SpaceXAI's Grok 4.5 ranks fourth on Artificial Analysis benchmark
The benchmarker put Grok 4.5 behind Fable 5, GPT-5.5 and Opus 4.8, with cost as the clearest part of SpaceXAI's pitch.
By Ryan Merket · Published
Why it matters
Grok 4.5 gives SpaceXAI a credible price-performance story, but Artificial Analysis' hallucination data shows the tradeoff behind the benchmark gains.
Elon Musk (@elonmusk)'s SpaceXAI released Grok 4.5 on Wednesday, July 8th, and Artificial Analysis (@ArtificialAnlys) ranked the model fourth on its Intelligence Index, behind Fable 5, GPT-5.5 and Opus 4.8.
The score is the cleanest outside read so far on SpaceXAI's launch claim. SpaceXAI is selling Grok 4.5 as its strongest model for coding, agentic tasks and knowledge work. Artificial Analysis gave Grok 4.5 a score of 54 on the Artificial Analysis Intelligence Index, placing it fourth on the leaderboard cited in the thread and listing the model as number 4 out of 170 in the model summary on its Grok 4.5 model page.
Musk framed Grok 4.5 before release as an "Opus-class model" with better speed, token efficiency and cost. That is the commercial claim SpaceXAI needs to prove. Frontier models are increasingly sold through the economics of work completed, not only a top-line benchmark number, and SpaceXAI's own launch page puts Grok 4.5 at $2 per million input tokens and $6 per million output tokens.
Artificial Analysis' numbers give SpaceXAI useful support on that price-performance argument. The evaluator said Grok 4.5 cost $0.31 to run per Intelligence Index task, less than GLM-5.2 and Kimi K2.6, and five times lower than Claude Sonnet 5 (max) while scoring higher on the Intelligence Index. Its public model page separately lists $600.92 as the total cost to evaluate Grok 4.5 on the Intelligence Index, with input pricing of $2 per million tokens and output pricing of $6 per million tokens.
SpaceXAI also gets a more specific win in coding-agent economics. Artificial Analysis said Grok 4.5 in Grok Build ranks on par with GPT-5.5 in Codex on its Coding Agent Index, which combines DeepSWE, Terminal-Bench v2 and SWE-Atlas QnA. The cost gap is the sharper datapoint: Artificial Analysis measured Grok 4.5 in Grok Build at $2.49 per task, compared with $11.80 for Fable 5 in Claude Code and $5.07 for GPT-5.5 in Codex.
That matters because SpaceXAI is packaging Grok 4.5 as an agent model rather than a chatbot refresh. The SpaceXAI launch post says Grok 4.5 was trained alongside Cursor, is the default model in Grok Build, and is available through Grok Build, Cursor and the SpaceXAI console. SpaceXAI also says EU availability is expected in mid-July, leaving a major enterprise market outside the first wave.
Cursor described Grok 4.5 as a mixture-of-experts model trained jointly with SpaceXAI and said the training included trillions of tokens of Cursor data capturing developer interactions with codebases and software tools. Cursor also said Grok 4.5 is its first model built for a broader set of work than software engineering, including data science, finance and legal work.
Cursor's post includes a caveat that should travel with any benchmark-heavy reading of the release. Cursor said Grok 4.5 had an advantage on CursorBench because an earlier snapshot of the Cursor codebase was accidentally included in training, and that CursorBench was excluded while the data is removed for future models. That caveat does not cover Artificial Analysis' full Intelligence Index, but it is a useful reminder that model releases increasingly depend on benchmark selection, harness choice and data contamination controls.
Artificial Analysis' own breakdown also gives SpaceXAI a reliability problem to watch. Grok 4.5 improved by 8 points on the AA-Omniscience Index, moving from 18 for Grok 4.3 to 26 for Grok 4.5. Accuracy rose from 35% to 52%. The hallucination rate also rose from 25% to 54%, according to the evaluator's thread.
That split is the uncomfortable part of the result. Grok 4.5 appears to know more and answer more, while also being more willing to produce wrong answers. Artificial Analysis said larger models commonly know more and hallucinate more, but for enterprise work the second half of that pattern has direct product consequences. A coding agent that finishes tasks cheaply still has to avoid introducing errors that cost more than the tokens saved.
SpaceXAI's release page says Grok 4.5 was trained across tens of thousands of Nvidia GB300 GPUs and used reinforcement learning over hundreds of thousands of tasks centered on multi-step software engineering and technical work. SpaceXAI claims the model delivers roughly twice the token efficiency of comparable leading models and runs at fast-model speeds. Artificial Analysis measured output speed at 91.3 tokens per second on the model page.
The market read is straightforward: SpaceXAI is competing on unit economics while trying to stay within striking distance of the highest-scoring frontier models. Grok 4.5 did not take the top spot on Artificial Analysis' leaderboard. It did give SpaceXAI a benchmark-backed argument that a fourth-place model can be commercially dangerous if it is cheap enough, fast enough and integrated where developers already work.