Summary
- On July 8, SpaceXAI launched Grok 4.5, priced at $2 per million input tokens and $6 per million output, significantly cheaper than similar offerings from Anthropic and OpenAI.
- Elon Musk announced on X that Grok 4.5 is "roughly comparable to Opus 4.7," Anthropic's prior model now replaced by Opus 4.8, emphasizing its speed and cost over performance metrics.
- Currently, Grok 4.5 is not accessible in the EU, but SpaceXAI anticipates European availability by mid-July.
Elon Musk's SpaceXAI introduced Grok 4.5 on Wednesday, marking its first public model release since the merger with SpaceX-xAI concluded in February and amid SpaceX's ongoing $60 billion acquisition of Cursor. This model targets software developers, engineers, and what SpaceXAI refers to as "knowledge workers," encompassing roles from legal professionals to finance teams.
Rather than claiming to be the top model, the company emphasizes Grok 4.5's affordability compared to other Western models. The pricing is set at $2 per million input tokens and $6 per million output tokens, while Anthropic's Claude Opus 4.8 costs $5 for input and $25 for output. OpenAI's GPT 5.6 Sol, which also debuted on the same day, is priced at $5 for input and $30 for output.
In his post on X, Musk detailed the model's standing, describing it as "roughly comparable to Opus 4.7, but much quicker." The previous model, Opus 4.7, has been succeeded by Opus 4.8, and Claude Fable 5 is now Anthropic's leading model.
Musk indicated that the design focuses on speed and cost rather than sheer performance, citing engineers from Tesla and SpaceX as examples of practical application.
Benchmark Performance
Upon its launch, SpaceXAI released four benchmark results, presenting a mixed performance landscape. In the DeepSWE 1.1 benchmark, which evaluates an AI's ability to resolve actual software bugs reported by developers, Grok 4.5 achieved a score of 53%, trailing behind Claude Opus 4.8 at 59% and GPT 5.5 at 67%. The leading model, Claude Fable 5, scored 70%.
In another benchmark, SWE Bench Pro, which assesses various software engineering tasks based on resolution rates, Grok 4.5 recorded 64.7%, outperforming GPT 5.5's 58.6% in this specific test. Nevertheless, Opus 4.8 led with 69.2%, while Fable 5 achieved 80.4%.
SpaceXAI opted to compare Grok 4.5 against GPT 5.5 instead of the newly launched GPT 5.6, which was announced shortly after Grok 4.5.
Grok 4.5 was trained in collaboration with Cursor AI using tens of thousands of Nvidia GB300 GPUs within the Colossus supercomputer, which boasts over 200,000 GPUs. Competing models do not possess similar computational resources, making Grok 4.5 competitive, albeit not the leading model.
This trend has characterized Grok's previous releases, where SpaceXAI consistently utilizes substantial computational power but ranks third in performance. The notable differences with Grok 4.5 are its pricing strategy and training methodology.
Cost Efficiency
The stronger case for Grok 4.5 lies in its efficiency rather than raw performance. In SWE Bench Pro tasks, Grok 4.5 utilized an average of 15,954 output tokens per task, compared to Opus 4.8, which consumed 67,020 tokens for similar tasks, reflecting a 4.2x efficiency gap.
This disparity translates into significant savings for teams deploying AI at scale, especially considering the already lower token costs. Despite Grok 4.5's lower performance in quality benchmarks, its cost-effectiveness and efficiency allow for more iterations without incurring high expenses.
The model operates at a speed of 80 tokens per second, positioning it within the fast-model category. Grok 4.5 was developed using data from developer sessions at Cursor, which included debugging logs and actual code modifications rather than static repositories. Musk has faced scrutiny regarding xAI's training methods in the past, but this time the training pipeline is linked to a platform SpaceX is acquiring.
For developers engaged in high-volume coding tasks, the economics favor Grok 4.5, delivering approximately Opus 4.7 capabilities at a 60% lower cost per input token. However, for those pursuing cutting-edge advancements, Claude Fable 5 remains the top contender in all the categories SpaceXAI chose to analyze. Initial tests using Grok on Hermes yielded disappointing results for creative writing and adequate performance in basic coding tasks.
The model can be accessed via API, on Hermes, and through Grok build, offering a context window of half a million tokens (equivalent to nearly 400,000 words). European users will have to wait, as SpaceXAI has stated that Grok 4.5 will be available in the EU by mid-July.