Musk’s SpaceXAI launches lower-cost Grok 4.5 for coding tasks

Elon Musk’s SpaceXAI released Grok 4.5 on July 8, putting a cheaper coding-focused AI model into a market where pricing and performance are becoming investor-level questions. For retail investors watching AI spending, the launch is a reminder that model makers are competing on more than leaderboard scores: cost per use can shape adoption.

SpaceXAI priced Grok 4.5 at $2 per million input tokens and $6 per million output tokens. Tokens are the small text units an AI model reads and generates, so input tokens are what users send in and output tokens are what the model writes back. By comparison, Decrypt reported that Anthropic’s Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens, while OpenAI’s GPT 5.6 Sol is priced at $5 and $30, respectively.

The company is aiming Grok 4.5 at coders, engineers and “knowledge workers,” a category Decrypt said SpaceXAI applies to users such as software developers, lawyers reviewing contracts and finance teams working in spreadsheets.

Musk described the model on X as “roughly comparable to Opus 4.7, but much faster,” according to Decrypt. That comparison matters because Opus 4.7 is no longer Anthropic’s leading model. Decrypt reported that Claude Opus 4.8 has replaced it, while Claude Fable 5 is Anthropic’s top offering.

Benchmarks show a mixed position

SpaceXAI released four benchmark results with the launch, according to Decrypt. Benchmarks are standardized tests used to compare AI systems, though they do not capture every real-world use case.

On DeepSWE 1.1, a test that measures how often a model can fix real software bugs submitted by developers, Grok 4.5 scored 53%, Decrypt reported. That put it behind Claude Opus 4.8 at 59%, GPT 5.5 at 67% and Claude Fable 5 at 70%.

Grok 4.5 performed better on SWE Bench Pro, another software-engineering test measured by problem resolution rate. SpaceXAI reported a 64.7% score for Grok 4.5, ahead of GPT 5.5 at 58.6%, according to Decrypt. Claude Opus 4.8 led that test at 69.2%, while Claude Fable 5 scored 80.4%.

Decrypt noted that SpaceXAI’s comparisons used GPT 5.5 rather than GPT 5.6 because OpenAI’s newer model launched later the same day as Grok 4.5.

Compute scale meets pricing pressure

SpaceXAI trained Grok 4.5 with Cursor AI on tens of thousands of Nvidia GB300 GPUs inside Colossus, the Memphis supercomputer that Decrypt said has total capacity across more than 200,000 GPUs. GPUs, or graphics processing units, are the chips used to train and run large AI models.

Despite that hardware base, Decrypt reported that Grok 4.5 does not top the published software benchmarks. The company’s case is instead built around lower prices and faster responses, based on Musk’s comments.

Access is also limited by geography for now. Grok 4.5 is not available in the European Union, and SpaceXAI expects European access in mid-July, according to Decrypt.

This story draws on original reporting from Decrypt.