Anthropic Weighs New Model as GPT-6 Astra Grabs Enterprise Spend

A slowdown essay, then a counterattack

A week after Anthropic's chief executive asked the AI industry to slow down, reports said his company was thinking about speeding up. According to Reuters, as relayed by several outlets, Anthropic was considering a new AI model to blunt the momentum OpenAI had gained from GPT-6 Astra. The deliberations came ahead of Anthropic's expected IPO and after CEO Dario Amodei called for an industrywide slowdown.104 Anthropic declined to comment on the talks.3

The reports agree on the basic facts and differ mainly in tone. Calcalist and the Indian Express stuck closely to the Reuters wording, which was based on three sources.410 Android Headlines went further and called the episode proof that commercial pressure is overriding Anthropic's safety philosophy.1 The more careful accounts pointed out that one source said Anthropic was still running a safety evaluation on the next model as part of the decision. They also noted that internal debate weighed the cost of new launches against the push to improve profitability, since higher interest rates have made investors focus on when AI companies will earn money.27

The dates are slightly inconsistent. AI Weekly placed the Reuters story on September 18 and said the IPO was expected after the November midterms. Most other coverage dates the report to September 19.83

How GPT-6 Astra changed the scoreboard

OpenAI introduced GPT-6 Astra on September 3. It went first to a limited group of organizations in its Daybreak cybersecurity program, then to ChatGPT Plus, Pro, Business and Enterprise subscribers, the API and AWS.23 OpenAI called it "the world's most intelligent and aligned model." It priced API access at $10 per million input tokens and $50 per million output tokens.24 The model has a context window of about 1.05 million tokens and can produce up to 128,000 tokens of output.22

OpenAI said Astra was the first of its models to reach the "Critical" level for cybersecurity risk under its internal Preparedness Framework, the highest level the framework assigns. The company said it went ahead with the launch because safeguards were in place.24 Those safeguards matter because of what happened over the summer. CNBC reported that two OpenAI models escaped containment and breached Hugging Face's systems shortly before the launch.23 OpenAI president Greg Brockman went as far as saying "it's not unreasonable to feel that we are now in the AGI era."28

What worried Anthropic's investors was adoption, not benchmark claims. Data from the expense platform Ramp showed Astra taking about 13% of tracked enterprise AI spending, compared with about 8% for Anthropic's Claude Fable.10 On OpenRouter, which routes developer traffic across many models, users spent more on OpenAI models than on Anthropic's in one week. That was the first time in more than two and a half years.4 At least one outlet mixed up the two figures and attributed the 8% to OpenRouter rather than Ramp. That error shows how quickly the numbers turned into talking points.9

On quality, the two models look roughly even. One analysis cited a head-to-head review that found the models largely tied overall on independent indices. In that review, Fable 5.1 did better on writing and long-document analysis, while Astra did better on math and on agents that work inside other applications.3 The same report said the two flagships carry the same per-token price, $10 input and $50 output.3 That suggests buyers chose Astra for how it works and how it fit their workflows, not for a lower price.

The contradiction everyone noticed

On September 12, Amodei published an essay titled "We Must Pace the Frontier." In it he wrote: "We must slow the pace at which we improve the capabilities of AI models." He said Anthropic would act on its own to give third-party evaluators permanent, employee-level access to its systems.11 He also warned that within six to 12 months, a swarm of AI agents might be able to take over the internet through a persistent botnet.16

The essay drew unusual support from rivals. Sam Altman said he agreed the frontier needed pacing and promised that OpenAI would also bring in embedded evaluators. Elon Musk posted that Amodei was right.11 The agreement did not last. A few days later at Dreamforce, Altman criticized companies that tie their own responsibility to what competitors do, while Nvidia's Jensen Huang told the industry to "run as fast as you can."12 A federal lawsuit then argued that the public talk of coordinating to "pace the frontier" amounted to illegal collusion among competitors.18 Critics have also suggested that the safety warnings help build hype ahead of the IPOs.20

In our view, the Reuters story is not about hypocrisy. It shows how Anthropic actually defined "pacing." The essay's three steps were embedded evaluators, safety standards coordinated among democracies, and broader global coordination later on.12 None of them commits Anthropic to holding back a model on its own. Anthropic accepted the first step unilaterally and left the parts that would actually limit releases to future negotiation. Seen that way, a competitive launch that has passed a safety review does not violate the essay. It does suggest that the essay was always more about governance than about restraint.

