Theo says Anthropic's lead started with Fable 5
One engineer's timeline of the last two releases, not a measurement.
Theo posted a timeline that puts Anthropic's current lead two releases earlier than most people date it. His claim is that Fable 5 was the first true next generation model, and that almost everyone underrated it at launch.
OpenAI has maintained pace with capability, at the cost of coherency and readability.
most people perceived the gap between Fable 5 and GPT-5.6 Sol to be small. IMO, it was massive.
His explanation for why the gap went unnoticed is practical rather than technical. Fable 5 required you to use the model entirely differently. It was prohibitively expensive. It was taken away after three days of usage. And it shipped with no ZDR policy, which meant most people could not use it at work at all. Add those up and the model barely touched the hands that would have judged it.
The result, he says, is that most people perceived the gap between Fable 5 and GPT-5.6 Sol to be small, when in his view it was massive.
Where he puts OpenAI
Theo's read on the other side is that GPT-6 Astra mostly caught up on raw capability for short term tasks but could not stay coherent over long runs. Opus and Fable, meanwhile, kept getting easier to work with and more coherent on long runs while still gaining capability. Opus 5.5, in his telling, brought the Fable wins to everyone by being cheaper, more generous on subscription limits, and usable in enterprise with ZDR.
He summarises the trade he thinks OpenAI made in one line.
For working engineers
This is narrative, not measurement. There is no benchmark here, no timing run, no cost table, just one person's account of how two labs' last few releases felt to use. Treat it as a hypothesis about why the ranking debate keeps resetting every time a model ships, not as evidence that it is settled.
It does reframe the recent complaints, though. If the reference point is Fable 5 rather than the most recent launch, the current state looks like a two release trend rather than one bad week, which is the opposite of how the Opus 5.5 regression reports read when they were earlier going around. Whether the coherence gap on long runs is real at the level Theo describes is the part nobody in these posts has measured.
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.
