Grok 4.5: xAI's Bet That Token Efficiency Beats Raw Benchmarks

xAI shipped Grok 4.5 on July 8, live in the API, in Cursor on every plan, and through the SpaceXAI console. The pitch isn't "we're the smartest model." It's "we finish your task using a quarter of the tokens the smartest model needs."

That's a different argument, and it's worth taking seriously.

The pricing undercut

Grok 4.5 runs $2 per million input tokens and $6 per million output tokens. That's over 60% cheaper than both Claude Opus 4.8 and GPT-5.5, priced against a market that's already in freefall.

The exact Opus and GPT-5.5 numbers move depending on thinking budget and caching, but the gap xAI is advertising is real and it's the headline for a reason.

Where the efficiency claim comes from

xAI's actual argument isn't the sticker price. It's tokens burned per completed task. On SWE-Bench Pro, xAI reports Grok 4.5 resolves the average task using 15,954 output tokens, versus 67,020 for Opus 4.8 running at max reasoning effort. That's roughly 4.2x fewer tokens for a comparable result.

Multiply a 4x price gap by a 4x token gap and the effective cost difference for agentic coding work is closer to an order of magnitude than a discount.

[!NOTE] Token efficiency and intelligence are different axes. A model that thinks less per step isn't automatically worse, but it does mean fewer opportunities to course-correct mid-task. Whether that trade-off works depends entirely on the task.

The benchmark picture is more mixed than the pricing story

On raw capability, Grok 4.5 lands in the same tier as GPT-5.5 and edges ahead of it on SWE-Bench Pro specifically. It carries an LMSYS Elo of 1462 and runs at roughly 80 tokens per second.

On the Artificial Analysis Intelligence Index, it scores 54, good for fourth place, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8. On the two hardest engineering benchmarks in that suite, Grok 4.5 falls noticeably behind Fable 5 specifically.

So the honest summary: Grok 4.5 is not the smartest model available in July 2026. It's the cheapest model that's still competitive, with a context window (500K tokens) larger than most of the field.

What this means if you're picking a model

• Long agentic sessions with tight budgets: Grok 4.5's token efficiency is the strongest argument for trying it. If your workload is mostly "grind through a large codebase," the 4x token multiplier compounds fast.
• Tasks near the frontier of difficulty: the benchmark gap against Fable 5 on the hardest evals suggests you shouldn't swap blind on your hardest problems without running your own eval first.
• Anything already tuned for Cursor: Grok 4.5 being live across all Cursor plans on day one removes the integration tax that usually slows adoption of a new model.

xAI is running the DeepSeek playbook from a different angle: instead of undercutting on training cost, it's undercutting on inference efficiency while staying inside API-only, closed-weight distribution. Whether that's enough to hold share against three labs with more capital and more benchmark headroom is the thing to watch over the next two quarters, not the launch week.