The new Grok 4.5 version has been released. Elon Musk says he’s competing with Claude Opus last year

Elon Musk’s SpaceXAI released the Grok 4.5 on Wednesday, its first public model since the SpaceX-xAI merger closed in February and SpaceX’s pending $60 billion deal to acquire Cursor. It targets programmers, engineers, and what the company calls “knowledge workers” — a category that apparently covers everyone from software developers to lawyers who review contracts for funding teams building Excel models.

The company’s pitch is not that it is the best model. It’s cheap for the Western model at least. The cost of Grok 4.5 is $2 per million input codes and $6 per million output codes. Cloud Opus 4.8, Anthropic’s flagship, manages $5 inputs and $25 outputs. GPT 5.6 Sol, OpenAI’s new high-level model that also launched on Wednesday, is priced at $5 for input and $30 for output.

Musk posted on X and explained where his new model actually exists. He described it as “almost comparable to Opus 4.7, but much faster.” Opus 4.7 is Anthropic’s former flagship. Opus 4.8 It has been successful ever since. Cloud Fable 5 It is now Anthropic’s best deals.

He framed it as a deliberate trade-off: speed and cost at the expense of raw capability, with engineers at Tesla and SpaceX as evidence of real-world utility.

What the standards actually show

SpaceXAI published four benchmark results at launch, and the picture was mixed. DeepSWE 1.1 measures how reliable the AI ​​is at closing real software bugs submitted by developers, using a standardized testing setup so that models can be compared fairly, and is scored by the percentage of issues fixed. Grok 4.5 scored 53%, behind Claude Opus 4.8 at 59% and GPT 5.5 at 67%. Claude Fable 5, the Anthropic Frontier Model, topped the chart with 70%.

On SWE Bench Pro, another benchmark that measures a set of software engineering problems scored by accuracy rate, the Grok 4.5 scored 64.7%, which is enough to beat the GPT 5.5 with 58.6% on this particular test. Opus 4.8 still leads with 69.2%, and Fable 5 with 80.4%.

The company’s benchmarks are being compared with GPT 5.5, not GPT 5.6, because the latter was also launched on Wednesday, hours after Grok 4.5 was announced.

SpaceXAI trained Grok 4.5 in collaboration with its recently acquired Cursor AI on tens of thousands of Nvidia GB300 GPUs inside Colossus, the Memphis supercomputer with a total capacity across more than 200,000 GPUs. Labs whose models exceed those same standards don’t have anything close to those devices. The model that was introduced is competitive, but it is not the first.

This pattern was followed Grok across multiple versions. SpaceXAI has consistently emerged with tremendous computing results and secured third place. What has changed with Grok 4.5 is the pricing and training signal.

Where the case for her actually holds

The better argument isn’t raw performance; It’s the mathematics of efficiency. In SWE Bench Pro tasks, Grok 4.5 used an average of 15,954 output tokens to complete each task. Opus 4.8 consumed 67,020 tokens for the same work, a 4.2x difference.

For teams using AI in large quantities, this difference is compounded in real savings on top of the already low price per token. Even with Grok 4.5 scoring very low on the quality criteria, cheaper tokens and more efficiency of use allow for more iterations without spending a lot.

The form also runs at a rate of 80 characters per second, which is fast form territory. Grok 4.5 is trained on developer session data from the benchmark, including debugging and editing traces of real code rather than static repositories. As musk He confessed in courtXAI’s training practices have come under scrutiny before; This time the pipeline goes through SpaceX’s entire buyout platform.

For developers doing high-volume coding jobs, the math works: Opus 4.7 has approximately 60% less power per input code. For anyone chasing the boundaries, Claude’s Fable 5 tops every category SpaceXAI has chosen to deploy. Our quick test using Grok’s design on Hermes was: Disappointing For creative writing and acceptable In a simple coding task.

The template is available via the Application Programming Interface (API), on the Hermes and Grok designs with half a million context tokens (just under 400,000 words). European users will have to wait before using it; SpaceXAI says Grok 4.5 will arrive in the European Union in mid-July.