Grok 4.5 places fourth on AA index, hallucinations double

xAI shipped Grok 4.5 with Elon Musk describing it as Opus-class, and the first big independent benchmark returned a more complicated verdict. On Artificial Analysis's Intelligence Index, the model scored 54 and placed fourth, behind Claude Fable 5, OpenAI's GPT-5.5, and Claude Opus 4.8, the same Opus that Musk had just claimed to match.

The number that has people talking is not the ranking. On the AA-Omniscience benchmark, Grok 4.5's factual accuracy climbed from 35% on Grok 4.3 to 52%. In the same run, its hallucination rate more than doubled, from 25% to 54%. As The Decoder put it, accuracy rose but the share of wrong answers delivered with confidence rose faster. The model knows more, and it also asserts more incorrect things more confidently.

The counterweight is price. Grok 4.5 lists at $2 per million input tokens and $6 per million output, against roughly $5 and $25 for Opus 4.8, $5 and $30 for GPT-5.5, and $10 and $50 for Fable 5. On the Coding Agent Index it matches GPT-5.5's Codex at 76 points, at $2.49 per task versus $5.07 and $11.80 for the two frontier alternatives. For a coding agent or bulk workflow where a downstream check catches errors, that arithmetic is aggressive.

The honest caveat is that a 54% hallucination rate is a hard constraint for any customer-facing surface, and pushing the model into those workflows without validation is asking for confident wrong answers. What the reporting doesn't tell you is whether the extra hallucinations concentrate in specific subject areas, or why accuracy and confident error climbed together in the same training run. The forward-looking read is that cost-sensitive buyers and coding tool vendors just got a much cheaper near-frontier option, and the shops selling validation and retrieval layers just watched their market get bigger.