One API for agents, $664 billion of obligations, and a public call to slow down

Three things happened this week that were not responses to each other and still read better together. OpenAI shipped the machinery that turns a frontier model into something others can build products on. Oracle published the clearest public number yet on what the infrastructure underneath costs to run. And two researchers, then the chief executive of Anthropic, argued under their own names that the rate of capability gain needs to come down. The common subject is control: who gets access to frontier AI, on what terms, and how fast it improves.

OpenAI spent the week on plumbing, not models

GPT-6 Astra itself landed on 3 September, just outside this window. What fell inside it was the support structure. On 8 September OpenAI made Prompt Cache Diagnostics generally available in the Responses API for GPT-5.6 and later, letting you compare cache reuse against a previous response and get a stated reason for a miss. The same day brought two image models, GPT Image 2.5 Sunburst and Flare, and GPT-Rosalind for life sciences research through a trusted-access programme.

The 10 September entries matter more. OpenAI released the Agents API in public beta, described in its own documentation as access to "the Codex harness through an OpenAI-managed API": OpenAI runs session orchestration, context compaction and recovery, while your application supplies tools and picks where the agent executes, either in an OpenAI-hosted sandbox or one you connect. The same day GPT-Live 1 reached general availability for full-duplex voice at $0.05 per minute, billed per second, with backend model and tool usage charged separately. Project API keys also gained expiration dates, with maximum key lifetimes enforceable at organisation or project level.

The throughline there is not capability. It is that the parts of agent-building that are tedious and stateful — compaction, session recovery, sandboxing — moved out of your codebase and into OpenAI's, with a meter attached. Metering showed up at the consumer end too: Apple's rebuilt Siri, which we covered on Wednesday, arrives with daily caps on server-side features and more access to be sold later.

What running it actually costs

Oracle reported first-quarter fiscal 2027 results on 10 September. Revenue rose 30% to $19.3 billion, with cloud infrastructure up 121%. Oracle said it booked more than $30 billion of additional AI cloud contracts in the quarter, taking remaining performance obligations to $664 billion, and delivered more than 300,000 GPUs since the end of the prior quarter alongside 850MW of new datacentre capacity — close to triple the capacity it added in the quarter before.

The same release carries the other half. Operating cash flow reached a record $23 billion, up 184%, while free cash flow was negative $5 billion. Oracle completed a $20 billion at-the-market stock sale during the quarter and guided to at least $90 billion of revenue for the full year. A record backlog and negative free cash flow are not in tension; they are one sentence read from both ends. The demand is contracted and the capacity to serve it is still being bought, with equity, in advance.

Google's 9 September commitment of 13 billion euros to Finland over two years — its largest single investment in Europe — came with the detail that usually gets skipped. Alongside the data centres sit a 22-year agreement supporting the life extension of the Loviisa nuclear plant, new onshore wind, and a contracted 94-megawatt battery system to stabilise prices in cold, windless periods. Data centre announcements are energy announcements now, and the companies have stopped pretending otherwise.

And then, the brakes

On Tuesday evening, Anthropic researcher Jacob Coxon posted that he was resigning. As TechCrunch reported on 9 September, Coxon — who said he spent three years on pretraining research at OpenAI and then Anthropic — wrote that the labs "are racing straight to self-improving superintelligence and gambling with our lives". His Anthropic colleague Evan Hubinger did not walk it back, saying the team does "earnestly believe AI could kill all humans", putting the odds above 10% within the decade and acknowledging that Anthropic does not "have a plan to solve alignment for superintelligence". Anthropic did not immediately respond to TechCrunch's request for comment.

Today, 12 September, Anthropic chief executive Dario Amodei published an essay titled "We Must Pace the Frontier". A disclosure on sourcing: the essay is hosted on Amodei's personal site, which is unreachable from our environment, so we could not open the primary document. What follows is TechCrunch's account of it, not our own reading. By that account, Amodei laid out three strategies and said Anthropic is "unilaterally committing" to the first — embedded third-party evaluators from organisations such as METR, given company badges, desks and access "mostly comparable to what internal risk assessment teams have". He compared the arrangement to regulators stationed inside the banks they supervise. Sam Altman responded that he agrees "we need to pace the frontier" and said OpenAI would do the same.

Keep the scepticism attached to it. Amodei also asked the US government for a narrow antitrust waiver so labs can hold safety conversations, and argued that export controls plus a crackdown on distillation could widen America's lead over China by three to five years — positions that happen to suit Anthropic. Journalist Brian Merchant, quoted in the same piece, said proposals like it "would likely only wind up serving Anthropic and OpenAI". The part worth tracking is the falsifiable one: embedded evaluators either get badges and publish, or they do not.

Anthropic's own threat intelligence report, published 10 September, sits between the week's two halves. It catalogues seven categories of misuse of Claude, including a dedicated section on illicit distillation. That report is about who gets to use frontier models; Amodei's essay is about how fast they improve. They arrived four days apart.

What changes for you

• If you maintain a hand-rolled agent loop, the Agents API now rents you OpenAI's instead. You give up portability and gain someone else's compaction and session recovery. Note the billing shape: model rates plus tool rates plus container rates, not a single line item.
• GPT-Live 1's $0.05 per minute buys the voice session, not the intelligence. Whatever backend model does the reasoning bills on top of it.
• Prompt Cache Diagnostics is the first time OpenAI will tell you why a cache missed rather than only that it did. If you are on GPT-5.6 or later and your hit rate is worse than you modelled, start there.
• GPT-6 Astra has a 1,050,000-token context window, but prompts above 272,000 input tokens are billed at 2x input and cache rates and 1.5x output for the entire request. The big window and the cheap window are not the same window.
• Astra also removes the `none` reasoning effort level, rejects custom temperature, top_p and logprobs, and requires the Responses API for tool calling. A migration off Chat Completions is not a drop-in swap.

Two motions, one week. Frontier capability got easier to rent by the minute and more expensive to own by the gigawatt, and the people closest to it began asking out loud for the rate to come down. Whether pacing turns out to mean evaluators with desks and publishing rights, or a phrase that survives a single news cycle, is a question the coming months answer rather than this one.