News · 2026-07-09

SpaceXAI ships Grok 4.5, trained on trillions of Cursor coding sessions

SpaceXAI released Grok 4.5 on July 8, 2026 -- its first model since becoming a public subsidiary of SpaceX -- and its pitch is aggressive pricing backed by an unusual training set. The model is priced at $2 per million input tokens and $6 per million output, far below Anthropic's Opus 4.8 ($5/$25), and Elon Musk calls it "Opus-class, but faster, more token-efficient and lower cost," later clarifying it is "roughly comparable to Opus 4.7."

Key facts

• What: Grok 4.5, a ~1.5-trillion-parameter V9 foundation model, priced at $2/$6 per million tokens (under 200K context).
• When: July 8-9, 2026, SpaceXAI's first model as a public SpaceX subsidiary.
• Training data: Trillions of Cursor developer-interaction tokens, plus RL from SpaceX and Tesla environments.
• Primary source: TechCrunch and the Hacker News discussion (759 points).

The genuinely interesting bet is the training data. Grok 4.5 was trained alongside Cursor, the popular AI coding editor, on what SpaceXAI describes as "trillions of Cursor interaction tokens" -- real developers accepting, rejecting, and editing AI suggestions in real projects. As one Hacker News commenter noted, the Cursor team's earlier Composer 2.5 model was well-received, so "a good score here doesn't come out of nowhere. Cursor has enormous amounts of training data to train good coding models with." SpaceXAI claims "twice greater token efficiency" versus other leading models.

The community's most useful contribution was puncturing the pricing headline. On Hacker News, commenter HarHarVeryFunny pointed to x.ai's own docs to show the $2/$6 rate only applies below 200K tokens of context; above that, up to the 500K maximum, it doubles to $4/$12. And Grok's cache hits are priced at 25% of the input rate, versus the roughly 10% that US frontier labs typically charge (DeepSeek and Xiaomi push it to about 1%). Cached tokens dominate long agentic runs -- the kind with dozens of tool calls -- so the real "cheap versus Opus" advantage for exactly the workloads Grok targets is smaller than the sticker suggests.

How it works, by analogy: a coding model trained on Cursor sessions is like a medical resident who trained by watching thousands of real surgeries with the outcomes attached, rather than only reading textbooks. The signal is dense and outcome-linked. That is a plausible reason to expect strong coding behavior -- but it is also why independent verification matters, since a model tuned on a specific tool's workflows can look better in that tool than in the wild.

That independent check arrived the same day. In AkitaOnRails' blind-audited coding benchmark, Grok 4.5 "finally pulled the Grok family into Tier A" but landed at the bottom of it, scoring 87 -- the same score cheaper models like Kimi K2.6 ($1) and GLM 5.2 (subscription) achieved, while Grok cost about $5.10 per run. The author flagged Grok 4.5 as an "irrational pick" on his value-per-price frontier.

Why it matters: Grok reaching the top tier is real progress, but the story is less "new frontier leader" than "the floor keeps rising." Capable coding models are becoming a commodity, and price, not peak quality, is the battlefield. One HN theory -- that xAI has a huge cluster with less traffic and can pass savings on -- was immediately countered: "Why would having more costs and less income allow them to pass savings on to the end user?" The honest caveat: independent LMSYS and Artificial Analysis benchmarks were still pending as of July 10, so the "Opus-class" claim remains, for now, the vendor's.

Primary source, verified: read the paper →

Ground Truth. (2026, July 9). SpaceXAI ships Grok 4.5, trained on trillions of Cursor coding sessions. Ground Truth. https://groundtruth.day/news/spacexai-ships-grok-4-5-cursor-trained.html

@misc{groundtruth:spacexai-ships-grok-4-5-cursor-trained,
  title  = {SpaceXAI ships Grok 4.5, trained on trillions of Cursor coding sessions},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/spacexai-ships-grok-4-5-cursor-trained.html}
}

Topics: SpaceXAI · Grok · model-release · coding · pricing