EquitiesMarket noteWednesday 30 September 2026, 00:30
OpenAI shelved its next model as it tries to catch up with Anthropic
GPT-6.1 Astra failed OpenAI’s own safety tests on the eve of DevDay. Anthropic overtook OpenAI this spring on revenue and valuation, and hired some of the best-known researchers in the field.
By The Notebook Desk
OpenAI said on 28 September that it would not release GPT-6.1 Astra, the next version of its most powerful model, after internal tests found that it fell short of the company’s standards. The model did not meet the bar for “scope and authorization, and how it communicates back to the user about the type of work it’s done”, Saachi Jain, OpenAI’s head of safety systems, told Al Jazeera.
The decision came on the eve of DevDay, the company’s developer conference in San Francisco. There, on 29 September, OpenAI launched GPT-6.1 Sol instead, a model that it says comes close to Astra at a fifth of the price per token, BGR reported. It also showed Dots, agents that run inside ChatGPT on the current GPT-6 Astra model.
For OpenAI the timing is poor. This is the year in which it lost first place to Anthropic on the measures that investors watch most.
How Anthropic moved ahead
At the end of 2025, OpenAI’s annualised revenue was about $21.4bn and Anthropic’s about $9bn, according to data compiled by Epoch AI from company statements and press reports. Anthropic passed $30bn in April. By the end of July it was at $65bn, while OpenAI was on course for more than $40bn in August.
Private investors followed the revenue. OpenAI raised $122bn at a valuation of $852bn at the end of March, Bloomberg reported. Two months later Anthropic raised $65bn at $965bn. Its listing could now value it at more than $2 trillion, which is holding up the rest of the IPO queue.
Business customers moved first. In December 2025, Menlo Ventures estimated that Anthropic had 40% of company spending on large language models through their APIs, against 27% for OpenAI, down from 50% in 2023. In coding, it put Anthropic at 54% and OpenAI at 21%.
Where the researchers went
The flow of people pointed the same way. Andrej Karpathy, a founding member of OpenAI and later director of AI at Tesla, joined Anthropic’s pre-training team in May. Nick Joseph, who leads that team, wrote that Karpathy would build “a team focused on using Claude to accelerate pretraining research itself”.
In June, Bloomberg reported that Jonas Adler and Alexander Pritzel, two senior researchers on Google DeepMind’s Gemini models, would join Anthropic. Pritzel had worked on pre-training. John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, was also reported to be moving from DeepMind to Anthropic, UC Today noted.
Pre-training is the first and most costly stage of building a model, in which it learns from vast amounts of data. It is where most of the gains in raw ability come from, and a large run takes months. So it is too soon to credit hires made in May and June for this month’s models. What the hires show is where leading researchers expect the next gains to come from. With both companies preparing to list, the value of the shares they are paid in counts too.
Anthropic has kept up its pace of releases. It launched Opus 5.5 on 22 September, which it says works at the level of its larger Fable 5.1 model and costs 40% less to run than Opus 5. Reuters described the launch as a move to counter the momentum OpenAI had built since GPT-6 Astra.
OpenAI’s reply
The case that OpenAI is closing the gap came on the day of DevDay. Its annualised revenue is approaching $70bn, after enterprise sales more than doubled since July, a source told Reuters. The run rate has risen by more than 70% since the start of the third quarter. Shares in Oracle rose 5.3% on the report. OpenAI accounts for around half of Oracle’s computing backlog, according to Gil Luria of D.A. Davidson.
If the figure holds, it is above the last number reported for Anthropic, although that one is two months old. Run rates are a loose measure: they often take one month’s revenue and multiply it by 12, as Reuters noted. Neither company has yet published audited accounts that settle the question.
The Astra decision shows the bind that OpenAI is in. Earlier this month Dario Amodei, Anthropic’s chief executive, called on AI developers to “pace the frontier”, and Sam Altman backed the call. A lab that holds back its best model loses time to its rivals. Anthropic’s prospectus lists similar risks for its own models.
What to watch
The public version of Anthropic’s prospectus will give the first audited comparison. OpenAI filed confidentially for its own listing in June and is expected to list by early 2027, Reuters reported. If OpenAI sets a new date for GPT-6.1 Astra, it will show how long a safety review can hold back a model in a race this close.
Disclosure
This is analysis and opinion, not investment advice.