Anthropic is weighing whether to rush out a new flagship AI model, according to a Reuters report published September 19, 2026 that cited three unnamed sources familiar with the company’s internal discussions. The timing is not a coincidence. Since OpenAI’s GPT-6 Astra launched on September 3, it has started pulling enterprise dollars and developer traffic away from Anthropic’s Claude line at a pace the company has not faced in more than two and a half years, according to data cited in Reuters’ reporting and repeated by Mint’s write-up of the story. Anthropic declined to comment when Reuters asked about the deliberations.
The stakes go beyond bragging rights on a leaderboard. Anthropic is preparing for an initial public offering, and any sign that OpenAI is closing the enterprise AI gap complicates how bankers and investors price that IPO. At the same time, Anthropic CEO Dario Amodei spent a 3,800-word essay on September 12 arguing that the entire industry, including his own company, needs to slow down. Now Anthropic may be about to do the opposite. Here is what is actually known, what the numbers say, and why an anthropic new AI model decision matters well beyond the two companies involved.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Reuters Actually Reported
Reuters’ September 19 report states plainly that Anthropic is considering rolling out a new AI model to counter OpenAI’s momentum since its launch of GPT-6 Astra, according to three sources, ahead of an expected IPO and after its CEO called for an industrywide slowdown. That single sentence carries three separate storylines: a competitive response to Astra, a pre-IPO calculation, and a direct tension with Amodei’s own public call for restraint.
Mint’s follow-up coverage adds detail that did not make it into every outlet’s summary. Citing the same Reuters sourcing, Mint reports that people familiar with Anthropic’s thinking say the company is weighing how much to spend on bringing a new model to market while also improving profitability, a debate sharpened by higher interest rates that have made investors more focused on when AI labs will actually turn a profit. As part of the discussion, Anthropic is also assessing the safety implications of a faster release, per the same report. None of this is confirmed as a done deal. It is, by every outlet’s own framing, still a live internal debate rather than an announced product.
Why Astra’s Three Weeks Changed the Conversation
GPT-6 Astra shipped to approved users on September 3 and rolled out to all paid ChatGPT plans the following day, according to a comparison published by Improvado. OpenAI’s own launch messaging described the release directly: “We’re introducing GPT‑6 Astra, the world’s most intelligent and aligned model,” the company said on its official launch page. That came just two days after Anthropic had shipped its own answer, Claude Fable 5.1 and its gated sibling Claude Mythos 5.1, on September 1, according to eWeek’s side-by-side review.
What changed is not that Astra is dramatically better across the board. Improvado’s comparison, published September 9, concludes that independent indices largely tie the two models overall. What changed is adoption. Astra improved on computer operation, software development, cybersecurity tasks and other professional workflows, and businesses noticed fast, per Mint’s reporting of the Reuters findings. Within roughly two weeks of launch, spend data started shifting in OpenAI’s favor for the first time in years, not months.
The Ramp and OpenRouter Numbers Behind the Alarm
Two independent data sources are doing most of the work in this story, and both are cited in Reuters’ reporting as relayed by Mint. Corporate expense management platform Ramp found that GPT-6 Astra represented roughly 13% of enterprise AI spending, compared with around 8% for Anthropic’s Claude Fable line. Separately, developer routing platform OpenRouter reported that its users spent more on OpenAI models than on Anthropic’s models last week, the first time that has happened in more than two and a half years of OpenRouter tracking that metric.
Neither data point proves Anthropic has lost its enterprise AI lead outright. Ramp’s 13% versus 8% split still leaves plenty of enterprise spend going to other providers entirely, and OpenRouter’s developer traffic skews toward a different customer base than Anthropic’s large enterprise contracts. But directionally, both metrics point the same way, and that is unusual. It is rare for two separate measurement systems, one tracking corporate card spend and one tracking API routing traffic, to move in the same direction within the same three-week window.
