
GPT-6.2 Sol's Rework Bill: Cheap Per Token, Costly Per Fix
- OrcaNEWOrcaCyber: Hy4 Flash Abliterated / Uncensored2026-11-18$0.07 / $0.28 per 1M tokens · 185 tok/s
- tencentNEWTencent: Hy4 Flash2026-11-16$0.02 / $0.07 per 1M tokens · 826 tok/s
- AlibabaNEWQwen: Qwen4 Max2026-11-16$1.50 / $4.50 per 1M tokens · 407 tok/s
- GoogleNEWGoogle: Gemini 4 Carbon2026-11-10$2.00 / $10.00 per 1M tokens · 87 tok/s
- AnthropicNEWAnthropic: Claude Fable 5.22026-11-09$8.00 / $40.00 per 1M tokens · 150 tok/s
- OrcaOrcaCyber: GLM 5.3 Abliterated / Uncensored2026-11-04$3.00 / $5.00 per 1M tokens · 99 tok/s
- OrcaOrcaCyber: Hy4 Abliterated / Uncensored2026-11-04$0.80 / $1.60 per 1M tokens · 134 tok/s
- OrcaOrcaCyber: Kimi K3.1 Pro Abliterated / Uncensored2026-11-04$2.00 / $5.00 per 1M tokens · 83 tok/s
- tencentTencent: Hy42026-11-04$0.83 / $2.50 per 1M tokens · 120 tok/s
- OrcaOrcaCyber: MiniMax M3.1 Flash Abliterated / Uncensored2026-10-29$0.80 / $1.60 per 1M tokens · 108 tok/s
- OrcaOrcaCyber: GLM 5.3 Flash Abliterated / Uncensored2026-10-27$0.80 / $1.60 per 1M tokens · 67 tok/s
- OrcaOrcaCyber: Qwen3.8 Flash Next Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 90 tok/s
- OrcaOrcaCyber: MiMo V2.6 Flash Abliterated / Uncensored2026-10-26$0.80 / $1.60 per 1M tokens · 41 tok/s
- MoonshotMoonshotAI: Kimi K3.1 Pro2026-10-2447Intelligence
- OrcaOrcaCyber: DeepSeek V4.1 Flash Abliterated / Uncensored2026-10-23$0.80 / $1.60 per 1M tokens · 212 tok/s
- minimaxMiniMax: M3.1 Flash2026-10-1929Intelligence
- OrcaOrca: OrcaCyber Ultra 1.02026-10-17$5.00 / $10.00 per 1M tokens · 111 tok/s
- deepseekDeepSeek: DeepSeek V4.1 Flash 10172026-10-1733Intelligence
- DeepSeekDeepSeek: DeepSeek V4.1 Flash 09102026-10-17$0.15 / $0.60 per 1M tokens · 160 tok/s
- DeepSeekDeepSeek: DeepSeek V4.1 Flash2026-10-1740Intelligence
On the morning of 20 November, the Chinese-language AI commentator 宝玉 (@dotey) posted a verdict on GPT-6.2 Sol that anyone who has spent a week with the model will recognise. Simple tasks are fine. Complex coding tasks are not just a little worse than Claude Opus 5.5 — the work Sol finishes still needs Opus to clean up after it, so you would have been better off handing the job to Opus from the start. The post is a usage report and nothing more: no harness, no task list, no diff, 124 likes at the time of writing and a comment thread arguing about it. It is also checkable in a way most launch-week impressions are not, because the thing it describes has a shape. A model that is fine on easy work and falls apart on hard work should show a specific pattern in the evaluations — strength on execution, weakness on comprehension — and the published numbers either have that pattern or they do not.
They largely do. OpenAI shipped GPT-6.2 Sol on 17 November 2026 at $2 per million input tokens and $10 per million output. On Artificial Analysis's independent Coding Agent Index it lands below the model it replaces, and the component that drags it down is the one where the model has to read a codebase and say something true about it. That is the mechanism behind "it needs cleaning up after", stated as a number. What the tweet does not do, and what this piece tries to, is price the cleanup — because when you actually run the arithmetic, the rework thesis turns out to be much stronger on wall-clock time than on money, and the two lead to different decisions.
