SpaceXAI has launched Grok 4.5, their first Opus-class model following the acquisition of Cursor. The lab claims to be moving faster than any other frontier AI company.
SynthePulse Insight · AI deep reading
Version 1 · 4 sources
On July 8, 2026, xAI released Grok 4.5, positioning it as its first model specifically trained for coding and agent tasks, emphasizing cost efficiency and speed over absolute benchmark leadership.
On July 8, 2026, xAI (under the name SpaceXAI) officially released Grok 4.5, its first model specifically trained for coding and agent tasks, co-trained with Cursor. xAI officially states it offers "frontier intelligence with leading speed and cost efficiency." Elon Musk previously teased on X that Grok 4.5 is "Opus-level," but faster, more token-efficient, and lower cost.
xAI clearly positions Grok 4.5 as the "best balance of capability and cost," rather than absolute benchmark leadership. Musk stated the model is "roughly equivalent to Opus 4.7, but much faster," and emphasized its utility for Tesla and SpaceX engineers. Cursor calls it "our most powerful model" and "the first model built to go beyond software engineering."
On the Artificial Analysis Intelligence Index, Grok 4.5 scores 54, ranking 4th, behind Fable 5, GPT-5.5, and Opus 4.8, but 16 points higher than Grok 4.3. On the Coding Agent Index, it scores 76 in the Grok Build environment, tying with GPT-5.5 in Codex, but lower than Fable 5 in Claude Code.
On the DeepSWE 1.0 (pass@1) benchmark, Grok 4.5 scores 62.0%, lower than Fable max (66.1%) and GPT 5.5 xhigh (64.31%), but higher than Opus 4.8 max (55.75%) and Opus 4.7 max (40.12%). Additionally, it leads on the τ³-Banking task with a score of 33%, higher than GPT-5.5 (xhigh)'s 31%.
Notably, the OpenAI evaluation team has pointed out that SWE-Bench Pro is saturated/has fundamental flaws, so benchmark comparisons should be interpreted with caution.
Grok 4.5's official pricing is $2 per million input tokens and $6 per million output tokens, significantly lower than GPT-5.6 ($5/$30) and Opus 4.8 ($5/$25). Cache hits receive a 75% discount to $0.5 per million tokens, but long inputs exceeding 200k tokens incur double the cost.
Efficiency metrics are notable: on Intelligence Index tasks, average output tokens are about 14k, over 60% lower than Opus 4.8; on Coding Agent tasks, average total tokens are 1.9M, far below Fable 5 in Claude Code's 7.2M and GPT-5.5 in Codex's 6.2M. Cost per Intelligence Index task is $0.31, and per Coding Agent task is $2.59.
xAI emphasizes that Grok 4.5 uses half the tokens per task, has higher output speed, and costs less than half of comparable models.
Grok 4.5 has 1.5T parameters, 3 times that of Grok 4.3, placing it in the flagship model category. The context window is currently 500k, lower than Grok 4.3's 1M, but Musk stated it is expected to be restored to 1M next week. The model supports vision input and configurable reasoning.
At launch, Grok 4.5 is available in the SpaceXAI console, Grok Build, and Cursor, with all limits reset to provide free trials. Hermes Agent offers day-one support, accessible via Nous Portal, Grok/X subscription, and OpenRouter. The EU region is not yet available, expected to open later this month.
Cursor clarified that Grok 4.5 is in a different weight class from the Composer series (below 1.5T), and Composer 2.5 will continue to be available.
Primary information comes from xAI's official X account, Cursor's official statements, and third-party evaluator Artificial Analysis. Benchmark data is from Geek's X post, whose original source is not specified. Some details (e.g., parameter count, context window restoration plan) come from Musk's personal X account, which are first-party statements but not independently verified.
Grok 4.5 represents xAI's strategic entry into the coding agent space, leveraging significant cost efficiency advantages. While not the absolute performance leader, it occupies a favorable position on the price-performance curve.
Primary report
Primary source
Also covered by 2 sources