Grok 4.5 Ships: SpaceXAI's Cheap, Fast Workhorse

Grok 4.5 is SpaceXAI's newest frontier model — a mixture-of-experts system co-trained with Cursor on trillions of tokens of coding data, launched on July 8, 2026. It is the first model the Cursor–SpaceXAI team built for more than software engineering, and it lands at a deliberately disruptive price. But the honest headline is not benchmark supremacy: on the raw coding evals Grok 4.5 sits mid-pack, and Fable 5 leads all four.

That distinction matters more than it sounds. For a small agency running agentic coding day in and day out, the number that shows up on the invoice is not a two-point benchmark gap — it is cost per finished task. Grok 4.5 resolves a task with 4.2× fewer output tokens than Opus 4.8 at its maximum setting, at $2 in and $6 out per million tokens ($0.50 cached). Multiply that across a month of agent runs and the arithmetic bends hard.

This guide maps what actually shipped, the spec sheet that matters, the benchmark standings read honestly, why cost per task is the real story, what Grok 4.5 does beyond code, where it fits alongside Fable 5, Opus 4.8, and GPT-5.5, and the caveats worth flagging first — starting with the fact that it is not yet available in the EU.

On July 8, 2026, SpaceXAI announced Grok 4.5 with a single line: "Today, we're launching Grok 4.5." It is a mixture-of-experts model, trained jointly with SpaceXAI, and Cursor frames it as its most intelligent model and the first it has built for more than software engineering. The training set is the tell: trillions of tokens of Cursor data — real user and developer-agent interactions with codebases and tools — plus a deliberately broader mix of STEM tasks, research papers, and knowledge work.

That breadth is the deliberate break from its predecessor. Cursor's earlier coding model, Composer 2.5, was a specialist — fast, cheap, tuned for software engineering and nothing else. It remains on offer as a different weight class. Grok 4.5 keeps the coding strength but widens the aperture, which is why SpaceXAI positions it for coding, agentic tasks, and knowledge work in one model rather than three.

The compute story is characteristically SpaceXAI. The model was trained on tens of thousands of NVIDIA GB300 GPUs in SpaceXAI's Memphis data centers, with large-scale asynchronous reinforcement learning across hundreds of thousands of tasks. For the backstory on how a rocket company came to own a coding-agent lab, see our coverage of the $60B Cursor / Anysphere acquisition.

Strip away the marketing and the spec sheet is what an engineering team actually plans around. On the SpaceXAI API, Grok 4.5 is priced at $2.00 per million input tokens, $0.50 per million cached input tokens, and $6.00 per million output tokens. Inside Cursor, a "fast" variant is offered at $4 in / $18 out per million — the premium buys priority serving, not a different model.

The context window is 500,000 tokens, with a higher pricing tier for requests that exceed 200K. The model is served at 80 tokens per second — what x.ai calls fast-model speeds — accepts text and images and returns text, and exposes a reasoning_effort parameter, structured outputs, and native parallel tool calling. The model id is grok-4.5 (aliases grok-4.5-latest and grok-build-latest), it runs in the us-east-1 and us-west-2 regions, and the default rate limit is 150 requests per second.

Here is where the honest framing earns its keep. SpaceXAI reports Grok 4.5 against the strongest modes of its rivals on four coding evals, and on every one of them it is mid-pack. Fable 5 at its maximum setting leads all four. Opus 4.8 and GPT-5.5 each beat Grok on some. The table below shows the vendor-reported standings; Grok's column is highlighted, and the leading score in each row is bold.

Read the columns, not the highlight. On DeepSWE 1.1 — run independently by DataCurve on the mini-swe-agent harness rather than each vendor's own — Grok drops to fourth of five, behind Fable 5, GPT-5.5, and Opus 4.8. On the two benchmarks where it looks strongest, Terminal Bench 2.1 and SWE-Bench Pro, it is still third. This is a capable model that loses the accuracy race and wins a different one.

Price per token is a vanity metric; cost per finished task is the one that hits the invoice. Grok 4.5 resolves a SWE-Bench Pro task with roughly 15,954 output tokens on average, against about 67,020 for Opus 4.8 at its maximum setting — 4.2× fewer. x.ai puts it more broadly: roughly 2× the token efficiency of comparable leading models, solving tasks in under half the number of steps. Combine 4.2× fewer output tokens with a $6 output rate and the per-resolved-task math pulls apart fast, even after conceding a two- to four-point accuracy gap.

Project that forward and the structural point is the interesting one. SpaceXAI now owns the full stack — Colossus compute, the Grok model, Cursor as the tool and the data source, and Grok Build as the distribution surface. A lab with a captive stream of real coding-agent interactions can bend its price-performance curve faster than one buying data on the open market, which is exactly the flywheel we mapped in the pre-launch analysis of Grok 4.5's Cursor data engine — and the same concentration that makes it a lock-in risk worth pricing in. For teams weighing where a cheap, fast workhorse belongs in a multi-model setup, that trade-off is exactly what our AI transformation engagements are built to work through.

The reason Grok 4.5 is priced and positioned as a workhorse rather than a code specialist is what it does off the code path. Grok Build — the coding agent where Grok 4.5 is the default model — does not stop at repositories. It builds multi-sheet Excel models with working formulas and web research, leaving notes and "stickies" as it goes; it assembles PowerPoint decks using native shapes and diagrams; and it drafts Word documents in prose. Microsoft Office plugins for Word, PowerPoint, and Excel put the model inside the tools knowledge workers already live in.

The knowledge-work training shows up on the benchmarks that matter for that audience. x.ai reports Grok 4.5 at #1 on Harvey's Legal Agent Benchmark, and the broader data mix — data science, finance, legal — is what lets Cursor claim a model built for more than software engineering. For the mechanics of the agent itself, and how Grok Build runs parallel coding agents from the CLI, see our Grok Build CLI deep-dive.

So where does a cheap, fast, mid-accuracy workhorse actually belong? Not as a replacement for your top-accuracy model — as a second tier you route the high-volume, cost-sensitive work to. The matrix below is the practical decision.

On distribution, SpaceXAI has been aggressive. Grok 4.5 is available in Cursor on all plans, across desktop, web, iOS, CLI, and SDK, with included usage doubled for the first week. It is the default in Grok Build, reachable from the SpaceXAI console, and free to use for a limited time at launch in both Grok Build and Cursor. Cursor also added new cybersecurity safeguards reflecting the model's capabilities. For the wider field it now competes in, our Q2 2026 agentic-coding platform matrix maps twenty tools side by side.

Four things to weigh before you wire Grok 4.5 into anything that matters. The first is regional, and for an EU-based team it is the big one.

The fourth caveat deserves more than a line. One owner now controls the compute (Colossus), the model (Grok), the tool and its data (Cursor), and the distribution surface (Grok Build). That vertical integration is what makes the price-performance story credible — and what makes it a lock-in risk. We argued both sides in the pre-launch flywheel and lock-in analysis; the launch confirms the flywheel is real, which means the counterweight is too.