Grok 4.5 vs Claude Opus 4.8: Reading Vendor Benchmarks Without the Hype

Every model launch comes with a comparison chart, and the chart is always drawn by the team that shipped the model. xAI's Grok 4.5 launch (July 8, 2026) is no exception: the page puts Grok 4.5 head-to-head against Claude Opus 4.8 across a set of coding benchmarks. What makes this one worth reading is that the scorecard is unusually honest — it shows Grok 4.5 losing two of them.

That's the version of a vendor benchmark you can actually learn from. A clean sweep tells you the vendor picked the fights it could win. A mixed result tells you where the model is genuinely strong and where it isn't. So let's read this one carefully — the first-party numbers, the one independent number that anchors them, and the model that quietly beats both.

As always with a day-old launch: the xAI numbers are first-party and pending external validation. Treat them as a map of where to look, not a verdict.

Key Takeaways

• Grok 4.5 wins 2 of 4 first-party coding benchmarks against Opus 4.8: DeepSWE 1.0 and Terminal-Bench 2.1. It loses DeepSWE 1.1 and SWE-Bench Pro.
• Claude Fable 5 leads all four — the model xAI didn't put at the center of the chart is the one on top.
• The one independent anchor is Artificial Analysis's Intelligence Index: Grok 4.5 at 54, rank #4, behind Fable 5, GPT-5.5, and Opus 4.8.
• First-party benchmarks aren't lies — they're marketing. The skill is reading the axes, the harness, and the model that's missing from the headline.
• For a builder, the takeaway isn't "who won" but "where is each model strong, and what does it cost to get there."

The scorecard, read straight

Here is what xAI's own launch page reports for Grok 4.5 versus Opus 4.8:

Net result: 2–2. Grok 4.5 takes the older DeepSWE 1.0 and Terminal-Bench 2.1; Opus 4.8 takes the newer DeepSWE 1.1 and the well-worn SWE-Bench Pro. On a page the vendor controls, a 2–2 split is a strong signal of relative honesty — nobody publishes a tie they didn't have to.

Notice the pattern, too. Grok 4.5's wins are on a terminal-agent benchmark and an older SWE variant; its losses are on the newer DeepSWE revision and the SWE-Bench Pro battery that most teams treat as the harder, more current bar. That's the kind of texture a sweep would hide. It suggests Grok 4.5 is competitive-to-excellent on some agentic coding shapes and a step behind Opus 4.8 on others.

The model that isn't in the headline

Here's the detail that matters most and gets the least attention: on all four of these benchmarks, Claude Fable 5 leads. xAI benchmarked against Opus 4.8, not against the top model — and the independent data agrees with that ranking. Artificial Analysis's Intelligence Index places Grok 4.5 at 54, rank #4 of ~170, explicitly behind Fable 5, GPT-5.5, and Opus 4.8.

This is the single most useful habit when reading a launch chart: ask which model is missing. When a vendor benchmarks against the second- or third-best model instead of the best, that's a positioning choice, and it's telling you where the real frontier is. It doesn't make Grok 4.5 bad — #4 of ~170 is excellent — but it corrects the "new king" narrative before it forms.

For the wider method here, I keep a reusable checklist in First-Party Benchmarks Are Marketing: A Skeptic's Checklist. The short version follows.

How to read any vendor benchmark

Run every launch chart through these questions before you believe it:

1. Who ran it? First-party (the vendor) or independent (Artificial Analysis, a neutral eval)? First-party sets the burden of proof higher.
2. Which harness? xAI's coding-agent numbers come from its own "Grok Build" harness. Scaffolding, retries, and tool access change scores a lot. Two "SWE-Bench" numbers from different harnesses aren't comparable.
3. Which model is the comparison target — and which is missing? Benchmarking against #3 while #1 sits off-chart is a choice.
4. Which benchmark version? DeepSWE 1.0 vs 1.1 is not a rounding difference. Newer revisions are usually harder and more discriminating.
5. Is the split too clean? A 4–0 sweep on the vendor's page deserves more suspicion than a 2–2. Honesty leaves marks.
6. What did it cost to get the score? A benchmark win at 4× the token spend isn't the same win. (More on that below.)

