Claude ‘Onigiri’ appears in Artificial Analysis code

Expand

Open original image ↗

A Claude model labelled Onigiri has turned up in Artificial Analysis’s public website code. The records list four effort variants, each with Opus 4.8 configured as its fallback. 1

We shared the finding on X on 15 October and checked the live JavaScript file again while preparing this article. The entries were still there. Our initial suspicion is that this could be Fable 5.5, though the name does not appear in the Onigiri records.

What the entries show

The full identifier is claude-onigiri-eap-2, with four versions ending in -low, -medium, -high and -xhigh. There is also an entry without an effort suffix. All five point to claude-opus-4-8 through a field named fallbackModelSlug. 1

The records contain per-evaluation data under rateByEval, as well as intelligenceIndexRate. Both use input, output and reasoning fields. Those numbers alone do not establish a benchmark score or a public price for the model.

Why Fable is a possibility

The Opus 4.8 fallback is the most useful clue. In its June report on Fable 5, Artificial Analysis described how that model could route flagged queries to Opus 4.8. It also said it had received access before the public release and benchmarked Fable with that fallback configured.

That gives us a precedent for both the fallback and advance testing. It still leaves the naming question open: an Opus 4.8 fallback does not uniquely identify Fable 5.5. The Onigiri records contain no release date, and their presence in the frontend does not tell us whether benchmarking is underway or finished.