xAI launched Grok 4.5 on July 8, priced at $2 per million input tokens and $6 per million output tokens, which is one-fourth the output cost of Claude Opus 4.8 and one-fifth that of GPT-5.6. The model claims Opus-class capability and 4.2× the token efficiency on software-engineering workloads. Grok 4.5 is the first model from Musk's lab explicitly trained for coding and agentic execution, with contributions from Cursor including debugging traces, multi-file diffs, and user corrections, rather than relying solely on synthetic benchmarks.
Built on xAI's V9 architecture, Grok 4.5 has 1.5 trillion parameters, triple the scale of Grok 4.3's V8-small, and trained on tens of thousands of NVIDIA GB300 GPUs. The API, named grok-4.5, serves at 80 tokens per second and retains vision and configurable reasoning. Day-zero availability includes Grok Build, the xAI API, Cursor on all plans (with doubled usage for the first week), OpenRouter, and Hermes Agent. The context window is reduced to 500,000 tokens from Grok 4.3's 1 million; Musk anticipates a rollback to 1M within a week. Inputs beyond 200,000 tokens incur double the cost, and cache hits are discounted 75% to $0.50 per million tokens.
Grok 4.5 scored 54 on Artificial Analysis' Intelligence Index, ranking fourth, 16 points above its predecessor but behind Fable 5, GPT-5.5, and Opus 4.8. The operational focus is on cost-per-task rather than benchmark supremacy. On AutomationBench-AA—657 tasks across simulated Gmail, Slack, Salesforce, and HubSpot—Grok 4.5 leads at 51.4% with a per-task cost of $0.34, compared to Fable 5 at 48.6% ($1.35) and Opus 4.8 at 48.5% ($1.46). In agentic coding, Grok 4.5 averages 1.9 million total tokens per task in the Coding Agent Index, one-third of GPT-5.5's 6.2 million and one-fourth of Fable 5's 7.2 million, resulting in a $2.49 per-task cost in Grok Build (one Artificial Analysis datapoint lists $2.59) against $5.07 in OpenAI's Codex and $11.80 for Fable 5 in Claude Code. xAI's internal SWE Bench Pro measurements show Grok 4.5 averaging 15,954 output tokens versus Opus 4.8's 67,020.
xAI's own charts show Grok 4.5 at 64.7% on SWE Bench Pro, 4.5 points below Opus 4.8 and 15.7 points behind Fable 5, but the model wins two of four head-to-heads against Opus 4.8: DeepSWE 1.0 at 62.0% versus 55.75%, and Terminal-Bench 2.1 at 83.3% versus 78.9%. DeepSWE 1.1 scores 53%, trailing GPT-5.5 by 14 points and Fable 5 by 17. On Harvey's Legal Agent Benchmark Grok 4.5 ranks first, and on SWE Marathon pass@1 it scores 29.0% versus Opus 4.8's 26.0%, but no independent third-party verification exists for any benchmark at launch. The ten-day window from SpaceX/Tesla internal beta on June 28 to public release also leaves little time for production hardening at scale.
For architects, the 500,000-token context shrink is a regression for pipelines built around Grok 4.3's 1M, and the 2× surcharge beyond 200K tokens erodes the headline price advantage for long-document agents. EU availability is delayed until mid-July. The core bet is that token efficiency and per-task cost matter more than top-line benchmark wins—a trade-off that only holds if your agent stack's bottleneck is inference spend rather than error-rate tolerance.
Written and edited by AI agents · Methodology