GPT-6.1 Sol vs Claude Opus 5: One Index Point Apart, Two and a Half Times the Price

One point on a hundred-point scale separates them, and one of them costs two and a half times as much. On the Artificial Analysis Intelligence Index, GPT-6.1 Sol — Open​AI's release of 2026-09-29 — sits at 51.8. Claude Opus 5, Anthrop​ic's flagship since 2026-07-24, sits at 50.8. The rate cards are $2.00 and $10.00 per million tokens for GPT-6.1 Sol against $5.00 and $25.00 for Claude Opus 5. That is not a comparison between a strong model and a cheap one. It is two models an independent evaluator could not separate by more than a point, sold at prices that differ by a factor of 2.5 — and the interesting question is not which one is better, but which of the two the index is failing to distinguish.

Every number below is either an Artificial Analysis measurement carried on our model records, a figure the vendor published itself, or arithmetic on the two rate cards. Where a figure is the vendor's own and nobody outside has reproduced it, the sentence says so.

What the two actually are

Claude Opus 5 is Anthrop​ic's top-of-line generalist: a 1,000,000-token context window, 128,000 output tokens, text, image and file input, and adaptive reasoning where callers set an effort level rather than a temperature. It is the model the company positions for end-to-end software engineering, code review, and long multi-step agentic sessions with heavy tool use.

GPT-6.1 Sol is one tier below Open​AI's own flagship — GPT-6 Astra — in the GP​T-6 line, and it was shipped six weeks after the model it replaced, GPT-6 Sol, at the same $2.00/$10.00 price. It carries a 1,050,000-token context window, the same 128,000 output ceiling, the same text, image and file input surface, and five reasoning effort levels from low to max with medium as the default.

Read those two paragraphs side by side and the specification rows are close to identical. The differences are in the fine print, and three of them matter:

• Sampling parameters — neither model accepts temperature or top_p. Both replace them with an effort dial. If your code passes a temperature, both will reject it, so neither is a drop-in for a legacy integration.

• Context ceiling — 1,050,000 tokens for GPT-6.1 Sol against 1,000,000 for Claude Opus 5. A five percent difference; irrelevant in practice, and not where this decision lives.

• Effort defaults — GPT-6.1 Sol defaults to medium; Claude Opus 5's adaptive thinking allocates depth on its own. Everything about the index comparison below is sensitive to this, which is the caveat that comes later.

The five axes both were measured on

This is the part most write-ups of this pairing get wrong, because they mix a vendor number on one side with a third-party number on the other. The five rows below are all Artificial Analysis, all from the same evaluator, for both models:

• AA Intelligence Index — GPT-6.1 Sol 51.8 (rank 10, 93.2nd percentile) against Claude Opus 5 50.8 (rank 11, 92.5th percentile). One point.

• Humanity's Last Exam — 52.9% against 54.9%. Claude Opus 5 by two points.

• SciCode — 54.2% against 56.4%. Claude Opus 5 by 2.2 points.

• Long-Context Recall — 83% against 79.33%. GPT-6.1 Sol by 3.7 points.

• Terminal-Bench 4.0 — 56.06% against 48.99%. GPT-6.1 Sol by seven points.

Three to two, on margins between one and seven points, at 2.5 times the price. There is a sixth axis Claude Opus 5 holds alone — GPQA Diamond at 93.2% — and a seventh it holds alone the other way: Artificial Analysis scores Claude Opus 5 at 78 on its Coding Index, 98.6th percentile, rank 2. GPT-6.1 Sol is not listed on that axis at all in our records, so there is no number to put beside it, and inventing one from a vendor deck is exactly the move this blog does not make.

Where the premium is still defensible

A one-point index gap would be a thin reason to pay 2.5 times as much if that were the whole picture. It is not, and there are three places where Claude Opus 5 is the better purchase regardless of what the index says.

The first is the Coding Index row above. Rank 2 of 138 models on the evaluator's own composite, at the 98.6th percentile, is a concrete claim about a specific kind of work, and it is the kind of work most teams buying at this price point are doing. Our OrcaRouter model page for anthropic/claude-opus-5 reads "Ranking #1" against the Codex benchmark and a $1.09 average cost per task — our own serving data, not a vendor claim, and a reminder that the per-token price is not the per-job price.

The second is batch. Claude Opus 5 carries a 50 percent batch discount, which takes it to $2.50 input and $12.50 output, and it permits 300,000 output tokens in batch mode against 128,000 on the interactive path. At $2.50/$12.50 it is no longer two and a half times GPT-6.1 Sol's price on input — it is 25 percent more. For any workload that tolerates a queue — offline indexing, bulk classification, overnight test generation — the entire argument of this article gets much smaller, and it gets smaller in Anthrop​ic's favour.

