GPT-6.1 Sol vs Grok 4.6: The Accept-Header Problem, and a 35-Point Gap That Is Not Real

There is a pairing in the current model market where the cheaper option wins the hardest science benchmark outright, and the whole comparison still does not go its way. Grok 4.6, SpaceXAI's flagship since 2026-08-12, scores 94.9% on GPQA Diamond — graduate-level science questions — against 42.9% on Humanity's Last Exam for the same model, and it costs $2.00 per million input and $6.00 per million output. GPT-6.1 Sol, Open​AI's 2026-09-29 release, scores 52.9% on that same HLE set and charges $10.00 per million output. So Grok 4.6 is better at one hard thing, cheaper on every line of the rate card, and yet the second half of that comparison is where the interesting problem lives: the two models are not always being asked the same question.

The numbers below are Artificial Analysis measurements carried on our own model records for both models, vendor figures where labeled, and our own serving data. Evaluator and date are named wherever a figure could be ambiguous.

The version number doing the work

Start with the numbers that look like a rout and are not. Artificial Analysis runs two terminal-agent benchmarks, and it has scored both models on each:

• Terminal-Bench 2.1 — Grok 4.6 scores 88.39%; GPT-6.1 Sol is not listed on this version.

• Terminal-Bench 4.0 — Grok 4.6 scores 21.21%; GPT-6.1 Sol scores 56.06%.

Read only the second row and Grok 4.6 looks unable to complete a task in a shell. Read both and it becomes clear that the difference is the harness, not the model. Terminal-Bench 4.0 is a harder revision of the same idea, and the two models were scored on it at different times — GPT-6.1 Sol's Artificial Analysis record carries an evaluation date of 2026-09-28, Grok 4.6's carries 2026-08-12. A 35-point gap between two models measured six weeks apart on a benchmark that was revised inside that window is not evidence that one of them cannot use a terminal. It is evidence that you should not compare those two cells.

The honest version of the agentic comparison is the narrower one: Grok 4.6 posts 88.39% on Terminal-Bench 2.1, and there is no GPT-6.1 Sol figure on that version to put beside it. That is a hole in the evidence, and the rest of this article treats it as one.

What the two are

Grok 4.6 is the current flagship of the Gro​k line, listed as "Latest" in SpaceXAI's own developer documentation, succeeding Grok 4.5 with the same 500,000-token context window and the same text, image and file input surface. It supports configurable reasoning effort, native tool calling, structured outputs, and the full sampling surface — temperature, top_p, seed, logprobs, penalties — which is a meaningful detail: unlike GPT-6.1 Sol, it accepts a temperature.

GPT-6.1 Sol is a single snapshot — gpt-6.1-sol, no dated variant — with a 1,050,000-token context window, 128,000 maximum output, the same text, image and file input surface, and five effort levels defaulting to medium. It does not accept temperature or top_p; sampling is replaced by the effort dial.

The specification rows, side by side, on the dimensions where they differ:

• Context window — 1,050,000 tokens for GPT-6.1 Sol against 500,000 for Grok 4.6. GPT-6.1 Sol is just over twice the window.

• Maximum output — 128,000 tokens for GPT-6.1 Sol; Grok 4.6's output ceiling is not published in our catalogue record for it, which is a field that was never populated rather than a zero.

• Sampling parameters — temperature, top_p, seed and the penalty parameters work on Grok 4.6 and are rejected by GPT-6.1 Sol.

• Base price — $2.00/$6.00 for Grok 4.6 against $2.00/$10.00 for GPT-6.1 Sol, inside both models' first tiers.

Five axes, one clear direction

The two models share enough independent measurement to be compared properly, and the shape of the result is consistent:

• AA Intelligence Index — GPT-6.1 Sol 51.8, rank 10 of 147, 93.2nd percentile; Grok 4.6 44.3, rank 20, 86.4th percentile.

• Humanity's Last Exam — 52.9% against 42.9%. Ten points, GPT-6.1 Sol.

• Long-Context Recall — 83% against 80.33%.

• SciCode — 54.2% against 56.5%. Grok 4.6 by 2.3 points, the one axis of these four where it leads.

• GPQA Diamond — GPT-6.1 Sol is not listed on this axis in our records; Grok 4.6 scores 94.9%, the single strongest number on its card.

Two of Grok 4.6's axes go unmeasured on the GPT-6.1 Sol side, and one of GPT-6.1 Sol's goes unmeasured on Grok 4.6's. That is the standard shape of a comparison between models evaluated on different schedules, and the way to handle it is to leave the cells empty rather than fill them from a launch deck.

One reconciliation is owed on the Intelligence Index. Earlier coverage on this blog placed Grok 4.6 at 61 on the same index. Our current record carries an additional row with that value, alongside the 44.3 figure above — the index has been revised and rescored, and both readings exist for the same model. Anyone quoting Grok 4.6's index position should say which revision they mean, because the difference between 61 and 44.3 is larger than the gap this entire article is about.

