called any cut to the trillion-dollar figure a non-starter, a stance made more pointed by reports of a multibillion-dollar net loss last year and cash burn running ahead of revenue in the first quarter of 2026. China's three leading independent labs kept diverging rather than converging on one playbook: Moonshot AI turned its own IPO ambitions into a formal filing, submitting a confidential Form A1 application to the Hong Kong Stock Exchange in the first days of September targeting roughly $3 billion at a $50 billion valuation with Goldman Sachs, China International Capital Corporation and Deutsche Bank as underwriters, while separately pressing Microsoft, Amazon and Google for up to 30 percent of Kimi K3-related cloud revenue in talks still at an early, unresolved stage. Zhipu AI, the first LLM company anywhere to go public, continued supplying the clearest evidence that an early listing does not shield a lab from China's price war, while DeepSeek reached general availability with its V4-Pro flagship on August 13 and kept pressing cost per token, rather than scale, as its main competitive lever.
What decides it: whether Astra's safeguards hold now that it is actually shipping, as courts and regulators start ruling on both sides of the Atlantic
The open letter that OpenAI, Anthropic, Alphabet, Microsoft, Amazon and more than 100 other companies signed on August 27 warning that AI-driven cyberattacks could escalate within months reads differently now that Astra, the model behind the freeze that provoked the letter, has actually shipped: OpenAI's split release, unrestricted for everything except the Critical-tier cyber capability, is the industry's first real-world test of whether a Preparedness Framework can contain the exact risk it was written for once the model is live, and the answer will show up in whether Daybreak and Daybreak Blue access holds or expands as Gemini 4 and Grok 5 close in.
The training-data fair-use fight kept splitting rather than resolving. The US Justice Department filed a statement of interest on September 1 in the New York Times' copyright suit against OpenAI and Microsoft, arguing in a 20-page brief that training large language models on copyrighted material can be fair use and that constraining the practice would hinder American AI leadership and national security, a position the Times called an invitation to take content without permission or payment. Anthropic's exposure on the same underlying question moved the opposite direction the same week: Sony Music Publishing and Warner Chappell sued the company on August 29 over alleged large-scale piracy-sourced training of Claude on copyrighted songs, naming CEO Dario Amodei and co-founder Benjamin Mann individually and seeking statutory damages up to $150,000 per composition, a year after Anthropic's $1.5 billion settlement of a similar author claim, evidence the fair-use question is still being litigated company by company rather than settled sector-wide. In Europe, the AI Office's first formal enforcement step, information requests sent August 29 to a number of general-purpose model providers reportedly including OpenAI, Anthropic and Google covering model security, independent evaluations and post-market monitoring, remains a request for documents rather than a finding, so the sector's first real fine or corrective order is still ahead.
The money and the rules: infrastructure spending pushes toward $800 billion for the year as startup funding stays concentrated in two American labs
Capital kept flowing at the historic scale documented earlier this year: Crunchbase counted a record $510 billion in global startup funding in the first half of 2026, already ahead of all of 2025, with OpenAI and Anthropic alone absorbing roughly $217 billion of it, about 43 percent of every startup dollar raised worldwide in the period. That startup-funding figure sits inside a far larger buildout, the five largest US hyperscalers, Microsoft, Amazon, Alphabet, Meta and Oracle, are projected to spend roughly $660-725 billion on AI infrastructure in 2026 alone, and global AI infrastructure spending overall is now tracking toward roughly $800 billion for the year, with the US accounting for more than 80 percent of it. Anthropic's compute commitments across Amazon, Nvidia-backed Lambda and other partners and OpenAI's roughly $105 billion Nvidia-backed Ohio Stargate financing are both individual line items inside that same wave rather than outliers.
China's capital picture kept splitting along the same public-versus-private line as its labs: Moonshot's newly formal Hong Kong filing and DeepSeek's continued private fundraising are both funding compute buildouts and, in Moonshot's case, distribution deals with US clouds, while Zhipu's post-IPO numbers remain a live data point on how much appetite public markets actually have for a Chinese lab still burning cash against domestic price competition. In Europe, Mistral AI's reported roughly 3 billion euro round at a 20 billion euro valuation, first reported in June, still had not been confirmed closed as of early September, the same pattern of reported-but-unconfirmed mega-rounds recurring across the sector everywhere except the two largest US labs, whose numbers keep getting confirmed in real time.
What to watch through 2027
Watch GPT-6 Astra's broader rollout actually reach ChatGPT Plus, Pro, Business and Enterprise users plus the API and AWS in the coming days as promised, and watch whether its Critical-tier cyber capabilities stay contained to Daybreak and Daybreak Blue as competitive pressure builds. Watch whether Gemini 4 ships in the November-December window its pretraining cadence implies now that Gemini 3.5 Pro has reportedly been shelved rather than fixed, and watch Grok 4.7's mid-September target as the nearest real proof point ahead of any actual Grok 5 date. Watch Anthropic's prospectus surface after Labor Day and the offering land in its newly narrowed late-September-to-October window, and watch whether Sam Altman and OpenAI's board choose a 2027 listing at a full trillion-dollar valuation over a faster, cheaper one this year.
Watch Moonshot AI's formal Hong Kong listing application, due by September 30, and whether its Kimi K3 revenue-share talks with Microsoft, Amazon and Google produce the first such agreement between a Chinese lab and a Western cloud giant. Watch whether the EU AI Office converts any of its information requests into a first real fine or corrective order, and watch how the Sony Music Publishing/Warner Chappell suit against Anthropic and the DOJ's fair-use intervention in the NYT case against OpenAI develop as the sector's two clearest tests of whether training-data litigation ends in settlements or in case law. And watch whether global AI infrastructure spending actually reaches the roughly $800 billion pace now projected for 2026, the scale every individual funding round in this sector is now just one piece of.
Sources
- NBC News - OpenAI debuts GPT-6 Astra, says it triggered security measures
- MarketScreener (The Information) - Anthropic plans to publicly unveil IPO prospectus after Labor Day
- Business Standard (Reuters) - Chinese AI firm Moonshot files confidentially for Hong Kong IPO
- Music Business Worldwide - Sony Music Publishing and Warner Chappell sue Anthropic
- Tokenstead - EU AI Act enforcement begins: the AI Office starts asking
- Crunchbase News - Global startup investment hit record $510B in H1 2026 as AI boom accelerates funding
