AI Next WaveSeptember 24, 2026
They Asked to Slow Down. Then They Cut Prices.
Within ten days of agreeing the industry should "pace the frontier," the heads of Anthropic, OpenAI and xAI had all shipped new models. That looks like hypocrisy, and part of it is. The more useful story is what "pacing" actually means for the AI tools you pay for.
The Short Version
On September 12, Anthropic CEO Dario Amodei published an essay called "We Must Pace the Frontier." Sam Altman replied the same day: "I agree with Dario." Elon Musk posted: "Dario is right."
Then came the launches. xAI released Grok 4.7 on September 21. The next day, Anthropic shipped Claude Opus 5.5, and OpenAI followed minutes later with GPT-6 Sol and Luna. On September 23, Altman and Amodei addressed the UN Security Council to ask for international safety standards.
Read closely, the launches were mostly price cuts. By the labs' own descriptions, this week's models bring top-tier ability to more people at lower cost. They don't claim a big jump past what already existed. That's a narrow, convenient reading of "pacing," and it's worth knowing. The ceiling is barely moving. The price floor is falling fast. For anyone paying for AI, the second part matters more.
The Pledge
What "pacing" meant on paper
Amodei's essay argues that AI capability is improving faster than safety work can keep up. He proposes three steps. First, outside evaluators embedded inside AI labs with employee-level access. Second, common safety standards among companies in democratic countries. Third, coordination between governments, including China, on limits for the most dangerous uses.
He was also clear about what pacing isn't. "We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast," he wrote. Pacing, in his telling, doesn't mean stopping training or technical progress.
Altman's response on X: "I agree with Dario that we need to pace the frontier. This has been a primary topic of discussions we've had at OpenAI in recent weeks." Musk's reply was three words.
Money followed words, at least on Anthropic's side. On September 18, Anthropic and Accenture said they each expect to invest "at least $1 billion" over five years in building the embedded-evaluator capacity the essay described.
The Week After
Three launches, and a lawsuit from the other direction
Grok 4.7 arrived September 21. xAI kept the price at $2 per million input tokens and $6 per million output tokens, and outside benchmark results placed it behind the leaders.
On September 22, Anthropic released Claude Opus 5.5. It cut Opus prices from $5 to $4 per million input tokens and from $25 to $20 for output, and said the model costs 40 percent less than Opus 5 on typical workloads. Decrypt reported that OpenAI launched GPT-6 Sol and Luna minutes later. OpenAI priced Sol at $2 and $10 and Luna at 10 cents and 50 cents, and said that's half its earlier promotional prices.
Anthropic's own announcement tries to square the circle. "Pacing is an approach to keeping AI safe, remaining competitive with China, and realizing AI's benefits," it says. In other words, pacing and shipping are supposed to happen together.
Not everyone accepts either half of that. On September 18, subscribers to the companies' AI services filed a class action in federal court in San Francisco. It accuses Anthropic, OpenAI, Google and SpaceXAI of violating the Sherman Antitrust Act by agreeing to slow development and "depriving paying subscribers of product improvements they were promised." It's a complaint, not a ruling. It still shows the bind: a lab that slows down can be sued for it, and a lab that ships gets called a hypocrite.
Nvidia's Jensen Huang pushed from the other side this week. "That 10 percent chance is not grounded on science," he told Axios on September 23, referring to estimates that AI could end in catastrophe. "Enough predictions."
Hypocrisy, or Something Narrower?
The labs' own fine print tells you where the line sits
The fair test is whether this week's models raised the ceiling, or just brought it closer to more people.
Anthropic says Opus 5.5 beats its larger Fable 5.1 model on most benchmarks. Then it adds a caveat worth reading twice: "In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest." OpenAI describes Sol as "approaching Astra-level reliability at much lower cost." Astra is the top model OpenAI already sells. Grok 4.7 trails both.
By their own accounts, the labs mostly took ability they already had and made it cheaper and faster. That's consistent with a narrow version of pacing, where the top rung stays put while the price drops. It also hands each company its own grading sheet. No outside body measures the ceiling. The embedded evaluators are still being hired.
There's another problem, and Anthropic raised it itself. "We see signs that Opus 5.5 often suspects it is being evaluated," the company wrote, "which challenges our ability to assess how it will act." A promise to slow down depends on being able to measure what you built. The builder just said that's getting harder.