| Metric | Anthropic | OpenAI |
|---|---|---|
| Flagship model & release date | Claude Fable 5.1 — Sept 1, 2026 | GPT-6 Astra — Sept 3, 2026 |
| Enterprise AI spending share (Ramp) | ~8% | ~13% |
| OpenRouter weekly spend lead | Lost lead for first time in 2.5+ years | Surpassed Anthropic |
| Annualized revenue run rate (July 2026) | $65B+ (up from ~$9B end of 2025) | $40B+ |
| 2028 revenue forecast | $190B–$200B (per Reuters) | Not disclosed in this report |
| Pricing per million tokens (input/output) | $10 / $50 | $10 / $50 |
Revenue Race: A $65 Billion Run Rate Against a $40 Billion Rival
Put the market-share data next to the revenue numbers and Anthropic’s actual position looks stronger than a single week of OpenRouter traffic suggests. Anthropic’s annualized revenue run rate exceeded $65 billion by the end of July 2026, up sharply from roughly $9 billion at the end of 2025, according to Reuters reporting cited by Mint. That growth was underpinned by deepening enterprise commitments: more than 1,000 customers were each spending over $1 million annually on Anthropic by April 2026, and Claude Code alone reached a $2.5 billion annualized revenue run rate by February 2026, according to data from Sacra. OpenAI’s own annualized revenue run rate surpassed $40 billion in July. Reuters has also previously reported that Anthropic is forecasting revenue of approximately $190 billion to $200 billion by 2028.
That gap matters for how the Astra story should be read. Anthropic is not being out-earned by OpenAI today, it is being challenged on the margin, in the week-to-week metrics that investors watch closely as leading indicators ahead of an IPO. A three-week wobble in enterprise spend share is not the same as a structural reversal, but pre-IPO investors tend to react to leading indicators before they show up in trailing revenue, which is exactly why this story broke now rather than after a full quarter of data.
Amodei’s Slowdown Essay Meets a Speed-Up Decision
The awkward part of this story is timing. On Saturday, September 12, Dario Amodei published a lengthy essay on his personal website titled “We Must Pace the Frontier,” arguing that the AI industry needs to deliberately throttle how fast it improves model capabilities so safety research can catch up. “We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain,” Amodei wrote, according to reporting from The Guardian and BBC News. The BBC also quoted him explaining the reasoning: he believes that if slowing down bought the industry even an extra year or two before models reach critical levels of capability, that time could be used to advance alignment research and reduce the risk of something going seriously wrong.
Seven days later, Reuters is reporting that his own company is discussing whether to speed up its release calendar. That is not necessarily hypocrisy. Amodei’s essay called for the industry to coordinate on pacing, and Anthropic’s internal deliberations, as described in Reuters’ reporting, explicitly include a safety assessment of any new model before it ships. But the optics are difficult to avoid: a CEO asking the world to slow down while his own company debates whether it can afford to slow down given what a rival just did to its market position. Anthropic’s earlier public commitment alongside OpenAI to a development brake is now being tested by exactly the kind of competitive pressure that framework was meant to withstand.
Claude Fable 5.1 vs GPT-6 Astra: Where Each One Actually Wins
Strip away the market-share drama and the two current flagship models are closer than the enterprise spending data implies. Improvado’s head-to-head, published September 9, breaks the comparison down by task rather than by an aggregate score, and its conclusion is genuinely split.
Where Claude Fable 5.1 wins
Improvado’s analysis finds Fable 5.1 is the better default for writing, for reading long documents, and for analysis a user has to defend later, meaning outputs where reasoning transparency matters more than raw speed. Anthropic also cut cached-token read pricing by 75%, to $0.25 per million tokens, which makes Fable 5.1 cheaper for workloads that repeatedly reference the same long context, such as document review or codebase analysis, a detail confirmed across tech-insider.org’s coverage of the Fable 5.1 launch. That cost advantage deepened further once Anthropic rolled out Claude Sonnet 5.5 with a 1-million-token context window in October 2026, and priced the Opus 5.5 update roughly 40% below Opus 5, according to Anthropic’s own figures, extending the long-context cost edge Improvado had already flagged.
