Anthropic spent September 1 cutting the price of running Claude at scale rather than announcing a smarter model. The same two weeks brought a near-doubling of AI coding startup Cognition's valuation and a real, metered government contract for OpenAI's ChatGPT. None of the three stories is about what a model can do. All three are about what it costs to run one, and who is now willing to pay for it on commercial terms.
Anthropic repriced Claude for agents, not benchmarks
On September 1, Anthropic released Claude Fable 5.1 to general availability and Claude Mythos 5.1 to a smaller set of vetted partners in cybersecurity and life sciences, alongside new privacy safeguards for business customers (Axios). The headline change was not the models themselves. Anthropic cut the price of cache reads, the mechanism agents use to reference information they have already processed, by 75%, while leaving base token pricing at $10 per million input tokens and $50 per million output tokens (Forbes).
That single change lowers typical bills by roughly a quarter and cuts costs for agent-heavy workloads by up to 45%, according to Anthropic. Any enterprise running a Claude-based coding assistant or long-running workflow agent sees the reduction without switching models or renegotiating a contract.
The cut also lands against a competitive backdrop. Anthropic confidentially filed a draft S-1 with the Securities and Exchange Commission on June 1, 2026, days after a $65 billion funding round valued the company at $965 billion (Fortune). Making Claude cheaper to run makes it easier to grow the same enterprise revenue base investors will scrutinize whenever that filing becomes public. For a company selling access to frontier intelligence, the competitive move that mattered most this month was a bill, not a benchmark.
Cognition's valuation nearly doubled in four months
Cognition, maker of the autonomous coding agent Devin, closed a $2 billion Series E on September 8 at a $48 billion valuation, up from $26 billion just four months earlier (TechCrunch). The round was led by Andreessen Horowitz, Accel, Founders Fund, General Catalyst and Avenir. Cognition said its annualized run-rate revenue grew from $492 million in May to roughly $900 million by the close of the round (SiliconANGLE).
Revenue and valuation grew at almost the same pace, so the implied revenue multiple barely moved. That is a different signal than the valuation jump alone suggests. Investors are not paying up for a story here. They are paying a steady multiple against genuinely growing enterprise and developer spend on a coding agent, which is a narrower and more testable bet than backing a general-purpose model lab.
Washington signed a metered contract, not a pilot
The US General Services Administration and OpenAI agreed to a 27-month OneGov deal on September 10, replacing the nominal $1-per-agency pricing that was set to expire September 30 (GSA). Starting October 1, the new agreement cuts token-based usage costs by 50%, drops the standard $15-per-user monthly license fee, and sets no minimum spend or platform-access charge (Bloomberg). The deal runs through December 31, 2028.
OpenAI said more than a million federal employees already use ChatGPT under the outgoing arrangement, and the new deal extends eligibility to roughly 23 million government workers (FedScoop). The shift from a symbolic dollar to a discounted, usage-based government contract with a multiyear term is the clearer signal here. Federal procurement is treating ChatGPT as a budgeted line item rather than a pilot program that needed a token price to justify itself.
What ties these together
Each story lands on the same point from a different direction. Anthropic cut what customers pay to run agents continuously. Cognition's investors priced a coding agent company on revenue growth that kept pace with its valuation, not on model capability claims. The US government moved from free access to a metered contract with a fixed multiyear term. Enterprise AI adoption is no longer being sold on what a model can theoretically do. It is being priced, budgeted and renewed like any other piece of core infrastructure.
None of the three developments proves the economics are settled. Anthropic's own price cut lowers its per-agent margin even as revenue grows. Cognition's flat revenue multiple only holds if the company hits the run rate its investors are underwriting. And a federal discount that starts October 1 has not yet been tested against real agency usage.
What to watch next
Two things are worth tracking over the next two quarters. Whether Anthropic's confidential S-1 becomes a public filing, which would show for the first time whether real revenue supports a valuation near $965 billion. And whether Cognition's own investor projection, an annualized revenue run rate of $4 billion to $5 billion by year-end, holds up against results rather than forecasts.