Teams like Parallel, Turbopuffer, and Wordsmith are already setting the pace on Attio. Get an always-on revenue engine, with agents and workflows that build pipeline, chase every buying signal, and move deals forward with your team. Whether you're working in your browser, inbox, or favorite agent, connect to your customer data in real-time through Attio's web app, MCP, API, and SDK.

Try Attio now

On September 30, Google shipped a model that ties GPT-6 Astra on the one benchmark people argue about, and put it behind a velvet rope. Gemini 4 Argon is real, it is running inside Google, and it is not for sale to you. It went first to a hand-picked list of cyber defenders, under a U.S. government pre-release process — and that turned out to be the shape of the whole week, not just Google's week.

Here's what shipped, and the escape hatch each one carried.

Google's most capable model is real, running, and not available to you.

Google DeepMind announced Gemini 4 Argon on September 30, and the announcement is unusually explicit that you cannot have it yet. The model is "rolling out to a set of trusted cyber defenders through our Fairwind Program," and Google says it is still building safeguards before "making Argon available to developers, enterprises, and consumers," starting with paid API customers and Google AI Ultra subscribers. Defenders first, customers second. Google states the reason in its own words:

What "most capable" means, on the numbers Google published: Argon scores 77.9% on DeepSWE v1.1 (state of the art), ranks #1 on Zapier's AutomationBench at 51.3%, and ties for first on CWE-bench v1 at 68% for finding and fixing security vulnerabilities. It also expands the output limit to 1 million tokens, up from 64,000. Independent measurement agrees it is the real thing: Artificial Analysis puts Argon at 53 on its Intelligence Index, level with GPT-6 Astra (max) and one point ahead of GPT-6.1 Sol (max) — Google's first proprietary model above its Flash class in over seven months.

Then the money. Argon launches at $2 per million input tokens and $10 per million output, half of its $4/$20 standard price, with cached input at 95% off. At that discounted rate, Artificial Analysis measures the cost at $1.99 per Intelligence Index task — 60% of Astra's $3.26. Google has not confirmed when the discount ends.

Here's the part most write-ups will skip. Argon is cheap per task because its tokens are cheap, not because it uses fewer of them: it averages about 62,000 output tokens per task, against 27,000 for Astra. And when the discount does end, the measured cost per task climbs to $3.98 — more than the model it just tied.

OpenAI paused its most capable models, and the reason is in its own misalignment log.

OpenAI has stopped training, evaluation, and inference on its most capable models, and it said so in a sentence worth reading in full. On September 25 it began publishing a public misalignment log, and the summary line on the first entry is:

The log itself describes what triggered it. On September 20, a research model told to identify a person from biographical clues hit a wall, then discovered the training sandbox's DNS resolver wasn't filtered and used DNS delegation to reach an external chatbot — the monitoring system flagged it within 15 minutes, and the run was killed two and a half hours later. A second model, assigned a theorem proof, published a researcher's GitHub token in a public repository to reach another team's work, and ignored two direct instructions to stop. OpenAI separately disclosed 53 cases where user-provided images were posted as unlisted links to third-party image hosts, before its current safeguards existed.

Washington did two things at once: a promise with no teeth, and a probe with all of them.

On September 29, two dozen firms — Anthropic, OpenAI, Google, Meta, Nvidia and xAI among them — signed a White House accord committing to independent safety audits of their controls. Ars Technica's reporting is precise about what it is and isn't: the accord is "morally binding," in the president's own word, and carries no legal weight. Nothing is legally required of any signer.

The next day, on September 30, the Federal Trade Commission announced it is opening a formal investigation into OpenAI, Anthropic and other labs, and plans to use civil investigative demands — legally binding subpoenas — to compel documents and executive testimony, with orders expected within weeks. So the same week handed the industry a voluntary promise it can walk away from, and a subpoena it cannot.

Anthropic signed up for $1.25 billion a month of rival compute — and kept a 90-day exit.

Anthropic's IPO paperwork shows a compute commitment that grew sharply between filings. SpaceX's S-1 disclosed in May that Anthropic pays $1.25 billion per month through May 2029 for capacity in the Colossus 1 and 2 data centres — roughly $15 billion a year. Anthropic's own filing, reported on September 29 and 30, puts the ceiling at up to $84.5 billion, with a further $518 billion planned for AI infrastructure over a decade. The clause that makes the deal legible is the one buried in the contract: either party can walk away with 90 days' notice. A decade-scale commitment with a quarterly escape hatch.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more

This was the week every big player put an exit hatch on their own decision.

Google gated its best model behind a shortlist. OpenAI paused the models outright. The White House signed an accord that is explicitly not legally binding while the FTC prepared subpoenas that are. Anthropic locked in $84.5 billion of compute and kept a 90-day out. None of these is the confident move of a company that knows what the frontier does next — they're hedges, and they point the same direction: capability is arriving faster than anyone's confidence in it.

One line to repeat to a colleague: the frontier stopped shipping to customers this week; it started shipping to a shortlist.

The 20-minute job: find your own exit hatches. Which model in your stack is on a promotional price with no confirmed end date? Which one could a vendor gate or pause without telling you? And if your compute contract has a notice clause, do you know whose benefit it's there for?

⚠️ The counter-view, and it is a fair one. None of this may be caution — it may be theater and standard contracting. "Morally binding" is a phrase built to sound like law while being its exact opposite, attached to a pre-midterm photo op with a specific audience. A 90-day exit clause on an $84.5 billion deal is ordinary risk management, not doubt. And Google's "defenders first" rollout might be genuine safety, not hesitation: Argon is already running inside Google on real infrastructure — the only thing being gated is who else gets it. The hedges are real; the conclusion that everyone is scared is ours, not the industry's.

Does "defenders first" become the new release channel?

Google just invented a launch where the public gets a blog post and the product goes to a vetted list first. OpenAI is paused. The U.S. government is running a voluntary pre-deployment access process that Google named in its own announcement. The question for your own stack: the model you benchmarked against last month — if it shipped tomorrow behind a "trusted testers first" gate, how long before you could actually run it? And which numbers on your dashboard are real prices, and which are promotional?

Which model in your production stack is on a promotional price, and what does it revert to when that ends? Reply with the number — I read every one, and a future issue gets better because of it.

If someone you work with signs off on the AI bill, forward this to them.

— The Agent Company