Labs

SpaceXAI ships Grok
4.5 at $2/$6,
one day before
OpenAI's GPT-5.6 launch

Elon Musk's newly public SpaceXAI released Grok 4.5 on Wednesday — the first model co-trained with Cursor since the $60 billion acquisition — pricing it well below Anthropic's Opus 4.8 and calling it "Opus-class, but faster."

SpaceXAI released Grok 4.5 on Wednesday at $2 per million input tokens and $6 per million output tokens, roughly 24 hours before OpenAI is expected to publicly launch GPT-5.6. The timing isn’t subtle, and the pricing is the news.

Anthropic’s Claude Opus 4.8, the model Elon Musk is explicitly positioning against, runs $5 in and $25 out. That makes Grok 4.5 more than four times cheaper on output. Anthropic Fable, the higher tier, sits at $10 and $50. OpenAI’s incoming GPT-5.6 Sol is priced at $5 and $30, with the smaller Luna variant at $1 and $6. Read as a matrix, Grok 4.5 is undercutting the frontier tier while sitting a hair above OpenAI’s small tier on input.

“It is an Opus-class model, but faster, more token-efficient and lower cost,” Musk posted on X, adding that the “internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster.” The benchmarks tell a more mixed story. Grok 4.5 scores 62.0% on DeepSWE 1.0, behind Fable’s 66.1% and GPT-5.5’s 64.31%. On SWE Bench Pro, it lands at 64.7%, trailing Opus 4.8’s 69.2% and Fable’s 80.4%. It leads on SWE Marathon pass@1 (29.0% vs. Opus 4.8’s 26.0%) and posts 83.3% on Terminal Bench 2.1. The pattern is legible: competitive on agentic and terminal work, behind on the harder software-engineering evals, priced as if the gap doesn’t exist.

The model is also the first co-trained with Cursor since SpaceX’s all-stock acquisition of Anysphere at a $60 billion valuation, per Bloomberg. “We’ve partnered with SpaceXAI to train Grok 4.5,” Cursor said in a statement reported by Reuters. Training ran on what SpaceXAI describes as “tens of thousands” of Nvidia GB300 GPUs. Availability is live on the SpaceXAI console and Grok Build, with EU access expected mid-July.

The competitive backdrop is where this gets interesting. Reuters reports GPT-5.6’s launch was pushed back last month after U.S. government requests tied to national-security concerns. That delay handed Musk a window, and he used it. Shipping the day before a rival’s launch is a familiar playbook. Google did it to Apple with Gemini in 2023, and Meta did it to OpenAI with Llama 3 in 2024. The move works less because it steals users and more because it forces the incumbent’s launch coverage to include a comparison table.

Which is, functionally, what the $2/$6 sticker is engineered to produce. The rebrand from xAI to SpaceXAI, the Cursor integration, the Opus-class framing, none of it lands without the price point doing the analytical work.

Sources