Grok 4.5

Grok 4.5 is SpaceXAI's July 2026 frontier model for coding, agentic tasks, and knowledge work. It was developed with Cursor and is the first Grok release after SpaceX agreed to acquire Cursor's parent company, Anysphere, in a USD 60 billion all-stock transaction.

Positioning

SpaceXAI markets Grok 4.5 as an "Opus-class" model. The more precise description is Opus-adjacent: it competes with Claude Opus 4.8 on some coding-agent evaluations but does not consistently beat it across SpaceXAI's own selected benchmark set.

The model's practical advantage is cost and output-token efficiency rather than undisputed benchmark leadership. It is priced at USD 2 per million input tokens and USD 6 per million output tokens, compared with Claude Opus 4.8 at USD 5 input and USD 25 output.

Benchmark profile

SpaceXAI's launch comparison showed Grok 4.5 beating Claude Opus 4.8 on Terminal-Bench 2.1 and DeepSWE 1.0, while trailing it on SWE-bench Multilingual and SWE-bench Pro. Grok 4.5 scored 83.3% on Terminal-Bench 2.1, 78.0% on SWE-bench Multilingual, 62.0% on DeepSWE 1.0, and 64.7% on SWE-bench Pro. The comparison used different effort settings across models, including high for Grok 4.5, max for Opus 4.8, and xhigh for GPT-5.5 where labeled.

Benchmark	Grok 4.5	Opus 4.8	GPT-5.5	Composer 2.5	Fable 5
Terminal-Bench 2.1	83.3%	78.9%	83.4%	73.0%	84.3%
SWE-bench Multilingual	78.0%	84.4%	77.8%	71.6%	Not reported
DeepSWE 1.0, Artificial Analysis	62.0%, high	55.8%, max	64.3%, xhigh	18.0%	66.1%, max
SWE-bench Pro	64.7%, high	69.2%, max	58.6%, xhigh	54.0%	80.3%

Artificial Analysis independently gave Grok 4.5 a score of 54 on its Intelligence Index, placing it fourth at launch behind Fable 5, GPT-5.5, and Claude Opus 4.8. It also measured output speed at roughly 92 tokens per second through the SpaceXAI API.

Benchmark results remain sensitive to harnesses, effort settings, and test contamination. Cursor disclosed that an earlier snapshot of its codebase had accidentally appeared in Grok 4.5's training data, giving the model an unclear advantage on CursorBench. Cursor said the data would be removed from future model training and that the benchmark would be expanded.

Token efficiency

On SWE-bench Pro, SpaceXAI reported that Grok 4.5 used about 15,954 output tokens per task, compared with 67,020 for Claude Opus 4.8 at maximum effort. That is roughly a 4.2-fold difference before accounting for Grok's lower output-token price.

This matters because benchmark pass rates do not capture the full economics of an agentic coding system. A model that reaches a slightly lower success rate with far fewer tokens may be more economical for production workloads, provided that its concise execution does not create additional review or retry costs.

Availability

At launch, Grok 4.5 was available through Cursor on all plans, Grok Build, and the SpaceXAI API console. European Union availability was expected in mid-July 2026.

Its strongest initial distribution is Cursor, where developers can select it alongside models from Anthropic, OpenAI, and other providers. That placement also creates a strategic tension: Cursor built its enterprise position around model neutrality, while its prospective owner now competes directly with the model providers that make neutrality useful.

Relationship with Cursor

Grok 4.5 is an early product of SpaceXAI's vertical-integration strategy. Cursor contributes developer distribution, coding workflows, evaluation infrastructure, and training signals. SpaceXAI contributes model research and large-scale compute.

The merger agreement values Cursor at USD 60 billion and is expected to close in the third quarter of 2026. Once completed, Cursor becomes a developer platform owned by one of the frontier-model companies it previously treated as interchangeable suppliers.

That ownership does not automatically remove competing models from Cursor, but it changes the incentives around defaults, feature access, infrastructure priority, and data. Enterprises evaluating Grok 4.5 therefore have two separate questions: whether the model is good enough for their workloads, and whether Cursor can preserve credible model neutrality under SpaceXAI ownership.

Practical use

Grok 4.5 is best suited to coding-agent workloads, terminal tasks, repository work, tool-using automation, and high-volume knowledge tasks where output-token cost matters. Its low price makes it attractive as a default or high-throughput model, with more expensive frontier models reserved for escalation.

The main caveats are limited independent auditing, benchmark contamination on CursorBench, and uncertainty about whether the reported token efficiency persists across messy production repositories and long-running agent loops. Teams should compare cost per accepted task, not just API rates or benchmark pass percentages.