What the post actually claims
Translated plainly: "Today I stress-tested GPT-6.2 Sol. Simple tasks are all right. On slightly more complex coding tasks it is not just a little worse than Opus 5.5 — the work it finishes still needs Opus 5.5 to wipe up after it, so you'd be better off letting Opus 5.5 do it from the start."
Three separable claims sit inside that. First, a quality claim: Sol is materially worse than Opus 5.5 on complex coding. Second, a workflow claim: the gap shows up as rework rather than as an obvious failure — you get output, it looks finished, and a stronger model has to repair it. Third, an economic claim: the rework wipes out whatever the token price saved.
The first two are testable against published evaluations. The third is arithmetic, and it depends on a number nobody has published — how much a repair pass costs relative to doing the task once with the expensive model. Everything below keeps those three claims separate, because they do not all survive equally.
The clean comparison runs inside one harness
One warning before any numbers, because it is the most common way this comparison gets made dishonestly. Artificial Analysis measures coding agents as a pairing of model and harness: Codex for OpenAI's models, Claude Code for Anthropic's. Cost per task, token spend and wall time all move with the harness as much as with the model, so a raw "Sol costs $1.22, Opus costs $12.61" comparison is measuring two different machines.
What survives that objection is any comparison within a harness. Inside Codex, GPT-6.2 Sol can be set against GPT-6.1 Sol and GPT-6.2 Astra directly, on identical tasks, with the same tooling and the same grading — and that is where the regression lives. The cross-harness pair is still worth printing, clearly labelled, because it is the comparison a reader is actually making when they decide between an OpenAI subscription and an Anthropic one.
• Codex, GPT-6.2 Sol (max) — Coding Agent Index 48.6, $1.22 per task, 4.25M tokens
• Codex, GPT-6.1 Sol (max) — index 52.2, $1.51 per task, 4.33M tokens
• Codex, GPT-6.2 Astra (max) — index 53.5, $5.12 per task, 4.67M tokens
• Claude Code, Claude Opus 5.5 (max) — index 58.2, $12.61 per task, 15.89M tokens
Read the first two lines together. The new model scores 3.6 points below the one it replaces, on the same harness, at a token spend within 2% of it. The price did not change either — Sol has held $2 and $10 since 29 September, and GPT-6.2 Sol kept the rate card exactly. So this is not a cheaper model trading quality for price. It is the same price, the same harness, and less capability on the thing the index measures. @dotey's comparison is to a different vendor's model, but the regression he is feeling is visible without leaving OpenAI's own product line.

Where the complex-task gap is real
The Coding Agent Index is a composite of three evaluations, equally weighted: FrontierCode-AA 2.0 (132 software engineering tasks, written by Cognition and implemented by Artificial Analysis), Terminal-Bench 4.1 (66 agentic terminal tasks, from the Laude Institute), and SWE-Atlas-QnA-AA (124 technical Q&A tasks over a codebase, from Scale AI with grading adjusted by Artificial Analysis). Each score is the mean of three attempts per task.
Split Sol against its predecessor by component and the shape @dotey describes appears immediately:
• SWE-Atlas-QnA-AA — GPT-6.2 Sol 46.2% vs GPT-6.1 Sol 59.9%, a 13.7-point fall inside one harness
• FrontierCode-AA 2.0 — 41.5% vs 42.5%, essentially flat
• Terminal-Bench 4.1 — 58.1% vs 54.0%, a 4.1-point gain
So the model got better at doing and worse at knowing. Terminal work is execute-and-observe: run the command, read the output, adjust. Codebase Q&A is retrieve-and-report: find the relevant code, understand why it is the way it is, and state it accurately. A model that improves on the first and regresses 13.7 points on the second will pass a smoke test and produce edits that look right and miss the reason the code was written that way. That is precisely the failure mode that generates rework rather than an error message, and it is why the complaint arrives as "it needs cleaning up" rather than "it doesn't work".
