When SpaceXAI put Grok 4.5 in front of the public this week, the hole in the story was independent proof. Elon Musk called it “an Opus-class model, but faster, more token-efficient and lower cost,” and the company’s own charts implied it could beat Anthropic’s Opus 4.8 outright — but the numbers were Musk’s own, with no outside scorecard to check them against. That scorecard is now in. Artificial Analysis, which runs one of the most-watched independent model rankings, placed Grok 4.5 fourth on its Intelligence Index at 54, trailing Anthropic’s Fable 5 (60) and Opus 4.8 (56) and OpenAI’s GPT-5.5 (55), while clearing every Google and open-weight model on the board. Fourth, not first. Priced the way SpaceXAI priced it, that might not matter.
Look at the sticker. Grok 4.5 runs $2 per million input tokens and $6 per million output. Opus 4.8 — the model Musk measured himself against — runs $5 and $25, so on the output tokens that dominate real workloads, Grok is close to a quarter of the price. Then a second discount stacks on top: on the SWE Bench Pro coding test, Artificial Analysis clocked Grok 4.5 resolving tasks with about 15,950 output tokens on average, against roughly 67,000 for Opus 4.8 — a bit over 4x fewer. Cheaper per token and fewer tokens per job is the same discount applied twice, and on token-heavy agent work the two compound. On the same firm’s Coding Agent Index, Grok 4.5 landed on par with GPT-5.5 — for a fraction of the spend to get there.
That math matters now in a way it didn’t a year ago. AI stopped being flat-rate software and became a metered utility, billed by the token like electricity — which is why companies from Tesla on down have started capping what employees can burn. Once usage is metered, “fourth-smartest at a quarter of the price” isn’t a consolation prize; it’s how you take the volume. The frontier is now a real four-way race — Anthropic, OpenAI and SpaceXAI all within six points on the same index, Google and the open field behind — and it’s moving weekly: OpenAI’s GPT-5.6 landed the very next day and reshuffled the top again. The leaders can out-score Grok. The open question is whether they can afford to out-price it.
Our take: The benchmark headline is a decoy. The number that reprices the industry is $6 per million output tokens for a model that scores within striking distance of the best on the board. SpaceXAI has effectively handed every finance team the story they want to hear — “frontier-ish for a quarter of the bill” — at the exact moment those teams started reading the bill. Anthropic and OpenAI can keep the intelligence crown. Defending their pricing is the harder job.
What to watch
- The price response. If Grok 4.5’s numbers hold up in production, the pressure lands on the output-token pricing of the models above it — the most profitable slice of the stack.
- Token efficiency in the wild. Lab tests like SWE Bench Pro are clean; real agent workloads are messy. The 4x token gap is the claim most worth stress-testing on your own jobs before you bank on it.
- The enterprise switch. “Good enough for a quarter of the cost” wins procurement, not leaderboards. Watch which teams quietly route high-volume, low-stakes work to the cheaper model.
- The pace. GPT-5.6 landed a day after Grok 4.5; Fable 5 sits on top for now. On this cadence, today’s ranking is a snapshot, not a standing.