Unconfirmed: Demis Hassabis reposted an X user's claim on November 21 that Gemini 4 Neon beats GPT-6.2 Astra, and Automatica has not confirmed it.
The material reviewed contains no statement from the developers of either model about the comparison. A repost is not an endorsement or confirmation by Hassabis or Google. The repost itself is verified, and the retrieved record shows 97 retweets of it.
@demishassabis reposted a message from @lyraxana at 22:53 UTC on November 21. The retweet, as retrieved, reads1: "After deeply testing Gemini 4 Neon for the past few days I can confidently say that this model beats GPT-6.2 Astra and Claude…" The retrieved text ends in an ellipsis after naming Claude. We have not seen @lyraxana's original post, so the full comparison and the testing method, including prompts, settings and tasks, are unknown.
The claim has some context in earlier coverage. Hassabis reposted a similar user claim about Fable 5.2, as Automatica Press reported on November 193. That report verified the repost but not the benchmark claim.
Separately, Google's model card, reported by Automatica Press on November 214, shows Neon ahead of Claude Opus 5.5 on 14 of 16 benchmarks and behind on two coding tests. Those results are vendor-reported and not independently verified. The coverage reviewed does not mention GPT-6.2 Astra, so it does not bear on this claim.
At 22:39 UTC the same day, Hassabis also reposted a reply by @OfficialLoganK to @SahilPanhotra2 describing the three models. It reads: "Carbon = big model smell / Neon = best all around / Hydrogen = quick and efficient model," with the original separated by line breaks. It does not mention Astra, and it is one user's description rather than a benchmark.
Automatica Press also reported a separate November 21 X post5 speculating that OpenAI could regain the lead after a Gemini turnaround. That post likewise has no independent corroboration in the available evidence.
Three things would confirm or kill the claim:
- An official statement from Google or from the developer of GPT-6.2 Astra.
- Published head-to-head benchmarks that name both model versions and state their configurations.
- Independent testing of the named versions, with the conditions disclosed, and the full text of @lyraxana's original post.
Until one of those appears, the claim rests on one user's testing, relayed by a high-profile account.