Grok 4.5 Fights Opus 4.8 to a Draw on Coding Benchmarks, at Under Half the Price

TL;DR

On July 8, xAI shipped Grok 4.5, its new flagship, and Elon Musk called it "Opus-class." On the coding benchmarks xAI chose to publish, it fights Opus 4.8 to a 2-2 draw and trails Anthropic's Fable 5 outright. The reason to keep reading anyway: it lists at $2 per million input tokens and $6 per million output, against $5 and $25 for Opus, and xAI says it resolves a task on SWE-Bench Pro using roughly a quarter of the output tokens. The leaderboard gap and the invoice gap point in opposite directions, and for most teams shipping code the invoice is the number that wins the meeting.

What actually shipped

Grok 4.5 is xAI's new top model, live now in Grok Build, inside Cursor on every plan, and through the xAI console, with the public rollout following a day later. It runs at about 80 tokens per second and carries a 500K-token context window. Musk's framing was blunt: "It is an Opus-class model, but faster, more token-efficient and lower" cost, and "roughly comparable to Opus 4.7, but much faster."

"Opus-class" is doing a lot of work in that sentence, and Musk applies the label the way some people apply "artisanal," so the honest test is the numbers. Here they are.

The benchmarks: solidly mid-pack

On the four coding evaluations xAI published against Opus 4.8, it is a genuine tie. Grok 4.5 wins DeepSWE 1.0 (62.0 vs 55.75) and Terminal-Bench 2.1 (83.3 vs 78.9), and loses DeepSWE 1.1 (53 vs 59) and SWE-Bench Pro (64.7 vs 69.2). Two-two. That is a credible flagship, not a record-breaker.

Widen the field and the picture is less flattering. On SWE-Bench Pro, the hardest of the set, Grok 4.5's 64.7 sits behind both Opus 4.8 and Fable 5, and Fable's 80.4 is in a different weight class. On the coding leaderboard, Grok 4.5 is a strong upper-middle model, full stop.

The number that flips the story: cost

Now look at the price tag. Grok 4.5 lists at $2 per million input tokens and $6 per million output. Opus 4.8 is $5 and $25. GPT-5.5 is $5 and $30. Fable 5, the benchmark leader, is $10 and $50. On output, the token you actually generate the most of when a coding agent is grinding through a task, Grok 4.5 is a quarter the price of Opus and roughly an eighth of Fable's.

Think of two contractors bidding the same renovation. One has a gleaming portfolio and a higher day rate; the other is a notch below on paper but quotes a lower rate and swears he will finish in fewer days. The portfolio is the benchmark. The number that hits your bank account is the day rate times the days. A few points on a leaderboard rarely survives contact with that arithmetic.

And then it uses fewer tokens too

Price per token is only half of a bill. The other half is how many tokens the model burns to finish the job, and this is where xAI leans hardest. It claims Grok 4.5 resolves a SWE-Bench Pro task in an average of 15,954 output tokens, against 67,020 for Opus 4.8, which it frames as using 4.2x fewer tokens. A cheaper token and fewer of them multiply, and that product is the actual cost per resolved task, which is the only cost that matters when you are paying the bill.

There is a real asterisk here, and we will get to it, because the honest version of this number is smaller than the poster version.

Where it came from

Two details explain the shape of this model. First, xAI trained it on tens of thousands of Nvidia GB300 GPUs. For a company whose founder talks up custom silicon at every turn, the frontier is still rented from Nvidia like everyone else's.

Second, Grok 4.5 was tuned on real developer sessions from Cursor, the AI coding editor SpaceX agreed to buy for $60 billion in June. That is the tell for why this model is built the way it is. xAI's whole coding push exists because Grok kept losing developers to Claude Code and Codex, and the fix was to buy the editor those developers already live in and train on what they actually do inside it. A cheap, token-frugal model that plugs straight into Cursor is exactly the weapon you build when your problem is retention, not raw capability.

What it means if you build

The takeaway is not "Grok 4.5 is the best model." It plainly is not. The takeaway is that the benchmark leaderboard and the cost-per-task leaderboard are different rankings, and if you have been picking models off the first one you have been overpaying. For an agent that runs thousands of iterations a day, a model that is a few points behind on evals but a quarter of the output price and materially more token-efficient can be the correct engineering choice by a wide margin.

So run the math for your own workload. Take your real task mix, measure tokens in and tokens out on two or three models, multiply by their list prices, and rank by cost per completed task, not by benchmark score. That number, not a bar on a vendor's slide, is what should decide your default model. Grok 4.5's launch is a good excuse to build that spreadsheet if you never have.

The honest caveats

Keep four things straight before you switch anything. First, every benchmark and efficiency figure here is xAI's own, self-reported, and not yet independently reproduced. Second, that 4.2x token claim compares Grok's average against Opus 4.8's maximum, which is not a like-for-like comparison; the real gap is narrower, so treat 4.2x as a best case, not a spec. Third, it loses to Fable 5 on every coding benchmark shown, so if you are optimizing purely for capability on hard tasks, this is not your model. Fourth, it is not available in the EU yet, with xAI targeting mid-July, and the Cursor-derived training data raises the usual quiet question of how much overlap there is between what it learned on and what it is graded on. No evidence of contamination, but it is the thing to watch.

Key Takeaways

• xAI released Grok 4.5 on July 8, its new flagship, at $2 per million input tokens and $6 per million output, with a 500K context window and roughly 80 tokens per second.
• On coding benchmarks it ties Opus 4.8 (2-2) and clearly trails Fable 5, whose 80.4 on SWE-Bench Pro leads Grok's 64.7. It is a strong mid-pack model, not a record-setter.
• Its edge is cost: output is a quarter of Opus 4.8's price and about an eighth of Fable 5's, and xAI claims it resolves SWE-Bench Pro tasks in far fewer tokens.
• The 4.2x token-efficiency claim compares Grok's average against Opus 4.8's max, so the honest gap is smaller. All numbers are xAI's own and unverified.
• The real lesson for builders: rank models by cost per resolved task, not by leaderboard score. The two rankings disagree, and the invoice is the one you pay.
• Grok 4.5 was trained on Nvidia GB300 GPUs and tuned on Cursor developer sessions, the editor SpaceX is buying for $60B, which explains its cheap, agent-friendly design.

Sources: xAI (Grok 4.5 announcement), TechCrunch (launch and Musk quotes), The Decoder (pricing, GB300 training, cross-model benchmarks), Scale SWE-Bench Pro (benchmark), TechCrunch (SpaceX-Cursor acquisition)