SpaceXAI unveiled Grok 4.5 for coding and agents, with emphasis on price and speed.
SpaceXAI unveiled Grok 4.5, a new model for coding, agent tasks, and knowledge work. The company called it its most powerful system.
Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.https://t.co/i8HpU7w64k pic.twitter.com/oBjGtTsoNc
— SpaceXAI (@SpaceXAI) July 8, 2026
The model is available via the API, in Grok Build, and in Cursor across all plans. It is not yet available in the European Union, with launch expected in mid-July.
The release centers on cost and speed. Grok 4.5 is priced at $2 per 1 million input tokens and $6 per 1 million output tokens. For comparison, Anthropic’s Claude Opus 4.8 costs $5 and $25, respectively; Claude Fable 5 is $10 and $50; and OpenAI’s GPT-5.6 Sol is $5 and $30.
GPT-5.6 is currently in limited preview: OpenAI has opened the model family to a small group of trusted partners and plans to make Sol, Terra, and Luna publicly available in the coming weeks. Within the same line, Luna is priced lower — $1 per 1 million input tokens and $6 per 1 million output tokens.
According to Elon Musk, Grok 4.5 is an “Opus-class model, but faster, cheaper and more token-efficient.” He later clarified that, by SpaceXAI’s internal assessment, the system is “roughly comparable to Opus 4.7, but much faster.”
Based on strong positive feedback from customers in our beta test program, @SpaceXAI will make Grok 4.5 available to the public tomorrow.
It is an Opus-class model, but faster, more token-efficient and lower cost.
— Elon Musk (@elonmusk) July 8, 2026
Test results
In results published by SpaceXAI, Grok 4.5 is a mixed picture: the model did not lead every benchmark but was not clearly behind competitors.
On DeepSWE 1.0, Grok 4.5 scored 62%. It trailed Fable at 66.1% and GPT-5.5 at 64.31% but outperformed Claude Opus 4.8 at 55.75% and Opus 4.7 at 40.12%.
Source: SpaceXAI.
On DeepSWE 1.1 the result was lower: Grok 4.5 scored 53% versus 59% for Opus 4.8, 67% for GPT-5.5, and 70% for Fable. On SWE Bench Pro the model posted 64.7%. That is above GPT-5.5 at 58.6% and nearly on par with Opus 4.7 at 64.3%, but below Opus 4.8 at 69.2% and Fable at 80.4%.
On Terminal Bench 2.1 Grok 4.5 scored 83.3% — nearly on the level of GPT-5.5 at 83.4% but below Fable at 84.3%. On SWE Marathon it took first place with 29% versus 26% for Opus 4.8 and 24% for Fable. The company also said Grok 4.5 led on Harvey’s Legal Agent Benchmark.
Where SpaceXAI’s case is stronger
The stronger angle of the release is economics, not top scores. According to SpaceXAI, on SWE Bench Pro tasks Grok 4.5 used an average of 15,954 output tokens per task. Claude Opus 4.8 used 67,020 tokens — 4.2 times more.
Source: SpaceXAI.
For teams running AI across large task volumes, this affects total cost. With lower output token pricing, Grok 4.5 may be more cost-effective in scenarios with many iterations: fixing bugs, generating edits, code review, and agent loops.
SpaceXAI also cited a generation speed of about 80 tokens per second and a 500,000‑token context window. The company positions the model as a compromise between quality, speed, and cost.
Cursor connection
SpaceXAI said Grok 4.5 was trained together with the Cursor AI service. The company did not disclose the full dataset but said the model was built for programming, tool-using agents, science, engineering, and math.
According to Decrypt, training used developer session data from Cursor, including debug traces and real code edits, not just static repositories.
In June SpaceX signed an agreement to acquire the AI service at a $60 billion valuation. The deal will be in SpaceX Class A shares, with closing expected in Q3 2026 after regulatory approvals.
Grok 4.5 is one of the first major models after Musk’s AI group combined with SpaceX. To train it, the company used tens of thousands of GPU Nvidia GB300.
In February, media reported plans by the then separate SpaceX and xAI to develop software for Pentagon autonomous weapons.
Follow ForkLog on social media
Found a mistake in the text? Select it and press CTRL+ENTER