Anthropic is weighing whether to rush out a new AI model after OpenAI’s GPT-6 Astra took a bite out of its enterprise business in the space of a single week, a shift first detailed by Analytics India Magazine. The move comes as OpenAI’s newest model, released September 3, 2026, climbed to roughly 13% of enterprise AI spending tracked by corporate expense platform Ramp, compared with about 8% for Anthropic’s Claude Fable, according to reports from Techstrong.ai. Developer-traffic router OpenRouter recorded a similar shift, with users spending more on OpenAI models than Anthropic models for the first time in more than two years.
The timing could hardly be worse for Anthropic. The company is heading toward a Nasdaq listing that bankers have floated at a valuation near $2 trillion, and its pitch to public investors rests heavily on revenue momentum, not on ceding ground to a rival mid-quarter. Reuters reported that Anthropic is now evaluating the safety profile of its next model as part of internal talks over whether, and when, to ship it. The standoff is the latest flashpoint in our ongoing AI and machine learning coverage, and it sits awkwardly next to CEO Dario Amodei’s own public calls for the industry to slow its pace of capability gains, a stance covered in depth after his “pace the frontier” remarks earlier this year.
GPT-6 Astra’s Fast Start Puts Pressure on Anthropic
OpenAI began rolling out GPT-6 Astra to approved users on September 3, 2026, with wider availability the next day, according to the company’s own system and safety documentation. OpenAI calls it “the most capable model we have ever broadly deployed” in its safety materials, and its launch communications go further, describing Astra as “the world’s most intelligent and aligned model.” Those aren’t small claims for a company that shipped GPT-5.6 Sol only months earlier, and the market reaction suggests enterprise buyers took the pitch seriously.
CNBC reported that the Astra rollout was tied to a broader safety and cybersecurity framing from the start, a departure from OpenAI’s usual product-launch playbook. That framing matters because it’s the same lens Anthropic now has to apply to its own roadmap. If OpenAI can put a Critical-risk model into general availability and still gain enterprise share, the calculus for what counts as an acceptable release just shifted for everyone chasing the same customers. For background on how GPT-6 Astra’s pricing and access tiers work, Astra launched at $10 and $50 pricing tiers with staged access before reaching general availability.
Inside OpenAI’s Critical Cybersecurity Classification
OpenAI’s own Preparedness Framework classifies GPT-6 Astra at the Critical level of cybersecurity capability, the highest tier the company tracks. Per OpenAI’s system card documentation, that means the model, given appropriate tools and access, can find previously unknown security flaws and develop new ways to exploit them across well-protected systems without a person guiding each step. That’s a substantial jump from prior generations, and it’s the reason Astra shipped with extra scrutiny attached rather than a standard product announcement.
OpenAI also states that Astra is more robust to jailbreaks than GPT-5.6 Sol and better aligned overall, calling it a significant step forward in model alignment. The company frames the Critical label as a safety disclosure rather than a warning against use, but the classification has already shaped how the rest of the industry talks about its own releases. Readers who want the full breakdown of what the Critical tier actually restricts can find it in our earlier coverage of the two-week pause OpenAI applied before general release.
What GPT-6 Astra Can Actually Do
Beyond the safety framing, OpenAI positions GPT-6 Astra as state-of-the-art on computer use, browsing, professional work, software engineering, cybersecurity and science. Benchmark summaries circulating since launch show improved scores over both GPT-5.6 Sol and Anthropic’s Claude Fable 5.1 on common evaluation suites, though neither company has published a full head-to-head comparison. TechCrunch reported that Astra rolled out first through OpenAI’s Daybreak cybersecurity program before extending to Pro, Plus, Enterprise, Business and API customers, a staged approach that let OpenAI gather safety data before the widest release.
That staged rollout also gave OpenAI a marketing advantage: by the time Astra hit general availability, early enterprise pilots were already generating usage data that fed directly into the Ramp figures showing the 13% spending share. Anthropic doesn’t have a comparable pipeline running for its next release, which is part of why the decision to ship quickly carries real trade-offs.
GPT-6 Astra vs GPT-5.6 Sol vs Claude Fable 5.1
The table below lays out what’s confirmed about each model’s positioning as of September 21, 2026. Figures marked “not disclosed” reflect gaps in the public record rather than omissions on our part.
| Model | Developer | Release Date | Safety Classification | Enterprise Spend Share (Ramp) |
|---|---|---|---|---|
| GPT-6 Astra | OpenAI | September 3-4, 2026 | Critical (cybersecurity, Preparedness Framework) | ~13% |
| GPT-5.6 Sol | OpenAI | Prior generation (2026) | Below Critical per OpenAI comparisons | Not broken out separately |
| Claude Fable 5.1 | Anthropic | Earlier 2026 release | Not disclosed at Astra’s tier | ~8% |
Claude Fable 5.1 itself was a notable release in its own right, cutting cache costs by 75% when it launched. Details on that update are in our report on the Fable 5.1 and Mythos rollout. The gap in the table isn’t a knock on Fable’s engineering, it reflects that Anthropic hasn’t published a Preparedness-style Critical classification of its own, making a direct safety-tier comparison impossible with public data.
