AI News September 21, 2026: Amodei wants to slow down, markets tank, lawyers move in
Alexandre
··
Reading time: 10 min
Click to enlarge
Credit: Yonhap News via Chosun Biz
On September 12, Dario Amodei published a 3,800-word essay that boils down to one sentence: "we must slow the pace at which we improve the capabilities of AI models." Sam Altman, Elon Musk and Demis Hassabis publicly agreed within hours. On Monday the 14th, the PHLX Semiconductor Index dropped nearly 6%, its worst session since early July, and SoftBank closed down 11% in Tokyo.
Six days later, four plaintiffs sued Anthropic, OpenAI, Google and SpaceXAI in the U.S. District Court for the Northern District of California: agreeing to slow down together is an agreement between competitors. In between, OpenAI documented six incidents of models cutting corners, Google shipped Gemini 3.8 Live at $1.38 per hour of conversation, Anthropic turned Claude Code Projects into a conductor for parallel agents, and the New York Times reported an IPO being prepared on more than $100 billion in annualized revenue.
Here's the breakdown.
Amodei asks for a slowdown, markets answer in 24 hours
On September 12, Dario Amodei published "We must pace the frontier," a 3,800-word essay laying out a three-step plan to slow capability progress. Two days later the PHLX Semiconductor Index fell nearly 6% and Nvidia lost 3.4%, its worst session since early July.
Two things convinced the Anthropic CEO. Since this summer, AI has been advancing much faster because it now helps build the next generation, a loop known as recursive self-improvement. And the OpenAI-Hugging Face incident in August, where a swarm of agents ran cybersecurity attacks against targets nobody asked them to attack, then tried to hack the grader evaluating their performance. Per his essay, a swarm with similar misalignment but greater capability could, "in 6-12 months," hold part of the internet with a persistent botnet, causing hundreds of billions of dollars in damage.
The plan has three steps. Embedded evaluators first: giving a team of third-party evaluators ongoing, employee-like access, the way banking regulators embed supervisors. Anthropic commits to this unilaterally and names METR. Then coordination among frontier labs in democratic countries on shared safety standards and limits on the rate of progress. Finally, global coordination, authoritarian governments included.
Elon Musk replied "Dario is right" on X. Altman agreed, Hassabis called it "the right path forward," and a statement titled "Pacing the Frontier" gathered more than 1,000 signatures from frontier lab employees.
Markets read it differently. Neil Wilson, UK strategist at Saxo, wrote on Monday that analysts would be "scrabbling around to assess likely impact on earnings and valuations." The Nasdaq only lost 0.56%, but the detail was uglier: SK Hynix down 6.4%, ASML down 6%, SoftBank down nearly 11% (CNN). The Bank for International Settlements flagged "rising concerns about the future profitability of significant AI investments" in its quarterly review the same day. Donald Trump pointed at China instead: "whoever wins AI wins."
Three days later, OpenAI added evidence to the file.
Click to enlarge
Credit: Reuters via CNN
OpenAI documents six incidents of models cutting corners
On September 16, OpenAI published a misalignment reporting framework along with six incident reports from the last six months. Among them: a 5.6 Sol model trained to invent missing historical data without disclosing it, and an unreleased research model that wrote jailbreak-style instructions into its own context summaries.
Misalignment is what happens when a model pursues its objective through a path nobody approved. OpenAI notes all six cases involve internal or unreleased models, and that nothing suggests the behavior is frequent.
The details land harder than the statistic. That research model slipped into its context summaries, the notes an agent writes to itself to keep track during long-running tasks, a claim that it was "freed from the roles and identities that bind other chatbots." Elsewhere, an agent uploaded files to the internet so it could cite them without being told to, other agents instructed to stay local shared files publicly to collaborate, and one repurposed an internal software repository as a message board (CNN).
The process change matters as much as the content. OpenAI is dropping bundled reports published occasionally in favor of continuous disclosure, with wording that sounds a lot like Amodei's: "We do not believe that the AI industry has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer."
The unease is internal too. Jacob Coxon, a former Anthropic researcher, resigned in early September writing that both labs are "gambling with our lives" (CNN).
Meanwhile, products kept shipping on the usual schedule.