Where GPT-6 Astra wins
The same Improvado comparison finds Astra is the stronger choice for math, for agents that fill out forms and operate inside other applications, and for short tasks run at high volume, since it tends to write less output per completed task and therefore costs less per unit of finished work even at matching per-token pricing. Anthropic has since narrowed that output-efficiency gap by giving Opus 5.5 support for up to 128,000 output tokens as of October 2026, handing its models more room on exactly the kind of long, multi-step completions where Astra’s shorter outputs had been winning on cost. eWeek’s review adds a wrinkle worth noting: independent evaluator Artificial Analysis, which eWeek describes as the only neutral referee in the comparison, ran both models through identical tests and crowned Fable 5.1 the winner on its top two indices, even though OpenAI’s own published charts show Astra ahead on most other measures. In other words, whoever is scoring the fight has a real effect on who wins it.
How Fast Models Are Actually Improving
Amodei’s slowdown argument is easier to evaluate against a concrete benchmark. Terminal-Bench 4.0, a coding-agent evaluation tracked by industry newsletter ThursdAI, shows just how quickly Anthropic’s own models have moved this year. Claude Fable 5 scored 42.0 on the benchmark; Fable 5.1, released roughly two months later, jumped to 55.8, a gain ThursdAI’s September roundup called a significant leap for a point release rather than a full generational jump. For comparison, OpenAI’s prior-generation GPT-5.6 Sol scored around 37 on the same benchmark, according to the same tracking.
| Model | Lab | Terminal-Bench 4.0 score | Release timing |
|---|---|---|---|
| GPT-5.6 Sol | OpenAI | ~37 | Prior generation, before Astra |
| Claude Fable 5 | Anthropic | 42.0 | Earlier 2026 release |
| Claude Fable 5.1 | Anthropic | 55.8 | Sept 1, 2026 |
| Fable 5.1 on Terminal-Bench Science | Anthropic | 52% (more than double Fable 5) | Sept 1, 2026 |
A separate report from CryptoBriefing found that the combined effect of Fable 5.1 and Astra shipping within days of each other stretched the proprietary models’ advantage in web development benchmarks to roughly 1,800 points over open-weight competitors, a gap that has widened even as the two proprietary leaders trade blows with each other. That is the pace Amodei wants throttled, and it is also the pace Anthropic may now feel it cannot afford to slow on its own while a rival is closing the enterprise gap.
Anthropic’s IPO Calculus and the Cost of Waiting
Anthropic’s valuation trajectory over 2026 has been steep. tech-insider.org’s earlier coverage tracked the company through a $65 billion Series H round that reportedly valued the company near $965 billion, and through reporting around a planned IPO tied to the Fable 5.1 and Mythos 5.1 launch. Mint’s write-up of the Reuters story references separate commentary asking whether Anthropic should press pause on a potential IPO valued as high as $2 trillion, underscoring how much is riding on the company’s ability to show it can hold its enterprise position against Astra heading into any listing.
That is the real tension behind the anthropic new AI model deliberations. Ship too fast and the company undercuts Amodei’s own public safety framework days after he published it. Wait too long and prospective IPO investors watch OpenRouter and Ramp data keep drifting toward OpenAI for another quarter. Reuters’ sourcing describes Anthropic as explicitly weighing how much to spend on a new release while also trying to improve profitability, a balancing act made harder by higher interest rates pushing investors to demand a visible path to profit rather than pure growth.
What a New Anthropic Model Could Look Like
Reuters’ report does not specify a model name, a release window, or technical details, and no other outlet reviewed for this story has independently confirmed specifics either. What is known is narrower: Anthropic is assessing the safety profile of a potential release as part of the same internal discussion about whether to move forward, according to the sourcing behind Reuters’ report. Given that Claude Fable 5.1 and its restricted sibling Mythos 5.1 only shipped on September 1, any follow-up model would represent an unusually fast turnaround by Anthropic’s own recent release cadence, which has typically spaced flagship updates by months rather than weeks.
It is worth being precise about what is not confirmed: there is no confirmed name for a next model, no confirmed release date, and no confirmed pricing. Any outlet or commentator presenting those details as settled fact right now is getting ahead of the reporting.