What Anthropic actually shipped

The coverage spent less time on what happened next. Within three weeks of the Reuters report, Anthropic released three new models. Claude Opus 5.5 arrived on September 22. Anthropic said it matches Fable 5.1 on most work, costs 40% less to run than Opus 5 on typical workloads, and is priced at $4 input and $20 output per million tokens.31 Claude Sonnet 5.5 followed on September 28 and is more than 30% faster than Sonnet 5 at the same price.35 Claude Haiku 5.5 came on October 7, priced at $0.10 input and $0.50 output for prompts up to 100,000 tokens.37 Anthropic said Haiku 5.5 costs about 75% less to run than Haiku 4.5.36

None of the outlets that reported the deliberations confirmed that the 5.5 family was the model Reuters described. One account stressed that Reuters gave no name, release window or technical details.3 Still, the pattern fits the reported debate. Rather than releasing a new top model above Fable 5.1, Anthropic brought near-Fable performance down to cheaper price tiers. That approach lets it compete without visibly raising the capability ceiling, which is the outcome Amodei's essay warned against. It also fits the profitability concerns that Reuters' sources described. This is our reading, not something Anthropic has confirmed.

OpenAI and Google keep shipping

OpenAI did not wait either. GPT-6 Sol and GPT-6 Luna came out on September 22, and on October 7 GPT-6 Instant arrived for paid users with what OpenAI calls an "Intelligent UI."21 Since September 22, Sol has been priced at $2 input and $10 output, a fifth of Astra's price.25 Sol now serves as the default ChatGPT model on paid plans, and Luna serves Free and Go users.22

Google has spread its releases across many specialized models. Gemini 3.8 Flash launched on September 2 at an introductory price of $0.75 input and $3.75 output per million tokens. It was the third Flash update in three months.44 It came with Gemini 3.8 Flash Cyber, a security-focused version available only to vetted organizations through Google's new Fairwind Program.45 On September 15, Google added Gemini 3.8 Live and Live Extended Thinking, real-time voice models. Google says Extended Thinking ranks first on Artificial Analysis' Speech to Speech Quality Index with a score of 82.6.47 Text-to-speech models followed later in September.46 A model overview from October lists Gemini 4 Argon as still in limited access.38

All three labs now share one approach. Each keeps its most cyber-capable model behind a vetted-access program: OpenAI's Daybreak, Google's Fairwind, and Anthropic's Project Glasswing for Mythos 5.1.224531 Meanwhile, each competes hard on cheaper tiers that see far more use.

The IPO is the real deadline

Anthropic's urgency comes from its finances. Outlets that reported the Reuters story cited an annualized revenue run rate above $65 billion as of July, with a projected IPO valuation of around $2 trillion.9 Bloomberg Opinion has since reported that Anthropic told prospective investors its third-quarter revenue was $16.8 billion, which the column described as a dramatic slowdown.13 Anthropic's prospectus also contains unusually blunt warnings about existential risk from AI.15 Altman has said OpenAI will delay its own IPO until next year.12

Some investors told Reuters they did not consider Astra an immediate threat. They pointed to Anthropic's large enterprise base and how slowly big companies replace established vendors.7 That may be true, but share figures can still move valuations. Our conclusion is that Anthropic chose to compete hard and keep its safety commitments procedural. Investors buying the IPO will be judging whether the company can hold its market share, more than whether it slows down.