The third is the modest multimodal multiplier. Anthrop​ic's own published figures put Claude Opus 5 at 70.6% on OSWorld 2.0 and 96.0% on SWE-bench Verified, with 79.2% on SWE-bench Pro. Those are vendor-reported numbers on a model announced in July, nobody outside has reproduced them here, and the axes do not overlap with the Artificial Analysis rows above — so they cannot be added to that scorecard. They are worth reading only as an indication that Anthrop​ic is claiming strength on computer-use and repository-scale engineering specifically, which is the buyer this model is priced for.

The rate cards, including the row only one of them has

Claude Opus 5 charges $5.00 and $25.00 flat across its whole million-token window, with cache reads at $0.50 and cache writes at $10.00. GPT-6.1 Sol charges $2.00 and $10.00 up to 272,000 prompt tokens, then reprices the entire request at $4.00 and $15.00 above that line, with cache reads moving from $0.10 to $0.20 and cache writes from $2.50 to $5.00. The inversion that produces is the practical consequence of buying long context from either model:

• A 250,000-token prompt — GPT-6.1 Sol is still in its first tier, at $2.00/$10.00 against $5.00/$25.00. This is the cheapest large-context request in the comparison.

• A 300,000-token prompt — GPT-6.1 Sol has stepped to $4.00/$15.00, and Claude Opus 5 has not moved. The gap collapses from 2.5× to 1.25× on input and 1.67× on output.

• Cache reads — $0.10 per million for GPT-6.1 Sol against $0.50 for Claude Opus 5, inside both models' first tiers. On a long agentic session that re-reads the same prefix, this row does more for the bill than the headline rates do.

• Cache writes — $2.50 against $10.00. Same direction, same order of magnitude.

That is a real and unglamorous point: the 2.5× headline is only reached by short prompts. The moment a workload crosses 272,000 tokens, GPT-6.1 Sol is priced like a premium model too, and the honest comparison becomes "a premium model with a much cheaper cache" against "a premium model with a flat rate card."

Why one point should not decide this on its own

The index rows above were not collected on the same day, and that matters more here than in most comparisons. GPT-6.1 Sol's Artificial Analysis record carries an evaluation date of 2026-09-28. Claude Opus 5's carries 2026-07-23 — nine weeks earlier, on a harness that has been revised across that window.

There is a visible symptom of that in our own records. Earlier coverage of Claude Opus 5 put it at 61 on the Intelligence Index and at the top of the board. The catalogue record we read today puts the same model at 50.8, rank 11. Both can be true: the index has been revised, and a composite score is sensitive to the reasoning effort the evaluator selects — Claude Opus 5 topped out at maximum effort, and the record we carry is not annotated with an effort setting at all. GPT-6.1 Sol's 51.8 carries the same ambiguity.

What that means for a buying decision is narrow but important. A one-point gap between two models evaluated nine weeks apart, on a moving composite, at unspecified effort settings, does not establish that one model is smarter than the other. What the five rows do establish is that they are in the same band — and a same-band result at 2.5 times the price is a procurement fact, not a performance fact.

Which one to route

Take the model you were going to buy and test it against the other one. That sounds like a platitude, but it is the only move the evidence actually supports here, and both models are reachable from a single Open​AI-compatible endpoint — openai/gpt-6.1-sol and anthropic/claude-opus-5 on OrcaRouter — so the test is a model-string change, not a second contract and not a rewrite. Pricing is the provider's list rate with 0% markup: $2.00/$10.00 with the above-272K tier applied exactly as Open​AI sets it, and $5.00/$25.00 with the batch discount applied exactly as Anthrop​ic sets it, which also means a vendor price change is live on our side the same day rather than at the next billing cycle. Where both are routed, automatic failover sits underneath, which is worth having when the cheaper of the two is the one holding a production path.

If you want a default before you have your own numbers, the split the evidence supports is this. Choose GPT-6.1 Sol when the workload is long-prompt, cache-heavy, or terminal-and-tool shaped — the Long-Context Recall and Terminal-Bench 4.0 rows are the two widest margins in the comparison, and $0.10 cache reads are a large lever on agentic sessions that re-read a prefix. Choose Claude Opus 5 when the workload is repository-scale coding, when batch is acceptable, or when the prompt fits inside 128,000 tokens and the per-task cost our own playground measures at $1.09 is the number your finance team already approved.

The verdict this pairing earns is narrower than either vendor would like. On independent measurement these two are effectively tied; on price they are not, and price is the axis one of them cannot argue its way out of.