Grok 4.5 – a fast, cheap, token-efficient code model from XAI/Cursor

2026-07-09 Thu ⏱ 24 min t3dotgg

https://www.youtube.com/watch?v=U3uX115I9sY
Original title: Oh no (the new Grok model is good)

Theo reviews Grok 4.5 from SpaceX/XAI, now partnered with (and jointly trained by) Cursor. After 24 hours of real use he is genuinely impressed: it lands near-frontier intelligence at a fraction of the price and with unusually low token usage, making it a compelling default code model — though it lacks the orchestration/sub-agent sophistication of the newest generation (Fable 5, GPT-5.6).

The claims and benchmarks

Bold pitch: matches models like Fable 5 for dev work at a fraction of the cost. Cursor gave Theo early access thinking it was a new Composer; turned out to be Grok 4.5.
Artificial Analysis code index: neck-and-neck with GPT-5.5, just below Fable, beating Opus 4.8. Intelligence index score 54 → 4th place, behind GPT-5.5, ahead of Sonnet 5, crushing GLM-5.2.
Deepswe (SWE bench Theo trusts): 3rd behind Fable and GPT-5.5, massively beating Google models.
Whole new base model: 1.5T params (previous Grok ~500B), trained on tens of thousands of Nvidia GB300 GPUs, mixture-of-experts, ~80 tok/s. Trillions of tokens of Cursor data (real dev + agent interactions).

Price and efficiency

$2/M in, $6/M out under 200k context (doubles to $4/$12 for 200k–500k, which Theo finds a slightly greedy cliff). Compare Fable at $10/$50 — roughly 5–10x cheaper, and Grok is less token-hungry.
Token efficiency is the headline: ~2M tokens per coding task vs Fable 7.2M, Opus 9.2M, GPT-5.5-high ~6M. ~$0.31/task avg vs Fable $2.75. Feels faster, bills cheaper.
One of the most cost-efficient models for near-frontier intelligence.

The cursor-bench caveat

Cursor omitted Cursor Bench from marketing because an earlier snapshot of the Cursor codebase was unintentionally in training data, tainting the score. Theo credits them for being transparent rather than advertising the inflated number.

Real-world use

Hardening his "Lake Bed" cloud product: did a solid security/audit pass, found most of what Fable/GPT-5.6 found, and handled messy multi-turn back-and-forth (two lists with overlapping item numbers, vague prompts, screenshot input) without getting confused — where even GPT-5.5 tends to fixate on stale context. Made multiple PRs, addressed review comments, worked with Cursor's babysit skill.
3D: asked it to make his 2D fish game 3D — first model that is "almost decent" at 3D modeling in engines like three.js, beating Fable/GPT-5.6 despite rough creature models and swapped controls.

The verdict: great last-gen, not new-gen