Who is winning the AI race?
OpenAI released GPT-6 Astra on Thursday, two days after Anthropic made Claude Fable 5.1 generally available.
The benchmark table OpenAI published alongside the launch puts Astra ahead of Fable 5.1 on most rows: 97.6 percent against 87.8 on the hardest maths test, 96 against 93.7 on graduate-level science.
Greg Brockman, president of OpenAI, called it a generational leap and told reporters he believes the company may have reached AGI.
That is a big claim.
Some folks have sniffed at this as hyperbole and self-backpatting as the independent picture looks different.
Artificial Analysis, which runs the same nine evaluations across every model it can get its hands on, scores Fable 5.1 at 66 on its Intelligence Index and Astra at 61, which is the same score as OpenAI's previous model GPT-5.6 Sol and also behind Claude Opus 5 and Meta's new Muse Spark 1.3.
On the coding agent index, the one that matters for anyone paying for software work, Fable 5.1 leads at 70 and Astra sits at 67 alongside Opus 5 and Fable 5.
Another interesting number sits underneath all of this.
Astra does its coding work with a third of the tokens GPT-5.6 Sol needed and a fifth of the tokens Claude Opus 5 burns, so on Artificial Analysis's per-task costing, it matches Fable 5 for less than half the money.
GPT's list price is identical to Fable 5.1 at $14 per million tokens in and $70 out. The two labs have priced their flagships to the dollar and are now competing on how much work each dollar does.
We are well and truly into outcome economics now. If you are building AI systems, you will be looking at outputs per token cost soon if not already.
The price war has moved from list price to tokens per task. Both flagships cost exactly the same per million tokens, but ChatGPT uses a fifth of the tokens to do the same job.
Anyone negotiating an AI contract this quarter should be pricing per completed task and asking for token consumption data by workload.
In other news
Nvidia signed its definitive agreement to buy Hugging Face for $12.9bn, which is around 86 times the platform's annual revenue, with closing expected in the first half of 2027.
Commerce Secretary Howard Lutnick told the G20 innovation ministers that "we trust Anthropic", a few weeks after his department briefly banned foreign access to the company's top models and a federal judge ruled the Pentagon's Anthropic blacklist unconstitutional. The ban is still in place, however.
And Sony Music Publishing and Warner Chappell are suing Anthropic and its founders personally over song lyrics, seeking up to $150,000 a song.
Finally, John Ternus took over as Apple chief executive and inherited the only large technology company without a frontier model of its own. Although everyone seems to be buying Mac Minis to run open-source and Chinese AI models on.
Have a lovely weekend
Dan
Executive education is crucial for business leaders to stay on top of how AI will change their careers and business.
The Executive Summary is launching a series of Executive Education courses and cohorts, starting with the EXCO AI Hackathon - a day to learn, build and change the way you operate your business with AI at its heart.
If you would like to learn how to make agents work for your board or leadership team and create a future-ready business, please reply to this email for more details.
