Breakthroughs

Grok 4.5 in Cursor Cuts Token Costs, EU Blocked

By Alec Chizhik July 10, 2026 6 min read

Since this week Cursor offers a new model that changes the cost equation for day-to-day coding. Grok 4.5 from SpaceXAI, the provider previously known as xAI, is not promoting the highest benchmark score. It promotes token consumption per solved task. The catch for teams in the EU: At launch the model remains blocked here.

The Key Points at a Glance

Cost instead of records. According to SpaceXAI, Grok 4.5 solves tasks on the SWE-Bench Pro benchmark with around 15,954 instead of 67,020 output tokens. This reduces the bill per task to about a quarter.
Blocked in the EU. At launch the model is not usable in the European Union, neither within SpaceXAI products nor via the API. Release is expected for mid-July; a firm date is missing.
Trained with Cursor. Grok 4.5 is the first model that SpaceXAI trained jointly with the Cursor editor. It has been running there as a first-party model since launch.

Cost per Task Beats Record Scores

I work with coding models every day. The question is rarely which one scores the highest value on paper. It is which one delivers enough to carry in everyday use without blowing the inference budget. This is exactly where the new model sets in. SpaceXAI does not put the pure hit rate in the foreground, but the effort per task.

The provider itself supplies the key figure. On SWE-Bench Pro, a test for agentic software changes, Grok 4.5 consumes around 15,954 output tokens per task. For Opus 4.8 it is 67,020. In addition, the claim that many tasks are completed in under half the steps. Fewer steps means fewer intermediate states that need to be checked or reverted.

For operations this pays off directly. At API prices of around 1.85 euros per million input tokens and 5.55 euros per million output tokens, a quarter of the output volume has a noticeable effect on the monthly bill. The model operates as a Mixture-of-Experts with 500,000 token context and, according to the provider, around 80 tokens per second. In agentic workflows that require many consecutive steps, this speed counts more than a single benchmark point.

// Metric

one quarter

This is how far the output token consumption per task drops on SWE-Bench Pro according to SpaceXAI, compared with Opus 4.8: 15,954 instead of 67,020 tokens.

// Source: SpaceXAI, x.ai/news/grok-4-5

Why Cursor and Grok Were Trained Together

The context behind the model explains a lot. SpaceXAI is the company that was formerly called xAI. SpaceX acquired xAI in February 2026, including Grok and the X platform. The official rebrand followed on July 6. Domain and API continue to run under x.ai.

In mid-June, SpaceXAI announced the acquisition of Cursor, the AI code editor from Anysphere. Reports put the value at around 55 billion euros. Grok 4.5 is the first model trained jointly with Cursor under this setup. Since launch it has been available there as a first-party model: for tab completion, inline edits, chat, and the composer across multiple files. In the first week there is a double quota. This is not a subsequent API integration, but native support in the editor.

One detail belongs to an honest assessment. Cursor itself disclosed that an earlier snapshot of its own codebase unintentionally ended up in the training data. The affected benchmark value was therefore removed from the public charts. The case shows two things: Training data from real editor usage plays a role. And such incidents are documented today instead of being kept quiet.

In Europe the Door Remains Closed for Now

This is the part that weighs most heavily for readers of this magazine. At launch, Grok 4.5 is blocked in the European Union. Neither through the SpaceXAI products nor through the API can developers in the EU access it. Outside the EU the model is immediately ready for use. The reason is the compliance obligations of the EU AI Act for particularly capable models.

For teams in Germany, Austria and Switzerland this means waiting. The expected release is in mid-July. SpaceXAI does not name a binding date, nor a detailed reason. Anyone wanting to firmly integrate the model into their workflows must plan for this gap or stick with another model for now.

This is more than a side note. Anyone who aligns their workflows with a model that is not yet available in their own legal jurisdiction carries an availability risk. Most organizations will continue running other models in parallel and wait for the release before initiating larger changes. Sovereignty in this case simply means: not making oneself dependent on a launch date that one does not know.

What the Benchmarks Really Say

The numbers place the model in the upper field without putting it at the top. On Terminal-Bench, Grok 4.5 is almost on par with the strongest providers. On SWE-Bench Pro it falls behind Opus 4.8. It does not reach the leader Fable in either of the two tests.

Benchmark	Grok 4.5	Opus 4.8	GPT-5.5	Fable
Terminal-Bench 2.1	83.3	78.9	83.4	84.3
SWE-Bench Pro	64.7	69.2	58.6	80.4

Pass rate in percent on agentic coding benchmarks, Grok 4.5 highlighted. All values according to SpaceXAI.

All these numbers are reported by the provider. They come from tests that SpaceXAI itself ran or published. They are useful as a rough orientation. They do not replace an own measurement on the relevant code. For practice, the token consumption is often more meaningful than the pure hit rate anyway. A model that produces significantly less output at a similar success rate changes the cost structure at high volume more than a few percentage points more.

Who Benefits from Switching Now

Four questions help with the decision. First: Does the majority of daily work anyway run in Cursor? Then the benefit is greatest once the EU release arrives, because first-party integration and joint training reduce friction. Second: Are token costs a real line item in the budget? Then it is worth comparing the efficiency claims against your own measurements.

Third: Is the team based in the EU? Then the expected mid-July start belongs firmly in the planning, not in hope. Fourth: Does the organization anyway regularly run its own evaluations? Then Grok 4.5 belongs on the list as soon as access is available.

Anyone achieving stable results today with another strong model and without acute cost pressure does not need a quick switch. The advantage shows especially at high volume and with tasks involving many iteration steps. In Cursor the model is included via the existing subscription as part of the first-party pool. This shifts the calculation once more compared to pure API usage.

// holds

Output consumption drops to about a quarter, reducing ongoing costs at high volume
Native first-party integration in Cursor saves setup effort for editor teams
The disclosed training data case provides an honest look at the data situation

// open

Blocked in the EU at launch, no binding release date available
All efficiency and benchmark values are reported by the provider, not independently verified
Behavior on complex internal projects over longer periods remains open

Frequently Asked Questions

What is Cursor?

Cursor is the AI-powered code editor from Anysphere. It extends a familiar editor interface with code completion, chat-based editing, and agentic changes across multiple files. SpaceXAI acquired the company in mid-June 2026. Grok 4.5 has been running there as a first-party model since launch.

Why is Grok 4.5 not available in the EU?

SpaceXAI has blocked Grok 4.5 at launch in the EU. The reason is the compliance requirements of the EU AI Act for particularly capable models. A release is expected for mid-July 2026. The provider has not named a fixed date.

What does token-efficient mean in concrete terms?

It means that the model produces significantly fewer output tokens for comparable tasks. On SWE-Bench Pro, according to SpaceXAI, it is around 15,954 tokens per task instead of 67,020 with Opus 4.8. This corresponds to about a quarter of the consumption and shorter processing chains. Fewer tokens mean lower costs per solved task.

When will Grok 4.5 arrive in the EU?

The release is expected for mid-July 2026. There is no fixed release date. Teams in Germany, Austria and Switzerland must rely on other models until then.