SpaceXAI Releases Grok 4.5 With Performance Matching Claude Opus 4.8 at Lower Cost
Grok 4.5 undercuts leading competitors on price while matching or exceeding their benchmark scores in coding and specialized tasks, establishing a new price-performance baseline for developer-focused AI models.
Key Facts
- SpaceXAI announced Grok 4.5 on July 8, 2026, with a 1.5 trillion parameter base model further trained on Cursor data.
- API pricing for Grok 4.5 is $2 per million input tokens and $6 per million output tokens.
- Grok 4.5 scored 92.3% on the Harvey Legal Agent Benchmark, outperforming Claude Opus 4.5 (89.1%) and GPT-5 (87.6%).
- The model supports a 500,000-token context window and accepts text and image input.
- Grok 4.5 is available in Grok Build and Cursor plans, with EU release delayed to mid-July 2026.
Reporting from 3 sources: ASCII.jp, GameBusiness.jp, GIGAZINE.
SpaceXAI announced the release of Grok 4.5 on July 8, 2026, a top-tier AI model developed in collaboration with its subsidiary Cursor. The model is designed for coding, agent tasks, and knowledge work, and is priced below comparable models from OpenAI and Anthropic. According to SpaceXAI's benchmarks, Grok 4.5 scored 92.3% on the Harvey Legal Agent Benchmark, outperforming Claude Opus 4.5 (89.1%) and GPT-5 (87.6%). On the SWE-bench Verified coding test, it scored 85.4%, surpassing Claude Opus 4.5 (82.3%) and GPT-5 (80.1%). The model has a 500,000-token context window and supports text and image input. Elon Musk stated it is a 1.5 trillion parameter base model further trained on Cursor's data. API pricing is set at $2 per million input tokens and $6 per million output tokens. Independent tests by TryAI showed Grok 4.5 had the lowest latency and cost among tested models, though it failed on the first try for the most complex task. The model is available in Grok Build and Cursor plans, as well as via API, with EU availability delayed to mid-July 2026.
SpaceXAI's Grok 4.5 was announced on July 8, 2026, and began rolling out through partner services the next day. Japanese company SUPERNOVA Inc. added support for the model on its Stella AI and Stella AI for Biz platforms starting July 9. The model was developed with Cursor, a SpaceX subsidiary, and trained on NVIDIA GB300 NVL72 hardware. In addition to the Harvey Legal and SWE-bench benchmarks, SpaceXAI highlighted that Grok 4.5's average output token per task on SWE Bench Pro was 1/4.2 that of Claude Opus 4.8, indicating higher token efficiency. Third-party evaluation by Artificial Analysis ranked Grok 4.5's intelligence second only to Claude Opus 4.8 and GPT-5.5, above Claude Sonnet 5 and GLM-5.2. TryAI's hands-on test found that while Grok 4.5 failed on the first attempt to generate a 3D Rubik's Cube app, it succeeded on retry, and produced working particle gravity sandboxes and breakout games. The model's median latency was 2.8 seconds with a first-token time of 0.44 seconds, and throughput of 110 tokens per second, making it the fastest and cheapest in the comparison. SpaceXAI also demonstrated the model's ability to create a solar system simulation from a single short prompt. EU users will have to wait until mid-July for access due to regional release delays.
Synthesized by Yomimono from the 3 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.