Grok 4.5: xAI's Coding Model Co-Trained with Cursor, Starting at $2
xAI releases Grok 4.5, co-trained with Cursor and built for real engineering tasks. API pricing is $2/$6 per million tokens, and it's live on all Cursor plans.


Grok 4.5: xAI's Coding Model Co-Trained with Cursor, Starting at $2
xAI releases Grok 4.5, co-trained with Cursor and built for real engineering tasks. API pricing is $2/$6 per million tokens, and it's live on all Cursor plans.
On July 8, xAI released Grok 4.5 β a coding model co-trained with Cursor. The story here isn't "yet another general-purpose frontier model"; it's that xAI fed real coding data from Cursor's developer base into the training set, with a very clear goal: build a coding model aimed at real engineering tasks and cheap enough to deploy at scale.
Core Positioning
Grok 4.5 is a mixture-of-experts (MoE) model aimed at coding, agent tasks, and knowledge work. xAI's release notes stress two points:
- Trained on tens of thousands of NVIDIA GB300 GPUs
- The training data includes trillions of tokens of real Cursor coding data
That means the model wasn't trained on synthetic data or contest problems β it's aligned directly with developers' real workflows inside the IDE.
Benchmarks on Real Engineering Tasks
Across several engineering benchmarks, Grok 4.5 performs as follows (official xAI figures):
| Benchmark | Grok 4.5 | Comparison (Fable/Claude Opus 4.8/GPT-5.5) |
|---|---|---|
| SWE Marathon (pass@1) | 29.0% | Opus 4.8 26.0% / Fable 24.0% |
| Terminal Bench 2.1 | 83.3% | Fable 84.3% / GPT-5.5 83.4% |
| SWE Bench Pro | 64.7% | Fable 80.4% / Opus 4.8 69.2% |
| DeepSWE 1.0 | 62.0% | Fable 66.1% / GPT-5.5 64.31% |
Worth noting: on the most hardcore engineering benchmarks like SWE Bench Pro and DeepSWE, Grok 4.5 still trails Fable 5 (Anthropic). But on long-horizon tasks like SWE Marathon, it beats both Opus 4.8 and Fable.
π‘ How to read these numbers: if you purely want the strongest coding capability, Fable 5 / the Claude line still lead. Grok 4.5's edge is value for money β the same budget runs more tasks.
Token Efficiency: 4.2x
This is Grok 4.5's hardest cost advantage. On SWE Bench Pro tasks:
- Grok 4.5: an average of 15,954 output tokens
- Opus 4.8 (max): an average of 67,020 output tokens
In other words, Grok 4.5 solves the same tasks with about 1/4.2 the tokens of Opus 4.8. Layer on the lower unit price and the real cost gap widens further.
Pricing
Grok 4.5's API pricing:
- Input: $2 per million tokens
- Output: $6 per million tokens
Model speed is 80 TPS (tokens per second) β the "fast model" tier.
A third-party cost analysis (the-decoder) ran the numbers on "cost per solved task":
- Grok 4.5 (Grok Build): $2.49 per task
- GPT-5.5 (Codex): $5.07 per task
- Fable 5 (Claude Code): $11.80 per task
Grok 4.5 comes in at roughly half of GPT-5.5 and a fifth of Fable 5.
How to Use It
Grok 4.5 is already available through:
- Cursor: usable on all plans (the core beneficiary of the xAI-Cursor co-training)
- Grok Build: Grok 4.5 is the default model, free for a limited time
- SpaceXAI console: grab an API key and call it directly
# Call via the SpaceXAI API (illustrative; see the x.ai API docs for details)
# 1. Create an API key in the x.ai console
# 2. Call Grok 4.5 over the OpenAI-compatible protocolβ οΈ Regional restriction: as of release, Grok 4.5 is not yet live in the EU region (including SpaceXAI products and the API console), with availability expected in mid-July.
Office Capabilities
Beyond coding, Grok 4.5 in Grok Build also handles Office-style tasks:
- Excel: builds complex financial/data models, uses formulas across sheets, leaves sticky-note comments
- PowerPoint: builds complex charts from native shapes, designs intuitive slides
- Word: writes clear long-form documents
Official plugins are available for Word, PowerPoint, and Excel.
Who It's For
- Cursor users: free on all plans, near-zero switching cost
- Teams deploying coding agents at scale: at $2.49 per task, well suited to running large volumes of SWE-style work
- Budget-sensitive developers: 4.2x token efficiency plus a low price suits high-frequency calling
- Users chasing peak coding capability: for hardcore tasks like SWE Bench Pro, Fable 5 / Claude is still the recommendation
References
Related articles

Claude Design Hands-On: Professional-Grade Design from a Single Prompt
Claude Design can build web pages, PPTs, prototypes, and even animated videos. This article collects the full set of use cases plus official practical tips, with the access URL and example prompts.

CLAUDE.md Goes Viral on GitHub: Four Rules to Keep AI Coding in Line
The CLAUDE.md configuration file, born from Karpathy's coding experience, hit number one on the GitHub trending chart, with 60,000 developers copying it. Four core principles to substantially raise your AI coding quality.

Fun-ASR1.5: An AI Speech Recognition Model That Accurately Transcribes Dialects Too
Alibaba's Tongyi Lab has released Fun-ASR1.5, a large speech recognition model supporting 30 languages and seven major Chinese dialect groups, with dialect character error rate down 56.2% and classical poetry accuracy at 97%. Includes demo links and API sample code.

An Anthropic Lead's Vibe Coding Masterclass
Anthropic researcher Erik Schluntz shares hands-on lessons on using Vibe Coding responsibly in production, covering the 22000-line code merge case, the leaf node strategy, and advanced tips.

AI Memory Enhancement Tools: A Roundup and Hands-On Guide
From Claude-Mem to DeepSeek DSA, a roundup of the mainstream AI memory enhancement tools of 2026, with principle comparisons and selection advice.

A Hands-On, End-to-End Guide to Producing AI Short Dramas
From script to finished film: make AI short dramas from scratch with Doubao, Xiaoyunque, Vidu and other tools, with tool comparisons and cost references.