Early Access Isn't a Conspiracy
OpenAI’s launch post for GPT-6.2 Astra and Sol went up at 12:54 PM Eastern on Tuesday. By 1:02, Ridge’s timeline has Matt Shumer saying he wouldn’t write his usual review because the models “were essentially unusable”, and at 1:23 Theo Browne called Astra “the worst model OpenAI has ever released”. Dan Shipper said Every’s team had tested both for weeks and found “a disappointing release compared to Fable 5.2”. That’s three well-known testers reaching one verdict inside half an hour, two days before Anthropic’s shares are due to start trading.
It wasn’t the whole reception. Simon Willison called the buzz “mixed”, Artificial Analysis has Astra second on its index behind only Fable 5.2, and Derya Unutmaz, who also had early access, told Shumer his own testing was “exactly opposite”. The bad reviews were the loud ones, though, and they came in a cluster.
The suspicion arrived almost as fast. Shopify’s MJ wrote that the whole thing “feels weirdly coordinated. Apparently everyone has been using it for weeks and all decided to rant about it on the same day.” A reply under Shumer’s post reads, in full, “I’m glad to confirm you are an Ant shill”.
I don’t buy it, and MJ’s own sentence is the reason. Everyone had been using it for weeks because OpenAI gave them early access, and they all posted on the same day because, almost certainly, it was the first day they were allowed to. I haven’t seen the terms those testers agreed to, but that’s how pre-release access has worked for every model I’ve followed. Praise from early testers arrives on the same clock, and nobody calls that coordinated. The Anthropic half of the charge is thinner still. All three rate Claude models higher right now, which is a preference, and none of the posts I found alleging a campaign offers evidence of anything more.
The complaints also don’t read like a script. Shipper’s notes say Astra matched its predecessor on Every’s coding benchmarks but “seemed to stop too early frequently” on real builds. Browne’s long piece, Why I hate GPT-6.2 Astra, is a letter he says he sent OpenAI during testing, and it stops to call Astra a “Great reviewer” that “catches things other models miss”. He adds that he’s “likely overreacting (I am admittedly known for that)”. A shill doesn’t send his notes to the vendor first. None of this rules out a group chat, but it leaves the accusation with nothing left to explain.
OpenAI has supplied two explanations of its own, and neither involves bad faith. Its system card, as MLQ.ai reports it, says Astra “omitted a known obstacle to completing its work in 22.28 percent of trials, compared with 11.09 percent for GPT-6 Astra”, and adds “We are concerned by this regression”. That’s close to Browne’s complaint that he couldn’t tell afterwards what the model had done. Then OpenAI’s Thibault Sottiaux said the testers saw an older build: “we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days”. The shipped model may well be better than the one they reviewed. Either way, they reported what they were given.
Sources:
-
GPT-6.2’s first hour: the boards like Astra, the early testers mostly do not — Ridge
-
‘Essentially unusable’: OpenAI launches GPT-6.2 Astra and Sol to negative reception from AI power users — VentureBeat
-
Why I hate GPT-6.2 Astra — Theo Browne on X
-
GPT-6.2 Astra and GPT-6.2 Sol — Simon Willison
Filed under AI & machine learning
This post is timestamped using Blockchain technology. Verify