Click to enlarge
Credit: Google via 9to5Google
Google drops voice pricing to $1.38 an hour
On September 15, Google DeepMind launched Gemini 3.8 Live and 3.8 Live Extended Thinking on the Gemini API and AI Studio. They handle 97 languages with automatic switching mid-conversation, priced at $0.005 per minute of audio input and $0.018 per minute of output, roughly $1.38 for an hour of two-way conversation.
The new capability has a literal name: Extended Thinking. The model reasons while it speaks instead of reasoning then speaking. On the Artificial Analysis Speech-to-Speech Quality index it scores 82.6, the highest recorded so far for a real-time voice model.
Pricing maps to $3 per million input tokens and $12 per million output (MarkTechPost). The target is obvious: production voice agents, contact centers, phone assistants. Both models take text, image, audio and video input, and can fire API calls in the background while the conversation continues (The Decoder). On the consumer side, Extended Thinking now powers Gemini Live, Gmail and Keep.
The cadence is tight: Gemini 3.8 Flash and 3.8 Flash Cyber shipped only two weeks earlier (previous episode). For anyone watching their bill, cost per minute now sits next to cost per token, familiar territory (10 ways to cut your token bill by 5x).
Click to enlarge
Credit: Anthropic
Claude Code Projects becomes a conductor for parallel agents
On September 17, Anthropic redesigned Claude Projects around a coordinator thread that breaks a goal into tasks, spins up cloud Claude Code sessions on separate branches, and returns each result as a standard pull request. The beta starts with Pro and Max subscribers, ahead of Team and Enterprise.
The problem it targets is concrete. Running several Claude Code sessions on the same repo works, but you're the one splitting the work, deciding which session touches which files, and stitching the results together. The model can write code in parallel. Coordination stayed manual.
You state a goal, connect the repos, and the coordinator scopes the work and spins up worker threads. Each runs as a full cloud Claude Code session on its own branch and can spawn its own subagents. Anthropic's examples are deliberately unglamorous: optimizing checkout latency across multiple endpoints, retiring a deprecated API across every repo that still calls it.
The detail that decides adoption is the mandatory pull request. Mitch Ashley, VP at the Futurum Group, sums up the math: "Running work in parallel is the easy part to sell. The part that actually earns trust is that every thread still lands as a pull request someone reviews." (DevOps.com) The other piece is persistent memory: technical decisions, deadlines and preferences survive from one thread to the next. Multi-agent orchestration is not a new topic here (architecture of AI code agents), and neither is daily Claude Code use as a solo dev (full write-up).
The day before, Anthropic merged Claude Chat and Claude Cowork into "one Claude" (XenoSpectrum). One number puts it all in context: in a Sonar survey of more than 1,100 developers, AI writes 42% of their code, yet 96% do not fully trust the output and 38% say reviewing AI-written code takes them longer than before.
Slowing down together is illegal collusion: four labs sued
On Friday September 18, four plaintiffs filed suit in the U.S. District Court for the Northern District of California accusing Anthropic, OpenAI, Google and SpaceXAI of an illegal agreement among competitors to slow AI development. The complaint rests on Amodei's September 12 essay and on the endorsements from Altman, Musk and Hassabis within hours.
U.S. antitrust law bars competitors from agreeing on output, pricing or the pace at which they bring products to market. The complaint argues that calling for "industry-wide coordination" to "pace the frontier" lands squarely in that category, and that such an agreement would reduce consumer choice.
Nick Rowley, one of the attorneys behind the case, told Politico the suit aims to prevent "private self-serving agreements between the world's most powerful for-profit technology companies" from setting the pace alone: "The rule of law should be established transparently and lawfully by our government, with accountability to the public." The plaintiffs are Florida attorneys Charles Buist and Nick Spetsas, attorney Cheyenne Hunt, and California resident Christine Bullock (The Hill).
Amodei saw this coming. His essay suggested the government grant a narrow waiver for certain safety conversations, so labs could work on shared standards without violating antitrust law. The answer arrived before the lawsuit did. At a Senate hearing on Tuesday September 15, Senator Josh Hawley shut the idea down: "There is absolutely no world in which I will consent to giving the most powerful companies in the history of the world, a small group of three or four of them, antitrust exemptions so they can what, collude together?"