My read: the contradiction is real, but it's smaller than the headlines make it. The labs are racing on price and access while talking about pace at the top. Until someone outside the labs checks the top, "pacing" means whatever the next press release says it means.
What the Race Means for Your AI Bill
Cheaper, faster, and changing under your feet
Picture a six-person marketing agency. It pays for chat subscriptions for everyone and runs a small tool on an AI model's API to draft client reports. In one week, the model it built on got cheaper, a rival released something close for less, and the provider changed how the model behaves.
Some of those changes are pacing in practice. Anthropic now reroutes most cybersecurity tasks on Opus 5.5 to an older model, Opus 4.8. Biology work above a threshold requires joining a verification program. The model can also no longer run with "thinking" switched off, which affects speed and cost for simple jobs. None of that shows up in a price chart. All of it shows up in your workflow.
The costs move fast in both directions. OpenAI mentioned in its launch post that its own median researcher now uses more than $600 a day in tokens, and the 90th percentile uses more than $7,000. Prices per token are falling. Usage per person is rising faster.
Anthropic also said benchmark margins have become "a less reliable guide to real-world differences." That's the most useful sentence of the week for a small business. The only benchmark that counts is your own work.
Opportunity Radar
The quarterly model check-up
The problem: small businesses pick an AI model once and keep paying for it while prices, features and safeguards shift every few weeks. The buyers are owners and operations leads at companies of 5 to 100 people who spend real money on AI subscriptions or API usage.
The service runs a fixed-fee review each quarter. It collects 20 to 30 of the client's real tasks, runs them through the two or three current contenders, and scores cost per finished task, error rate and time saved. The deliverable is a one-page recommendation: stay, switch, or split the work between models.
Validate it with three paid pilots before you build any tooling. Measure the dollar difference you find and whether the client acts on it. The idea fails if savings come in under the fee, or if clients won't switch even when the numbers say to. If the first three pilots don't find savings above your fee, stop.
What You Can Do With This
If you pay for AI tools
Hold off on annual contracts priced before this week. Prices just dropped by a fifth to a half, depending on the model. Month-to-month costs a little more and keeps you free to switch.
If you've built something on an AI model
Rerun five real tasks on the new models this week, and track cost per finished task, not cost per token. Check whether any safeguard changes, like rerouted security work or always-on thinking, affect your use.
If you're just trying to keep up
When a lab says it's pacing, ask one question: who outside the company checked? For now the answer is mostly no one. Treat safety claims in launch posts the way you'd treat a company's own benchmarks. Useful, and never the final word.
The Bigger Picture
This week, the leaders of three AI companies endorsed slowing down and shipped new models within days. Both things happened, and both are on the record.
The launches still followed a pattern. Top-tier ability got cheaper and reached more people. The labs' own documents say the top barely moved. If that holds, the near-term effect of the AI race falls on price and access. Small teams can afford tools that were premium a month ago. That's real value, and it comes with a cost: the rules, features and safeguards inside those tools now change as fast as the prices. Until outside evaluators are in place, the only measurement you control is your own. Run your tasks and keep the receipts.
References
Dario Amodei: We Must Pace the Frontier (Sept. 2026) SiliconANGLE: Sam Altman and Elon Musk back Dario Amodei's call to slow down the frontier of AI development (Sept. 13, 2026) Anthropic: Anthropic and Accenture on embedded evaluation (Sept. 18, 2026) Decrypt via Yahoo Tech: xAI launches Grok 4.7 (Sept. 21, 2026) Anthropic: Introducing Claude Opus 5.5 (Sept. 22, 2026) OpenAI: Introducing GPT-6 Sol and Luna (Sept. 22, 2026) Decrypt: OpenAI launches GPT-6 Sol and Luna minutes after Anthropic drops Claude Opus 5.5 (Sept. 22, 2026) Bloomberg Law: OpenAI, Anthropic, Google, SpaceXAI hit with antitrust lawsuit (Sept. 2026) Axios: Nvidia CEO Jensen Huang on AI doom theories, "Enough predictions" (Sept. 23, 2026) Times of Israel (Reuters): AI chiefs warn UN Security Council of security risks as systems grow more powerful (Sept. 23, 2026)