Grok 4.5

Best for

Teams that need a documented SWE-bench Pro number from xAI's line rather than 4.6's unpublished claims

Price & access

$2 in / $6 out per 1M; cached input $0.30 (BenchLM notes this is a 10% cache-read estimate, not an xAI-published rate); higher-context pricing applies above 200K with an undisclosed surcharge

Evidence summary

A hard third-party SWE-bench Pro number plus two preference boards and first-party pricing, but the release date and AA index both have conflicting published values.

Edition status

No prior snapshot. Tracked in the September 2026 baseline.

Strengths

• SWE-bench Pro 64.7% (rank 10) beats GPT-5.6 Sol's 64.6% and Opus 4.7's 64.3%
• Roughly 2x token efficiency and under half the steps of comparable models per xAI
• Well-sampled Arena row (6,034 votes) and Design Arena 1301
• Same $2/$6 price as Grok 4.6

Tradeoffs

• Superseded by Grok 4.6, which scores higher on both preference boards at the same price
• AA publishes two conflicting Intelligence Index values (45 vs 56)
• 500K context with an undisclosed >200K surcharge
• Cached-input rate is an estimate, not an xAI-published price