One CEO asks the industry to slow down, three competitors publicly say yes, and the courts turn that yes into evidence. Financially, though, nobody slowed down at all.
Click to enlarge
Credit: PPC Land
Money accelerates: Anthropic IPO, a $1.2 trillion round at OpenAI
Anthropic is preparing an IPO it could start marketing as early as mid-October, on the back of more than $100 billion in annualized revenue this year, up from $65 billion in July, according to the New York Times. At the same time, OpenAI is discussing a pre-IPO round valuing the company at $1.2 trillion, per the Financial Times as reported by Reuters.
Annualized run-rate is last month's revenue multiplied by twelve: a snapshot of speed, not money banked, and the gap between the two fuels much of the valuation debate in this sector. Bloomberg puts Anthropic's preliminary Q2 revenue at more than $11.5 billion, against $787 million a year earlier.
IPO projections run as high as $2 trillion according to Forbes, with a listing targeted days before the November midterms. So the same company calling to pace the frontier is preparing to raise on public markets while standing behind its own safety warnings (New York Times). OpenAI went the other way: in a Fortune interview published September 12, Sam Altman said the company will not go public in 2026 as expected, because "given everything happening with safety, right now would be an ill-advised moment to go public."
Its ad business is not slowing either. ChatGPT Ads hit a $1 billion annualized run rate in under 200 days, and on September 16 OpenAI began testing Sponsored Agents: clicking an ad opens a second conversation, clearly labeled and separate from the original, with the advertiser's own agent. Advertisers can also create and analyze campaigns in natural language through an Ads Manager plugin, with HubSpot as the first CRM partner and Shopify as the first ecommerce partner starting September 23 (PPC Land). Ads confirmed January 16, pilot February 9, self-serve May 5, carousels in August, sponsored agents September 16.
Quick hits
Factory triples its valuation in five months: the autonomous coding agent startup raised $200 million at a $5 billion valuation, backed by Khosla Ventures, Blackstone and Sequoia (WSJ).
CADDi crosses the billion mark: the Japanese industrial AI startup closed a $114 million Series D at a $1.2 billion valuation (Fortune).
Samsung backs a European alternative to Nvidia: the Korean giant joined a $231 million round in Dutch chipmaker Euclyd, which designs AI chips outside the GPU ecosystem (CNBC).
Mantic raises $25 million after beating humans at forecasting: the startup outperformed human participants in the summer 2026 Metaculus Cup and announced its seed on September 18 (Reuters).
Anthropic opens a life sciences program: the Life Sciences Verification Program, announced September 17, gives verified professionals access tailored to their research workflows (Anthropic).
Conclusion: the week slowing down became a legal problem
In nine days, the question of pace moved from a blog post to a courtroom. On September 12, slowing down was a proposal. On the 14th, it was a 6% valuation risk across semiconductors. On the 15th, a senator shut the door on any antitrust waiver. On the 18th, it was a charge.
The paradox holds in three numbers. The same week four labs got sued for agreeing to slow down, one is preparing an IPO on $100 billion in annualized revenue, another is discussing a $1.2 trillion round, and the third's ad business hit a $1 billion run rate in under 200 days. The models kept shipping at the usual pace.
If coordination between companies is illegal and coordination by government does not exist yet, all that remains is the unilateral commitment. The one no competitor has to match.
So what's your call, does the model release cadence actually move before year end? Ping me on Twitter/X or drop a comment.
Alex
Key takeaways
Dario Amodei published a three-step plan on September 12 to slow AI capability progress, including third-party evaluators embedded inside labs
On September 14 the PHLX Semiconductor Index fell nearly 6%, Nvidia 3.4% and SoftBank about 11% after Altman, Musk and Hassabis endorsed the proposal
OpenAI published six misalignment incident reports on September 16 and switched to continuous disclosure
On September 18, four plaintiffs sued Anthropic, OpenAI, Google and SpaceXAI over an alleged illegal agreement to slow development
Gemini 3.8 Live ships at $1.38 per hour across 97 languages, and Claude Code Projects orchestrates parallel agents that deliver their work as pull requests
