Grok 4.6 Scores 26% and 88% on the Same Benchmark Line

If you have seen a number for Grok 4.6 in the last month, it was probably one of two: about 88%, or about 26%. Both circulate as Terminal-Bench results. Both are real. They are not the same benchmark, and the difference is a version number that most write-ups drop.

The primary source is xAI’s Grok 4.6 model card, dated August 12, 2026, revision 2026-08-17. It is worth reading directly, because it is more careful than its coverage.

The version problem

The card puts Grok 4.6 at 26.0% on Terminal-Bench 3.0, and explains what 3.0 is:

“Terminal-Bench 3.0 is the successor benchmark to Terminal-Bench 2.1, continuing the same terminal-agency evaluation line with an expanded task set and refreshed harness.”

A footnote adds that the suite “was formerly published under the name FrontierBench.” So a single evaluation line has carried three names and at least two incompatible versions. The ~88% figure in circulation comes from Artificial Analysis measuring 2.1. Quote either number without its version and you have said nothing.

Here is what 3.0 actually looks like, from the card’s own chart:

Model (effort)	Terminal-Bench 3.0
Opus 5 (max)	43.5%
GPT-5.6 Sol (max)	34.6%
Fable 5 (max, with fallback)	34.1%
Grok 4.6 (high)	26.0%
Opus 4.8 (max)	21.1%
Grok 4.5 (high)	15.7%
Sonnet 5 (max)	14.6%

Note the numbers in the left column as much as the right.

The effort-setting problem

Look at the parenthetical after each model name. Peers are reported at max. Grok 4.6 is reported at high. The card states that 4.6 “adds a new xhigh reasoning setting” above what 4.5 offered — so high is not this model’s ceiling.

The same pattern holds across the knowledge-work benchmarks:

Benchmark	Opus 5	Grok 4.6	Effort compared
SWE-Marathon v1.1	50.0%	31.9%	max vs high
AA GDPVal (Elo)	1849	1753	max vs high
AA-Briefcase (Elo)	1715	1577	max vs high
APEX-Agents	60.6%	57.5%	max vs high

Then look at the one benchmark where Grok 4.6 comes first — CursorBench 3.2 — and the effort setting changes: it “scores 70.8% at xhigh thinking effort, exceeding the other models tested, and 69.9% at high.”

There is a second footnote worth catching: on CursorBench, “Grok 4.5 was served with a maximum thinking effort of high.” The generational improvement from 4.5 to 4.6 on that chart is therefore partly a comparison between a model capped at high and one running at xhigh.

The Cursor problem

The card’s opening sentence discloses a relationship that shapes the whole coding section:

“Grok 4.6 is the latest release in SpaceXAI’s 1.5T-scale model family, developed in collaboration with Cursor.”

The attached footnote is more specific: “Grok 4.6 received supplemental training on anonymized Cursor workflow data to improve coding and agentic performance.”

The headline coding benchmark is CursorBench 3.2, which evaluates “realistic IDE-style tasks from production-like Cursor workflows.” And its footnote: “Results reported are taken from evaluations conducted by Cursor.”

So the model was trained on Cursor workflow data, evaluated on a benchmark built from Cursor workflows, by Cursor. None of that is hidden — all three facts are printed on the same pages — and none of it makes the result fake. It does make CursorBench a poor choice for the one chart you generalise from, and a good predictor of exactly one thing: how the model behaves inside Cursor.

To xAI’s credit, the card names an independent evaluator for nearly every other chart: Harbor for Terminal-Bench, Abundant AI for SWE-Marathon, Artificial Analysis for GDPVal and Briefcase, Mercor for APEX-Agents. That disclosure is better practice than most model cards manage.

While we are here: the parameter claims

Search results for Grok will hand you “6 trillion” or “10 trillion parameters.” The card says Grok 4.6 belongs to a 1.5T-scale model family. The trillions belong to Grok 5, which xAI confirmed was in training in January 2026 and has not released; the 6T and 10T figures come from roadmap talk, not a model card. Any current article attaching them to a model you can call today is describing something that does not exist yet.

Two more facts from the card that matter more than parameter counts for practical use: Grok 4.6 has a pretraining data cutoff of January 2026, with supplemental training data as late as June 2026, and it holds a 500,000-token context window.

What to actually do with this

The general lesson is not about xAI. Benchmark lines get renamed and re-versioned — this one went FrontierBench → Terminal-Bench 2.1 → Terminal-Bench 3.0 — and scores across versions are unrelated numbers that share a label. Vendors increasingly report at a non-maximal effort setting, which is defensible on cost grounds and invisible once the number is copied into a blog post.

So when a model number reaches you, three questions decide whether it means anything:

1. Which version of the benchmark? A major version bump makes the old number incomparable, not merely stale.
2. At what effort setting, and what were the comparisons run at? A high-versus-max table is measuring two different things.
3. Who ran the evaluation, and what is their relationship to the model? Self-run and partner-run results are still useful; they are just not independent.

xAI’s card answers all three, in footnotes, on the page. The failure is downstream, in every summary that keeps the number and drops the sentence under it.