In short
- SpaceXAI released Grok 4.5 on July 8 for $2 million in entry tokens and $6 million in output, less than half the price of similar models from Anthropic and OpenAI.
- Elon Musk wrote on X that the version is “similar to Opus 4.7,” the previous version of Anthropic, which has now been replaced by Opus 4.8, while highlighting the speed and cost of the comparison.
- The Grok 4.5 is not currently available in the EU; SpaceXAI says European access is expected in mid-July.
Elon Musk’s SpaceXAI released the Grok 4.5 on Wednesday, its first public prototype since the SpaceX-xAI merger closed in February and SpaceX is awaiting a $60 billion deal to buy Cursor. It targets coders, engineers, and what the company calls “knowledge workers” — a group that apparently includes everyone from programmers to lawyers reviewing contracts to finance teams building Excel models.
The tone of the company is not the best example. That’s because it’s cheap, for a western model at least. Grok 4.5 costs $2 per million input tokens and $6 million for output. Claude Opus 4.8, Anthropic’s chief executive, invests $5 and outputs $25. GPT 5.6 Sol, the new high-end version of OpenAI that was launched on Wednesday, is priced at $5 and costs $30.
Musk wrote on the X and explained where his new model resides. He called it “similar to Opus 4.7, but much faster.” Opus 4.7 is an Anthropic history; Opus 4.8 you have won. Claude Fable 5 now it’s Anthropic above the supply line.
He pitched it as a deliberate trade-off: speed and value over raw power, with engineers at Tesla and SpaceX as evidence of real-world applications.
What the benchmarks show
SpaceXAI published four test results for the launch, and the picture is mixed. DeepSWE 1.1 tests how well AI can catch real software bugs presented by developers, using a test setup so that samples can be compared correctly, scored and the number of issues raised. Grok 4.5 scored 53%, behind Claude Opus 4.8 at 59% and GPT 5.5 at 67%. Claude Fable 5, an example of Anthropic limits, topped the chart at 70%.
On SWE Bench Pro, another benchmark that tests the collection of technical problems of software developed by interest, the Grok 4.5 posted a 64.7%, good enough to beat the GPT 5.5’s 58.6% in that test. Opus 4.8 still leads at 69.2%, and Fable 5 sits at 80.4%.
The company’s indicators are compared to GPT 5.5, not GPT 5.6, because the latter was launched on Wednesday, a few hours after the announcement of Grok 4.5.
SpaceXAI trained Grok 4.5 in conjunction with the newly acquired Cursor AI on Nvidia GB300 GPUs inside Colossus, a Memphis supercomputer with a total capacity of over 200,000 GPUs. The labs whose models rank above the same benchmarks have nothing close to those devices. The model that came out is competitive, not the first.
That example has followed Hold on to multiple releases. SpaceXAI has been constantly coming up with massive computing and third-party space. What changed with the Grok 4.5 were the prices and the training sign.
Where his case holds
A good argument is not a good argument; and effective mathematics. For the SWE Bench Pro tasks, the Grok 4.5 used 15,954 output tokens to complete each task. Opus 4.8 burned through 67,020 tokens for the same task, a 4.2x difference.
For teams running AI at scale, this difference is made into real money on top of the lower price per token. Even with the Grok 4.5 scoring the lowest in high-end benchmarks, the low-cost tokens and excellent usability allow for more iterations without spending too much money.
The model also runs at 80 symbols per second, which is part of the fast model. Grok 4.5 was trained on the development phase information from Cursor, including bug tracking and real-time code changes instead of repositories. Like Musk he pleaded guilty in courtxAI training systems have already been reviewed; this time the pipeline runs through SpaceX’s platform while it’s being bought.
For developers with high-volume recording tasks, the math works: Opus 4.7’s average performance is 60% less for each input signal. For everyone who is pushing the limits, Claude Fable 5 leads every team SpaceXAI decided to publish. Our quick test using the Grok build on Hermes was frustration for creative writing and acceptable for a simple writing task.
The example is available via API, on Hermes and Grok build with half a million symbols (a little less than 400,000 words). Users in Europe will have to wait before using it; SpaceXAI says the Grok 4.5 will arrive in the EU in mid-July.
Daily Debrief A letter
Start each day with top stories right here, including originals, podcasts, videos and more.
