Grok 4.5: $2/$6 pricing undercuts Fable 5 and GPT‑5.5

xAI has released Grok 4.5, a model trained on tens of thousands of Nvidia GB300 GPUs and tuned for coding, agentic tasks, and knowledge work. The rollout pairs strong efficiency claims with mixed benchmark performance: Grok matches top rivals on some tests but trails on others, while offering substantially lower token pricing.

What does Grok 4.5 target and how was it trained?

Grok 4.5 targets coding, agentic tasks, and knowledge work, and xAI trained it using tens of thousands of Nvidia GB300 GPUs. xAI says it relied on heavy data filtering, deduplication, and domain-specific selection during training, then ran a reinforcement learning stage covering hundreds of thousands of tasks, mostly from software engineering, with automated scoring. The company also built infrastructure for asynchronous learning so agentic runs could stretch over many hours while training continued in parallel.

How does Grok 4.5 perform on benchmarks and where does it fall short?

Grok 4.5 posts mixed benchmark results: it nearly ties the leaders on Terminal Bench 2.1 but lags on software engineering tests. On Terminal Bench 2.1, Grok scores 83.3%, nearly matching GPT 5.5 at 83.4% and coming one point behind Fable 5 at 84.3%. On DeepSWE 1.1, which measures resolving real GitHub issues, Grok scores 53%, behind GPT 5.5 at 67% and Fable 5 at 70%. On SWE Bench Pro, a harder software engineering set, Grok scores 64.7% versus Fable 5 at 80.4%; Opus 4.8 posts 69.2% in its max configuration and GPT 5.5 xhigh scores 58.6% on the same test.

These splits show Grok can handle complex command-line and agentic scenarios competitively while conceding ground on deeper software engineering tasks. xAI also claims runtime efficiency advantages: Grok runs at 80 tokens per second and uses 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tasks, which the company says reduces overall cost per task.

How much does Grok 4.5 cost versus rivals?

Grok 4.5 charges $2 per million input tokens and $6 per million output tokens, substantially below competitors. Opus 4.8 is $5 input and $25 output per million tokens. GPT-5.5 and GPT-5.6 are priced at $5 input and $30 output per million tokens. Fable 5 sits at $10 input and $50 output per million tokens. Combined with Grok’s claimed token-efficiency, xAI positions Grok 4.5 as the cheapest option in its performance tier.

Why it matters

Lower absolute performance on some engineering benchmarks matters less if real-world users can complete tasks for a fraction of the cost. Grok 4.5 narrows or eliminates headroom on agentic and command-line tasks while exposing meaningful gaps on deeper software engineering challenges. For teams that run large volumes of automated tasks or integrate models into cost-sensitive workflows, Grok’s $2/$6 pricing plus the 4.2x token efficiency versus Opus 4.8 could materially reduce bills even if per-instance accuracy is lower on some benchmarks.

xAI’s training choices and asynchronous infrastructure indicate a focus on scalable, long-running agentic deployments rather than winning every engineering benchmark leaderboard.

What to watch

Grok 4.5 is available now through Grok Build, Cursor, and the xAI console, with plugins for Word, PowerPoint, and Excel. The model is not yet available in the EU; xAI is targeting a mid-July launch there. Watch for real-world adoption signals: enterprise usage via Cursor and Grok Build, whether Grok’s token-efficiency claims hold up at scale, and any EU launch details. Also watch whether improvements in software engineering tasks narrow the DeepSWE and SWE Bench Pro gaps over the next weeks.

Additional context: xAI trained Grok 4.5 alongside the code editor Cursor, which SpaceX acquired in mid-June for $60 billion in stock.