Ramp Data Ends a 2.5-Year Anthropic Streak
The detail that’s rattled Anthropic the most may be the OpenRouter figure. According to reporting picked up by Benzinga and Techstrong.ai, users routed more spend to OpenAI models than to Anthropic models on OpenRouter last week, the first time that’s happened in over two years. For a company that built its enterprise reputation on Claude being the default choice for coding and agentic workflows, a one-week reversal isn’t fatal on its own. But paired with the Ramp numbers, it’s a signal that switching costs for enterprise AI buyers are lower than either lab would like to admit.
Anthropic has previously touted its own internal usage figures to argue Claude is winning where it matters most. The company said Claude accounts for 26% of its own AI research and development work, a stat covered in our earlier piece on Anthropic’s self-reported R&D reliance on Claude. That kind of internal dogfooding argument doesn’t move Ramp’s dashboard, though, and it’s Ramp’s dashboard that enterprise finance teams are actually watching this month.
The Money Behind the Rivalry: Revenue and the $2 Trillion IPO
Anthropic’s underlying financial position is stronger than the spending-share numbers might suggest. CNBC reported that Anthropic told investors its annualized revenue run rate climbed to $65 billion in July 2026, up from roughly $47 billion in May and about $9 billion at the end of 2025, a sevenfold jump in seven months. By comparison, OpenAI’s own disclosed run rate stood at $40 billion at last report, meaning Anthropic still leads on this particular metric even after losing ground in weekly enterprise spending share.
That revenue trajectory underpins Anthropic’s IPO math. Bloomberg and Axios both reported the company’s last private valuation hit $965 billion in a May 2026 Series H round, and bankers are now discussing a public listing valuation of roughly $2 trillion, with marketing for the offering expected as early as mid-October and a Nasdaq debut targeted for November. The pitch to public investors leans on projected 2028 revenue of $190 billion to $200 billion, a figure that assumes Anthropic keeps growing enterprise share rather than watching it erode to OpenAI quarter over quarter.
Enterprise AI Spending and Revenue at a Glance
| Metric | Anthropic | OpenAI |
|---|---|---|
| Ramp enterprise spend share (Sept 2026) | ~8% | ~13% |
| Annualized revenue run rate (July 2026) | $65 billion | $40 billion |
| Revenue run rate, end of 2025 | ~$9 billion | Not disclosed at same interval |
| Reported/targeted valuation | ~$2 trillion (targeted IPO, Nov. 2026) | Not applicable (no near-term IPO confirmed) |
| OpenRouter weekly spend lead | Lost lead after 2.5 years | Took lead week of Sept. 2026 Astra launch |
Read together, the two tables tell a split story. Anthropic still generates more revenue overall and is scaling faster on a percentage basis, but OpenAI just proved it can pull enterprise workloads away from Claude with a single well-timed release. Sam Altman confirmed on September 12 that OpenAI itself would not pursue a public listing in 2026, citing AI safety concerns, which removes one point of direct competitive symmetry between the two companies’ near-term corporate plans.
Why Amodei’s Slowdown Call Complicates a Response
Dario Amodei has spent much of 2026 publicly arguing that AI labs should slow the rate at which they push capability forward, a position laid out in a three-step plan he shared earlier this year. That stance was easy to hold when Anthropic was comfortably ahead on enterprise mindshare. It gets harder to defend internally once a rival model starts pulling paying customers away in real time.
Reuters’ sourcing indicates Anthropic’s internal debate isn’t just about whether to ship a new model, but about how to evaluate its safety without abandoning the caution Amodei has championed publicly. Anthropic has leaned on outside partners for exactly this kind of check before, including bringing in independent evaluators to assess model behavior ahead of releases. Whatever model Anthropic ships next will be read as a referendum on whether Amodei’s slowdown rhetoric survives contact with a competitive threat that’s already showing up in the company’s own revenue dashboards.
Historical Context: Five Years of Leapfrogging
This isn’t the first time the two labs have traded the lead. OpenAI held the popular-attention advantage through most of ChatGPT’s early growth, Anthropic pulled ahead on coding and agentic workflows through 2025 as Claude became the default pick for many developer tools, and now OpenAI has clawed back enterprise spending share with Astra. Each swing has come faster than the last. Where the GPT-4-to-Claude 3 shift played out over roughly a year, the Astra-driven shift in Ramp’s tracked spending happened within a single week of general availability.
The compressed timeline says less about either model’s underlying quality and more about how quickly enterprise procurement teams now re-evaluate vendor contracts when a credible new option appears. Multi-model routing tools like OpenRouter have made switching a matter of updating an API endpoint rather than renegotiating a platform contract, which is exactly why the OpenRouter spend-share flip landed as such a visible signal this month.
How Other Labs Are Positioning Against Astra
Anthropic and OpenAI aren’t the only labs adjusting to Astra’s Critical classification. Google has continued pushing its Gemini line on multimodal and live-audio performance rather than competing directly on the cybersecurity-capability axis, an approach visible in Gemini 3.8 Live’s recent audio-benchmark push. Chinese lab Zhipu has taken a different tack, with GLM-5.2 posting benchmark wins over GPT-5.5 while still trailing Claude on the specific tests it was measured against, before Astra’s release reshuffled the leaderboard again.
