Unconfirmed: Demis Hassabis reposted on 19 November a user's claim that Gemini Neon beat Fable 5.2 in his testing, which Automatica has not verified (X post1).
Hassabis reposted a post by Victor Taelin on 19 November that said, @demishassabis on X1:
RT @VictorTaelin: Gemini Neon just beat Fable 5.2 in nearly everything I tried tonight. Is this supposed to be a cheap model? What is going…
That establishes only that Hassabis amplified Taelin's claim. The retrieved post does not include any added comment from Hassabis, and it does not prove the result itself (X post1).
The post also leaves out the details that would matter in any performance comparison. Taelin's original post was not available to Automatica, and the reposted text is truncated. There is no published test setup, no prompt set, no task breakdown, no model identifiers beyond the words "Gemini Neon" and "Fable 5.2," no pricing tier, and no scores. Without those basics, readers cannot tell what "nearly everything" covered or whether the comparison was repeatable (X post1).
The wording about a "cheap model" is part of the claim and points to another missing variable: access tier. If Taelin was comparing different plan levels, rate limits, tool access, or preview builds, those conditions could affect the result. None of that is supplied in the repost (X post1).
Automatica previously reported that Google DeepMind introduced Gemini 4 Neon4 as a model for complex tasks (Automatica Press earlier coverage4). But the post Hassabis reposted says only "Gemini Neon." Automatica cannot confirm from the retrieved sources that Taelin meant Gemini 4 Neon specifically, so any link between those names is an assumption rather than a sourced fact.
That earlier Automatica report is background on Google's naming, not confirmation of Taelin's comparison. It did not test any Gemini model against Fable 5.2, and it did not document the conditions of any head-to-head benchmark.
Separately, Google announced that the Gemini app for Mac can now perform "computer use" tasks across local files, desktop apps, and the web, operating in an isolated window with user-controlled permissions and approvals (blog.google2). The Verge likewise reported that the feature is rolling out in English in the US for the macOS app (The Verge3). That product rollout is distinct from, and does not verify, Taelin's performance claim.
What would confirm or kill this rumor is fairly direct. Confirmation would require Taelin or another tester to publish the exact conditions: the full prompts or workload, model versions, pricing or access tier, whether tools were enabled, and the raw outputs or scores. Independent replication under the same settings would matter even more. The claim would be undercut if those details showed a narrow one-off workflow, missing parity between settings, or results that other testers could not reproduce.
On the present record, the narrow fact pattern is that Hassabis reposted Taelin's statement on 19 November (X post1). Beyond that, the evidence does not show whether the claim reflects broad model performance, a specific workflow, or a single night's informal testing.
Automatica has not independently reproduced the comparison, and the sources reviewed here do not include any public statement from Google about this specific Neon-versus-Fable claim.