New Grok 4.5 released, Musk says it\'s comparable to last year\'s Claude Opus

SpaceXAI releases Grok 4.5: Targeting programmers and knowledge workers, focusing on price/performance

On Wednesday, Elon Musk\'s SpaceXAI released Grok 4.5, the first public model since the merger of SpaceX and xAI was completed in February and SpaceX\'s $60 billion acquisition of Cursor was underway. The model\'s target users are programmers, engineers, and what the company calls \"knowledge workers\"-a category that clearly includes everyone from software developers to lawyers reviewing contracts to financial teams building Excel models.

The company\'s selling point is not that its model is top-notch, but that it is cheap-at least for Western models. Grok 4.5 charges $2 per million input tokens and $6 per million output tokens. Anthropic\'s main flagship model, Claude Opus 4.8, charges $5 for input and $25 for output. Also released on Wednesday, OpenAI\'s next-generation top-of-the-line model GPT 5.6 Sol is priced at $5 for input and $30 for output.

Musk posted on X clarifying the actual positioning of his new model. He called it \"roughly equivalent to Opus 4.7, but faster.\" Opus 4.7 was Anthropic\'s previous flagship model and was subsequently replaced by Opus 4.8. Currently Anthropic\'s top product is Claude Fable 5.

He describes this as a deliberate trade-off: trading raw performance for speed and cost, and points to Tesla and SpaceX engineers as proof of its real-world practicality.

Benchmark performance is uneven

SpaceXAI announced four benchmark results at the time of launch, with mixed performances. DeepSWE 1.1 testing measures the ability of AI to reliably fix real software vulnerabilities submitted by developers, uses standardized test settings to allow for fair comparison of models, and scores are calculated as the percentage of issues fixed. Grok 4.5 scored 53%, lagging behind Claude Opus 4.8 (59%) and GPT 5.5 (67%). Anthropic\'s cutting-edge model Claude Fable 5 topped the list with a score of 70%.

In another benchmark test, SWE Bench Pro, which measures the resolution rate of software engineering problem sets, Grok 4.5 achieved a score of 64.7%, which was enough to beat GPT 5.5\'s 58.6% in this particular test. But Opus 4.8 still leads at 69.2%, while Fable 5 reaches 80.4%.

The company\'s benchmark compares GPT 5.5 rather than GPT 5.6, because the latter was also released on Wednesday, but several hours after the Grok 4.5 announcement.

Huge investment in computing power, but not the top

SpaceXAI has trained Grok 4.5 on the Colossus supercomputer using tens of thousands of NVIDIA GB300 GPUs in cooperation with the recently acquired Cursor AI. The Memphis supercomputer has a total capacity of more than 200,000 GPUs. Model laboratories that rank higher on the same benchmark have far less hardware. The final released model is competitive, but not the first.

This pattern has been repeated in Grok\'s multiple releases: SpaceXAI always has huge computing power, but only gets third place. What makes Grok 4.5 different is pricing and training signals.

The real advantage: efficiency and cost

The better argument is not raw performance, but efficiency mathematics. Among SWE Bench Pro tasks, Grok 4.5 uses an average of 15,954 output tags to complete each task. Opus 4.8 requires 67,020 tags to complete the same work, a 4.2-fold difference.

For teams that need to use AI heavily, this difference will further translate into real cost savings based on the already low price per label. Even with lower scores on quality benchmarks, cheaper labeling and greater efficiency mean that more iterations can be done without costing too much.

This model also runs at 80 tags per second, making it one of the fast models. Grok 4.5 is trained based on developer session data from Cursor, including debugging traces and real code edits, rather than a static code base. As Musk admitted in court, xAI training practices have attracted attention before; this time, the training process is conducted through a platform that SpaceX is embarking on a wholly-owned acquisition.

This is a good deal for developers who need to handle high-volume coding tasks: at a 60% lower cost per input tag, you get capabilities roughly equivalent to Opus 4.7. For those pursuing cutting-edge performance, Claude Fable 5 remains ahead in all categories SpaceXAI chose to announce. A quick test we conducted using Grok built by Hermes showed that it performed poorly in creative writing and was acceptable on simple coding tasks.

Availability and Regional Restrictions

The model is accessible via an API and has a context window of 500,000 tags (just under 400,000 words) on the Hermes and Grok builds. European users will still have to wait some time to use it;SpaceXAI said Grok 4.5 will arrive in the European Union in mid-July.