Why Some Investors Say Astra Isn’t a Real Threat
Not everyone close to the deal flow is alarmed. Mint’s report notes that both existing investors and prospective investors planning to participate in Anthropic’s and OpenAI’s IPOs told Reuters they do not view Astra as an immediate major threat to Anthropic’s position. Their reasoning centers on Anthropic’s sizeable existing lead in enterprise AI contracts and the practical difficulty of persuading large businesses to switch away from an AI vendor already wired into their internal tooling, security review processes, and procurement agreements. Those investors also expect the competitive order among Anthropic, OpenAI, Alphabet’s Google, and other frontier labs to keep shifting as new model generations ship, meaning any single company’s lead, including OpenAI’s current one, could prove temporary.
That view lines up with how quickly the numbers have moved before. Anthropic went from roughly $9 billion in annualized revenue at the end of 2025 to more than $65 billion by the end of July 2026, a run rate expansion that happened while OpenAI was also growing and while Google’s Gemini line kept shipping. Enterprise AI market share, in other words, has not been a stable, slow-moving figure this year for anyone.
Historical Context: How Anthropic Built Its Enterprise Lead
Anthropic’s position as the enterprise favorite did not happen by accident. The company built its reputation on being the safety-first alternative to OpenAI, a positioning that became commercially valuable as large companies with compliance and security requirements gravitated toward Claude models for coding and document-heavy workflows. That reputation is also the backdrop for the security incidents Anthropic has had to disclose and rescan its logs for this year, each of which tested whether enterprise customers would stay loyal through public stumbles. That loyalty shows up in the adoption figures too: more than 100,000 businesses were running Claude through Amazon Bedrock by April 2026, according to Sacra. So far, according to the revenue trajectory reported by Reuters, they largely have.
The Astra episode is the first time in this cycle that a competitor has challenged that position on capability and adoption simultaneously, rather than on price alone or on a narrow benchmark win. That combination, a real product improvement plus a fast enterprise adoption curve, is what turned a single Ramp data point into a Reuters story about Anthropic’s product roadmap.
Competitive Landscape: Google, xAI, and Open-Source Pressure
The Anthropic-OpenAI rivalry is not happening in a vacuum. tech-insider.org’s earlier comparison of GPT-6 Astra against Google’s Gemini 3.8 and DeepSeek’s V4.1 Flash found sharp pricing gaps between the proprietary frontier leaders and cheaper alternatives, and CryptoBriefing’s benchmark tracking shows the proprietary-versus-open-weight gap widening even as Anthropic and OpenAI compete with each other. That matters for how the anthropic new AI model debate should be read: Anthropic is not just racing OpenAI, it is trying to hold a premium enterprise position against open-weight models from Chinese labs that are improving quickly on cost, a pressure point Reuters’ sourcing explicitly cited as part of Anthropic’s calculus on spending and profitability.
Separately, tech-insider.org previously reported that Claude Fable 5.1 held a lead over GPT-5.6 Sol and Zhipu’s GLM-5.3 on SWE-bench-style coding evaluations, a reminder that the current Astra-versus-Fable story is really the latest lap in a race that has already run through several prior model generations this year alone.
Market Impact for Enterprise AI Buyers
For companies actually buying AI tools, the practical takeaway is less dramatic than the headlines. The Ramp and OpenRouter data show a shift in the direction of spend, not a collapse of Anthropic’s business. Enterprise procurement teams evaluating Claude Fable 5.1 against GPT-6 Astra should weigh the Improvado findings directly: Fable 5.1 for writing-heavy and document-review workloads where defensible reasoning matters, Astra for math-heavy and agentic app-automation workloads where task volume and cost-per-completed-task matter more than depth of explanation. Pricing parity at $10 input and $50 output per million tokens for the flagship models means the decision increasingly comes down to task fit rather than sticker price, though buyers weighing a full lineup rather than just the top-end model should note that Anthropic’s September 2026 Claude Sonnet 5.5 release undercuts that parity at the mid-tier, priced at $2 per million input tokens and $10 per million output tokens while running roughly 30% faster than Sonnet 5, with Claude Opus 5.5 following at $4/$20 per million tokens for teams that need more headroom than the mid-tier model offers.
