Three labs landed on $2/$10 this week. The differences moved to the fine print.

Three major AI labs spent this week landing on the same price. Anthropic's Claude Sonnet 5.5 (September 28), OpenAI's GPT-6.1 Sol (September 29) and the introductory rate for Google's Gemini 4 Argon (September 30) all list at $2 per million input tokens and $10 per million output tokens. Two of those are new mid-tier models pitched as near-flagship; the third is Google's new frontier model at a launch discount. Price stopped being the thing that separates them. What separates them now is in the fine print: API behavior, speed tiers, output limits and who gets access first.

Same price, different migration bills

The headline numbers are easy to compare. The switching costs are not.

• Sonnet 5.5 kept Sonnet 5's price and, by Anthropic's own testing, generates output 30%+ faster and costs up to 30% less per task. But its migration notes list breaking changes: forced tool use (`tool_choice` of `any` or `tool`) and `thinking: disabled` now return a 400. The common trick of forcing a single tool call to get JSON stops working.
• GPT-6.1 Sol matches GPT-6 Sol's $2/$10 but halves cached input to $0.10, per OpenAI's pricing page. OpenAI says it performs close to GPT-6 Astra, which costs $10/$50; that claim rests on OpenAI's own evaluations. The same day, OpenAI's changelog added an Ultrafast service tier for Astra that cuts time between output tokens, priced at $60 input and $300 output per million, six times standard.
• Gemini 4 Argon is the outlier. Google's announcement sets an introductory $2/$10 that rises to $4/$20 after an unspecified period, and a 1 million-token output limit. But it is rolling out first to cyber defenders in Google's Fairwind Program. Google gave no date for developer access.

Put together: the per-token rate is no longer where the decision lives. Total cost per task depends on how many tokens a model burns, how well your prompt prefixes hit the cache, and whether you pay extra for latency. Teams that copy settings across models (forced tool calls, effort levels, output caps sized for 64K) will find the breakage before they find the savings. The only reliable comparison is a run of your own tasks on each model.

Guardrails moved outside the model, and outside the law

A second thread ran through the week: who enforces limits on increasingly autonomous systems, and how.

NVIDIA's answer, on September 28, was architectural. Its Open Agent Safety Platform pairs OpenShell, an open-source runtime that sandboxes agents and checks their actions against policy, with Sentry, a watchdog on BlueField-4 DPUs that NVIDIA says can quarantine an agent stepping outside its boundary in milliseconds. The premise is that the model and its harness shouldn't be the last line of defense.

Washington's answer, on September 29, was a pledge. According to TechCrunch, President Trump and leaders of major AI companies signed a "Joint Commitment on Frontier Responsibilities" calling for internal controls and independent oversight of frontier models. It is voluntary, with no apparent legal consequences for breaking it. TechCrunch also reports an accompanying executive order directing agencies to say "superintelligence" or "SI" instead of "artificial intelligence." Promptea could not open the White House's own text from our research environment, so we are not characterizing its provisions beyond that reporting.

Google, meanwhile, said trusted defenders in Fairwind and its internal teams will get Argon without cyber guardrails. That's the same question approached from the other side: capability released to vetted users with fewer limits, everyone else waiting. Read together, the week's guardrails are mostly self-imposed, whether in code, in hardware or in a signed pledge.

Where the money and the hardware went

• AMD agreed to buy World Labs, Fei-Fei Li's world-model startup, for $8.2 billion, TechCrunch reported on September 28. Li becomes AMD's EVP and chief scientist; the deal is expected to close by year-end, subject to regulatory approval. A chipmaker buying a model lab says workloads now shape silicon roadmaps.
• Barclays said it expects Claude Code to reach 50% of its developers by the end of 2026, per Anthropic's announcement. That is a target. The live deployments are an internal assistant used by 16,000+ staff and a pipeline routing about 120,000 emails a day.
• Anthropic committed $100 million on October 2 to a Claude Frontier Academy aimed at training 10,000 deployment engineers by the end of 2027, with partners including Accenture, Deloitte and McKinsey. The bottleneck it's paying to fix is people who can put models into production, not models.
• Ai2 released Olmo-core 3, an open MoE training stack it benchmarked past a trillion parameters (with random routing, so a systems test, not a trained model).
• NVIDIA announced a 64GB DGX Spark for local AI, from $4,999 through six PC makers starting October 23.

What it adds up to

The cost of frontier-adjacent intelligence fell into one band this week, and the labs are now competing on behavior, latency and access rather than list price. For anyone choosing a model, that shifts the work from reading pricing pages to running evals and reading migration guides. On the governance side, the week produced more enforcement mechanisms that companies apply to themselves than obligations anyone can enforce on them.