Grok 4.5 launches: fast coding model at a low price

What's new

SpaceXAI launched Grok 4.5 on July 8, 2026. The model accepts both text and images and is aimed at coding and agent tasks, meaning tasks where the AI works its way through multiple steps on its own. According to xAI's documentation, it has a context window of 500,000 tokens. A token is a small piece of text, and the context window is how much the model can hold in mind at once.

At launch, the model was available in Grok Build, in Cursor on all plans and in SpaceXAI's console. According to search results, it also powers the assistant on grok.com, X and in mobile apps. It was not available in the EU at launch. Decrypt reports that EU access is expected in mid-July 2026. Whether it is available in Denmark is not stated directly in the sources.

What's smart

The model is fast: Decrypt puts it at around 80 tokens per second. It is also frugal. On the tasks that were measured, Grok 4.5 used an average of 15,954 output tokens per task, while Claude Opus 4.8 used 67,020. That is roughly 4.2 times fewer. Fewer tokens typically mean faster responses and a lower bill.

Elon Musk wrote on X that the model is an "Opus-class" model, but faster, more token-efficient and cheaper. According to Decrypt, he has also said it is roughly on par with Opus 4.7, but much faster. According to The Next Web, the model was trained on compute that SpaceXAI also rents out to rivals Anthropic and Google, and it builds on data from Cursor.

Cheaper or better?

The API price (for developers and businesses) for prompts under 200,000 tokens is $2 per million input tokens and $6 per million output tokens. Cached input tokens, meaning text the model has seen before, cost $0.30. For prompts of 200,000 tokens or more, the price is $4 for input ($0.60 for cached) and $12 for output, and the higher rate applies to all tokens in the request. The monthly price for regular users has not been disclosed.

For comparison, Claude Opus 4.8 costs $5 for input and $25 for output per million tokens. According to Decrypt, OpenAI's GPT-5.6 Sol costs $5 and $30, while The Next Web lists GPT-5.6 Luna at $1 and $6.

On quality, Grok 4.5 is close, but not at the front. On DeepSWE 1.1 it scored 53 percent versus 59 for Opus 4.8, 70 for Claude Fable 5 and 67 for GPT-5.5. On SWE-Bench Pro it got 64.7 percent versus 69.2 for Opus 4.8, 80.4 for Fable 5 and 58.6 for GPT-5.5. According to The Next Web, it beats Opus 4.8 on several benchmarks, but not the largest and newest models from OpenAI and Anthropic. The figures were selected and published by xAI itself.

What it's good for

Grok 4.5 is best suited to programming and agent tasks where speed and price matter. Developers can try it in Cursor, and companies with many automated calls can save money if the low token usage holds up in practice.

If you need the absolute strongest model for difficult code, Claude Fable 5 is ahead on both of the tests mentioned. Regular users will mainly encounter the model as the engine of the Grok assistant on X and in the app.