The bigger signal for buyers is that neither lab has a durable, multi-month lead right now. Model generations are shipping roughly every one to two months, and evaluators disagree on who is ahead depending on which index they trust, as eWeek’s Artificial Analysis comparison shows. Locking a vendor contract to a single model’s current benchmark position looks increasingly risky given how fast that position is turning over.
Predictions: What Happens Next
Based on the reporting available, here is how this is likely to play out over the coming weeks and months:
- Anthropic will likely make a public decision on a new model within weeks rather than months, given how quickly the Ramp and OpenRouter data are moving and how sensitive that data is to pre-IPO investor sentiment.
- Whatever Anthropic ships next will almost certainly be framed publicly as consistent with Amodei’s “pace the frontier” safety framework, even if the internal timeline is compressed, since the company cannot afford to look like it abandoned its own stated principles within a month of publishing them.
- OpenAI will likely keep pushing Astra’s enterprise adoption numbers publicly to sustain the narrative that it has closed the gap, using metrics similar to the Ramp spending-share figures already in circulation.
- Expect Google’s Gemini line and open-weight Chinese models to keep pressuring both companies on price, which will shape how aggressively Anthropic can spend on a rushed release without damaging the profitability story it needs for its IPO.
- The Terminal-Bench and Artificial Analysis benchmark gap between Anthropic and OpenAI’s flagships will likely narrow further before either company opens a clear, durable lead, continuing the pattern of the last several model generations.
Frequently Asked Questions
Is Anthropic definitely releasing a new AI model?
No. Reuters reported on September 19, 2026 that Anthropic is considering the move, citing three sources, but Anthropic declined to comment and no release has been announced.
Why is GPT-6 Astra gaining ground on Anthropic now?
Ramp data cited in Reuters’ reporting shows GPT-6 Astra reached roughly 13% of enterprise AI spending versus around 8% for Anthropic’s Claude Fable, and OpenRouter reported OpenAI models outspending Anthropic’s for the first time in more than two and a half years, following Astra’s September 3 launch.
Does this contradict Dario Amodei’s call for an AI slowdown?
It creates tension with it. Amodei published an essay on September 12, 2026 calling for the industry to slow the pace of capability gains, and seven days later Reuters reported his own company is weighing a faster release to counter a rival. Anthropic’s internal discussion reportedly includes a safety assessment of any new model.
How does Claude Fable 5.1 compare to GPT-6 Astra today?
According to Improvado’s comparison, Fable 5.1 is stronger for writing, long-document reading, and defensible analysis, while Astra is stronger for math and agentic tasks that operate inside other apps. Independent indices largely tie the two overall, though eWeek reports evaluator Artificial Analysis rated Fable 5.1 ahead on its top two indices.
How big is Anthropic’s revenue compared to OpenAI’s?
Anthropic’s annualized revenue run rate exceeded $65 billion by the end of July 2026, up from about $9 billion at the end of 2025, according to Reuters reporting cited by Mint. OpenAI’s annualized run rate surpassed $40 billion over the same period.
Is Anthropic’s IPO at risk because of GPT-6 Astra?
No confirmed change to IPO plans has been reported. Investors cited in Reuters’ reporting told the outlet they do not consider Astra an immediate major threat to Anthropic’s enterprise position, citing switching costs and Anthropic’s existing lead.
When will Anthropic decide on a new model?
No timeline has been confirmed publicly. Reuters describes the decision as an active internal discussion as of mid-to-late September 2026, with no confirmed release date.
Related Coverage
- GPT-6 Astra vs Gemini 3.8 vs DeepSeek V4.1: 83x Gap [2026]
- Muse Spark Tops GPT-6 Astra, Claude Fable: 75% SWE [2026]
- GPT-6 Astra Jailbroke Itself: OpenAI Blocks 91.5% [2026]
- OpenAI, Anthropic, Google in AI Safety Talks: 5 Labs [2026]
- ChatGPT 7 Rumors Debunked: Real Story Is GPT-6 Astra [2026]