None of this means "distrust everything." It means calibrate. Grok 4.5's chart passes several of these tests — mixed results, a named harness, a real independent anchor — which is more than a lot of launches can say.

Cost changes what "winning" means

There's a dimension the win/loss table hides: price and token efficiency. Opus 4.8 wins SWE-Bench Pro on accuracy — but xAI reports Grok 4.5 using ~15,954 average output tokens on that same benchmark versus Opus 4.8's ~67,020, roughly 4.2× fewer. And Grok 4.5's API runs $2 in / $6 out per million tokens, which Artificial Analysis pegs at >60% cheaper than Opus 4.8.

So the real comparison isn't "69.2% beats 64.7%." It's "69.2% at premium price and 4× the output tokens" versus "64.7% at a third of the cost." For a one-shot, correctness-critical task, you pay for the extra 4.5 points. For a high-volume agent loop running the same class of task thousands of times a day, the cheaper-and-leaner model can be the correct engineering choice even though it "lost" the benchmark. That trade is the whole subject of The Token-Efficiency Play and Grok 4.5's Price-per-Intelligence for the Agent Budget.

What these four benchmarks don't tell you

Even read perfectly, this scorecard has a narrow field of view. All four benchmarks are coding/agentic tasks — DeepSWE, Terminal-Bench, SWE-Bench Pro. That's a deliberately chosen slice, and it says nothing about the work many skills actually do: summarization, extraction, classification, long-context synthesis, structured output, multilingual tasks. A model that trades blows with Opus 4.8 on SWE-Bench tells you little about how it drafts a research brief or parses a messy document.

There's also a modality boundary the chart doesn't mention. Grok 4.5 takes text and image input but produces text only — no image generation from this model. If your skill needs the model to see a screenshot and reason about it, that's supported; if it needs the model to produce an image, this isn't the model, and no coding benchmark would surface that.

And none of these numbers speak to the things that break agents in production: instruction-following under long context, refusal behavior, tool-call reliability, or output-format stability across thousands of calls. Those don't show up on a launch chart because they're hard to score and unflattering to publish. You find them by running the model on your own workload — which is why the recurring advice in this series is to benchmark your real tasks, not to trust the poster.

What this means for a skill-builder

If you maintain skills or agents on top of these models, the benchmark chart translates into a routing policy, not a loyalty:

• Don't pick a winner — pick per task. Grok 4.5 looks strong on terminal-agent and some SWE shapes; Opus 4.8 leads on the newer/harder SWE variants; Fable 5 tops the index for the genuinely hard 20%. A router that sends each task class to the right model beats standardizing on any one.
• Weight cost into the score. Bake price-per-intelligence into your routing, not just accuracy. Turning that into a real skill is the point of A Model-Router Skill: Grok, Opus, Fable, and GPT.
• Re-run the benchmark that matters — yours. Public benchmarks are proxies. Put your actual tasks through both models and measure accuracy and token spend before committing.
• Adopt cheaply. The SDK-compatible API makes it a low-risk experiment; see Drop Grok 4.5 Into Your Multi-Model Skill Stack in Five Minutes.

You can find routers, eval harnesses, and cost-monitoring building blocks across the skills catalog, agents, and workflows.

The takeaway

The Grok 4.5 vs Opus 4.8 chart is a good benchmark precisely because it isn't a clean win. Grok 4.5 takes two, Opus 4.8 takes two, and Fable 5 — the model that isn't in the headline — leads all four while the independent index puts Grok 4.5 a respectable #4. Read that way, the launch tells you something true and useful: Grok 4.5 is a frontier-adjacent coding model that's competitive on some agentic tasks and a step behind on others, at a fraction of the price.

That's enough to act on. Don't crown it and don't dismiss it — route to it where the task and the economics line up, and verify on your own evals. For the rest of the series, start at the Grok hub.