The rate card, and the step inside it

Grok 4.6 is cheaper on every base line and has a longer list of lines. Its rate card:

• Up to 200,000 prompt tokens — $2.00 input, $6.00 output, $0.50 cached input.

• Above 200,000 prompt tokens — the whole request reprices to $4.00 input, $12.00 output, $1.00 cached input.

GPT-6.1 Sol's:

• Up to 272,000 prompt tokens — $2.00 input, $10.00 output, $0.10 cached input, $2.50 cache write.

• Above 272,000 prompt tokens — the whole request reprices to $4.00 input, $15.00 output, $0.20 cached input, $5.00 cache write.

Both step, and both steps land in the same region of prompt length — but note what the steps do to the headline gap. Inside the first tier Grok 4.6's output is 40 percent cheaper. Above the step, at $12.00 against $15.00, it is 20 percent cheaper. And on cached input, the ordering reverses completely: GPT-6.1 Sol's $0.10 per million is a fifth of Grok 4.6's $0.50 before the step and a tenth after it.

For an agentic workload that re-reads a long fixed prefix on every turn — a repo, a contract set, a long transcript — that cache row is worth more than the base rate, and it points the other way. For a workload that produces a lot of output from a short prompt, Grok 4.6's $6.00 is the line that decides it.

Our own serving data, and why part of it should be ignored

OrcaRouter's catalogue carries a seven-day serving panel for each model, and on one of these two it is an artifact rather than a measurement. Grok 4.6's panel reads a p50 latency of 6,250 milliseconds and an error rate of 56.76%, over a window in which it served 1,212,441 tokens. GPT-6.1 Sol's panel reads a p50 of 2,442 milliseconds, 95.57 output tokens per second, a 6.36% error rate, and 421,752,431 tokens served in the same seven days.

The traffic numbers are the ones to read first, because they explain the rest. A model that served 1.2 million tokens over a week has too thin a sample for an error rate to mean much — at that volume a 56.76% rate reflects a few dozen attempts, not a service quality, and treating it as evidence that Grok 4.6 is unreliable would be reading a measurement artifact as a measurement. Our GPT-6.1 Sol panel has roughly 350 times the traffic behind it, which makes it a usable number about our playground users and still not a property of the model.

What we cannot tell you

Three things are missing from this comparison, and none of them can be filled in by argument.

The first is a same-version terminal-agent score for both models. Until Artificial Analysis scores Grok 4.6 on Terminal-Bench 4.0, or GPT-6.1 Sol on 2.1, the agentic comparison rests on one model's number and one model's silence. Anyone who tells you the 56.06% against 21.21% spread is a capability gap is quoting a revision difference.

The second is Grok 4.6's output ceiling. The field is unmeasured in our record, so a workload that depends on very long single responses has no published number to plan against on that side.

The third is what both models actually cost per completed task. Our playground reports an average cost per task for Claude Opus 5; it does not report one for either of these two, and the token prices above are inputs to that number rather than the number itself. Open​AI's own vendor-reported figures for GPT-6.1 Sol describe cost-per-task improvements — Terminal-Bench Science 0.1 at $5.47 per task against $23.21 for Claude Opus 5.5 and $23.80 for GPT-6 Astra, at maximum effort — but those are Open​AI's numbers on Open​AI's harness, nobody outside has reproduced them, and they say nothing about Grok 4.6.

The choice

Both models are routed on OrcaRouter at the provider's list price with 0% markup — openai/gpt-6.1-sol at $2.00/$10.00 with $0.10 cached input and the above-272K tier applied exactly as Open​AI sets it, and grok/grok-4.6 at $2.00/$6.00 with the above-200K tier applied exactly as SpaceXAI sets it. One key reaches both, pricing follows the vendor the same day it changes rather than at the next billing cycle, and automatic failover sits underneath so the cheaper model can hold a path without being the only thing holding it.

If the workload is question-shaped and heavy on output — retrieval-and-answer, generation at volume, anything that never grows past 200,000 tokens of prompt — Grok 4.6 is the better buy on the rate card, by 40 percent on the line that matters, and its 94.9% GPQA Diamond is an independent figure, not a claim.

If the workload is agentic, long-prompt, or cache-heavy, GPT-6.1 Sol is the safer default: it leads on every independent axis the two share except SciCode, its cache reads cost a fifth of Grok 4.6's, and its published output ceiling is a number you can plan against. The 500,000-token window on Grok 4.6 is generous and the 1,050,000-token window on GPT-6.1 Sol is twice it — but on a genuinely long prompt both have stepped into their doubled tiers by then, which is where the price gap narrows to 20 percent and stops being the reason to choose either.

The verdict that survives the evidence: Grok 4.6 is the cheaper model and GPT-6.1 Sol is the more measured one, and the 35-point terminal-benchmark spread that makes the rounds is a comparison of two different benchmark versions, not of two models.