As Al Jazeera noted in its coverage of the launch, Astra arrived amid rising scrutiny over exactly this kind of disclosure, and no other lab has yet published a Preparedness-style Critical cybersecurity classification of its own. That leaves OpenAI as the only company that has formally told regulators and customers it shipped a model capable of independent exploit discovery at scale, a distinction Anthropic will have to address directly if its own next release approaches similar capability without matching the disclosure.
What Enterprise Buyers Should Watch
For IT and security teams currently running Claude in production, the immediate question isn’t whether to switch models overnight. It’s whether procurement contracts include the flexibility to route workloads to whichever model performs best on a given task, the same flexibility that let OpenRouter customers shift spend so quickly. Teams locked into single-vendor agreements have less room to react to a launch like Astra’s, for better or worse.
Security teams specifically should pay attention to how OpenAI’s Critical cybersecurity classification affects their own risk models, since a tool capable of independent exploit discovery changes the calculus for both defenders and attackers using the same technology. Anthropic’s eventual response model will likely face the same disclosure pressure once it reaches comparable capability, regardless of what Anthropic decides about release timing this quarter.
Finance teams evaluating multi-year AI contracts should also factor in the volatility this month exposed. A vendor that held a stable enterprise lead for 2.5 years lost visible ground within days of a single competing launch. That’s a different risk profile than the one most procurement policies were written for a year ago, when switching between frontier models required more engineering work than updating a routing configuration.
What an Anthropic Answer Model Could Look Like
Anthropic hasn’t confirmed specifications, pricing or a release window for whatever comes after Claude Fable 5.1, and any details beyond what’s been reported should be treated as speculation rather than fact. What is confirmed is that the decision is being driven by competitive pressure from Astra’s enterprise gains, filtered through Anthropic’s own safety evaluation process, and shadowed by the IPO timeline that gives the company strong incentive not to look like it’s losing the race in the months before it goes public.
Given Anthropic’s public commitments to safety review, any new model is more likely to ship with its own capability disclosure than to skip that step entirely, even under competitive pressure. That would put Anthropic in the position of either matching OpenAI’s Critical classification openly, which invites the same scrutiny OpenAI is currently facing, or shipping a model that stays below that threshold and risks looking less capable by comparison, at least on paper.
Five Predictions for the Next Two Quarters
- Anthropic will announce a new model or a significant Claude update before its Nasdaq listing, most likely timed to land ahead of the marketing period for the IPO rather than after it.
- Ramp’s enterprise spending tracker will show at least one more meaningful swing between OpenAI and Anthropic before year-end, given how quickly the September shift happened.
- Anthropic will address its own model’s cybersecurity capability level publicly if the new release approaches Astra’s classification, rather than staying silent on the comparison.
- OpenRouter’s weekly spend data will keep functioning as an early indicator of enterprise sentiment shifts faster than quarterly earnings reports from either company.
- At least one additional major lab, most plausibly Google, will publish its own capability-tier disclosure in response to the precedent OpenAI set with Astra’s Critical classification.
FAQ
Is Anthropic definitely releasing a new AI model?
No. Reuters reported that Anthropic is weighing the decision and evaluating safety questions, but the company has not confirmed a release date or specifications as of September 21, 2026.
What is GPT-6 Astra’s Critical classification?
It’s the highest tier in OpenAI’s Preparedness Framework for cybersecurity capability, indicating the model can independently find and exploit unknown security flaws in well-protected systems given the right tools and access.
How much enterprise market share has GPT-6 Astra taken from Anthropic?
Ramp data cited by Benzinga and Techstrong.ai put Astra at roughly 13% of tracked enterprise AI spending against about 8% for Claude Fable, as of mid-September 2026.
Does Anthropic still make more revenue than OpenAI?
Yes, based on the most recently disclosed figures. CNBC reported Anthropic’s annualized revenue run rate at $65 billion in July 2026 versus OpenAI’s disclosed $40 billion, though the two companies may calculate the metric differently.
When is Anthropic’s IPO expected?
Reports point to a Nasdaq listing targeted for November 2026, with a potential valuation near $2 trillion, following a $965 billion private valuation in the company’s May 2026 Series H round.
Is GPT-6 Astra safe to use in production?
OpenAI has released it broadly and describes it as better aligned and more jailbreak-resistant than GPT-5.6 Sol, but the Critical cybersecurity classification means organizations should apply the same access controls they’d use for any tool capable of independent exploit discovery.
Why does Dario Amodei’s slowdown stance matter here?
Amodei has publicly called for the industry to slow capability gains. A rushed competitive response to Astra would sit awkwardly next to that public position, which is part of why Anthropic’s internal deliberation is reportedly taking safety evaluation seriously rather than shipping immediately.
Will other AI labs follow OpenAI’s Critical disclosure model?
No lab has published a matching Critical-tier cybersecurity disclosure yet, but the competitive and regulatory pressure created by Astra’s classification makes it more likely that rivals, including Anthropic, will adopt similar public disclosures for future high-capability releases.
