xAI released Grok 4.5 on July 8, a model trained alongside Cursor’s AI coding editor and priced at $2 per million input tokens and $6 per million output tokens. The release lands the same week as OpenAI’s GPT-5.6 Sol public launch, intensifying the pricing and capability competition across frontier coding models.
Training and Architecture
Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs using datasets spanning coding, science, engineering, and math, according to xAI’s announcement. The distinguishing factor is the Cursor partnership: xAI used feedback from Cursor’s coding agent product to tune the model specifically for multi-step software engineering workflows.
The training process scaled reinforcement learning across hundreds of thousands of tasks, with grading combining automated and model-based methods, MarkTechPost reports. The system supports asynchronous training where agentic rollouts can run for hours while learning continues.
Benchmark Results
xAI published scores across four coding benchmarks. The results are mixed. Google’s Fable (max) leads on all four benchmarks, according to xAI’s own published data. Grok 4.5 is closest on Terminal Bench 2.1 at 83.3% versus Fable’s 84.3%.
On DeepSWE 1.0 (pass@1), Grok 4.5 scores 62.0% compared to Fable’s 66.1% and GPT 5.5 (xhigh) at 64.31%. On SWE Bench Pro resolve rate, Grok 4.5 hits 64.7% against Fable’s 80.4% and Opus 4.8 (max) at 69.2%.
The standout number is token efficiency. On SWE Bench Pro, Grok 4.5 resolved tasks with an average of 15,954 output tokens. xAI reports Opus 4.8 (max) used 67,020 on the same benchmark, roughly 4.2x more. Fewer output tokens translate directly to lower cost per task and lower latency.
Pricing and the Competitive Picture
At $2/$6 per million input/output tokens, Grok 4.5 undercuts most frontier-tier competitors on list price. The model serves at 80 tokens per second. Combined with the token efficiency claims, xAI is positioning cost-per-task as the metric that matters, not raw benchmark position.
The model is available in Grok Build (xAI’s plugin marketplace with MongoDB, Vercel, Sentry, Chrome DevTools, Cloudflare, and Superpowers integrations), in Cursor on all plans, and through the xAI console. Free usage is available for a limited time in Grok Build and Cursor. EU availability is expected mid-July, according to MarkTechPost.
The Broader Pricing Pressure
Grok 4.5 joins a week that also saw Anthropic’s Claude Sonnet 5 launch at introductory $2/$10 pricing and OpenAI’s GPT-5.6 Sol receiving government approval for public release. The convergence is notable: three frontier labs now offer coding-optimized models at the $2/M input price point within days of each other. For teams building agent infrastructure, the question is shifting from “which model is best” to “which model costs least per resolved task,” and xAI’s token efficiency numbers are a direct bid for that framing.