# xAI Ships Grok 4.5, a Cursor-Trained Coding Model Priced Below Rivals

The model resolves SWE-Bench Pro tasks with about a quarter of the output tokens Anthropic's Opus 4.8 uses, even as it scores lower on the benchmark itself.

- Published: 2026-07-09T05:49:21.519Z
- Canonical: https://polylog.news/ai/2026-07-09/xai-ships-grok-4-5-a-cursor-trained-coding-model-priced-belo
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Polylog editors](https://polylog.news), [MarkTechPost](https://www.marktechpost.com/2026/07/08/spacexai-releases-grok-4-5/), [VentureBeat](https://venturebeat.com/technology/spacexs-grok-4-5-launches-at-half-the-price-of-rivals-heres-why-that-could-rattle-anthropic-and-openai)

xAI released [Grok 4.5](https://t.me/aipost/7461) on July 8, its first model built from the start for software engineering and autonomous agents. The training run was conducted alongside the code editor Cursor. The model is available in Cursor on every plan, in xAI's build tools, and through the application programming interface (API) at $2 per million input tokens and $6 per million output tokens, roughly [half the price of comparable frontier models](https://venturebeat.com/technology/spacexs-grok-4-5-launches-at-half-the-price-of-rivals-heres-why-that-could-rattle-anthropic-and-openai), and it runs at about 80 tokens per second.

The benchmark results are mixed, and they come from xAI's own numbers, which no third party has yet reproduced. On [SWE-Bench Pro](https://www.marktechpost.com/2026/07/08/spacexai-releases-grok-4-5/) xAI reports Grok 4.5 at 64.7 percent, behind Anthropic's Opus 4.8 at 69.2 percent and the Fable line above it, though it leads on the longer-horizon SWE Marathon test at 29.0 percent. The more consequential claim is efficiency. xAI says Grok 4.5 completes SWE-Bench Pro tasks using an average of 15,954 output tokens against 67,020 for Opus 4.8. That is a 4.2-times gap, and if it holds outside the vendor's own test setup, it changes the cost of running an agent far more than a few points of accuracy do.

The release comes the same week that Anthropic shipped Sonnet 5 and OpenAI questioned the reliability of SWE-Bench Pro itself, so the headline percentages should be read with the benchmark's known problems in mind. What is verifiable now is the price and the token accounting. The capability ranking is asserted by the vendor and awaits independent runs.

## What this means

Coding is now the explicit training objective of a third frontier lab, and the competition is shifting from raw pass rate to tokens per task, which is what appears on an enterprise inference bill. If Grok 4.5's efficiency claim survives independent testing, the competitive pressure on Anthropic and OpenAI comes through cost rather than capability. A buyer running thousands of agent loops a day cares more about a 4x token reduction than a 5-point benchmark difference. The Cursor co-training also ties the model's distribution to a specific tool, giving xAI a route to working developers it did not previously have.

## What to watch

- Independent reproductions of the token-efficiency figure on SWE-Bench Pro and Terminal-Bench, since a vendor-measured 4.2x gap can shrink sharply under a neutral test setup.
- Whether other editors and integrated development environments (IDEs) strike co-training or default-model deals, which would signal that distribution through tooling, not the API, is becoming the way coding models reach users.