The absolute gap to Opus 5.5 on the same evaluation is larger still — 46.2% against 65.9% — but that one crosses harnesses, so treat it as directional. The within-Codex 13.7-point fall is not directional. It is the same harness, the same 124 tasks, the same grader, and the model simply answers questions about a codebase less accurately than the model it replaced.

Where Sol is genuinely fine
The tweet's first clause deserves the same treatment as its second. "Simple tasks are all right" is not a concession; on the composite Intelligence Index, which runs every model through one identical suite rather than through vendor harnesses, GPT-6.2 Sol at max effort beats Claude Opus 5.5 at max effort on several components outright:
• Terminal-Bench-Science 0.1 — 67.6% vs 59.0%, an 8.6-point win for Sol
• AutomationBench-AA — 71.3% vs 69.5%
• Terminal-Bench 4.0 — 60.1% vs 59.6%
• GDP.pdf all-pass — 30.0% vs 26.2%
And it loses on the ones that require holding a fact and reporting it:
• HLE-Diamond — 64.0% vs 67.9%
• AnalystAgent — 53.8% vs 56.3%
• AA-Omniscience — 40.4 at 61.6% accuracy vs 48.1 at 65.5% accuracy
• FrontierCode-AA 2.0 (Index version) — 44.4% vs 46.4%
The Omniscience row is the one worth sitting with. It is the evaluation that measures whether a model knows what it does not know: Sol answers correctly 61.6% of the time and its hallucination rate is 55.4%, against 65.5% and 50.3% for Opus 5.5. A model that answers correctly nearly four points less often and invents something five points more often is a model whose output a human has to check — which is the rework claim, arriving from a second direction.
Net, Sol scores 57.3 against 59.1 on the composite, and does it in 9.6 minutes per task rather than 25.7, at $1.01 rather than $10.04, while emitting 60.9 million output tokens across the suite against Opus's 218 million. Whatever else is true, the cheap model is not slow and it is not verbose.
The effort dial has already been swept
The obvious objection from anyone who has used an OpenAI reasoning model is that this is the wrong setting. Turn the effort down and the model stops over-thinking; turn it up and it stops cutting corners. Artificial Analysis has already run the full sweep on the Coding Agent Index inside Codex, and it does not rescue the situation:
• low — 37.2
• medium — 45.4
• high — 47.5
• xhigh — 49.8
• max — 48.6
The peak is at xhigh, not at max: pushing to maximum effort actually costs 1.2 points, which is the kind of non-monotonic result that shows up when a model starts second-guessing itself. But even the best setting, 49.8, sits 2.4 points below where GPT-6.1 Sol scored at max and 8.4 below Claude Opus 5.5. There is no dial position at which GPT-6.2 Sol becomes the stronger coding agent; there is only a choice about how much you want to pay for 49.8.
The cost of that choice is worth knowing, because it is steep in relative terms even if it is cheap in absolute ones. On the Intelligence Index, Sol runs $0.18 a task at low effort, $0.41 at medium, $0.56 at high, $0.65 at xhigh and $1.01 at max. Moving from medium to xhigh roughly doubles quality per task — 52.7 to 55.7 on that index — and costs 58% more. The default is medium, which is the setting that produces the 45.4 coding score. If you are evaluating Sol against Opus, do it at xhigh or max, or you are comparing the wrong configuration.
One number that predicts rework
There is a second data series in the same run that speaks directly to the "it needs cleaning up" experience, and it is not a quality score. Artificial Analysis records safety refusals per evaluation, and distinguishes refusals the harness recovered from ones that ended the attempt. On SWE-Atlas-QnA-AA, over 372 attempts:
• GPT-6.2 Sol (max) — 25 refusals, every one a hard stop, 6.7% of attempts
• GPT-6.2 Sol (xhigh) — 22 refusals, 5.9%
• GPT-6.2 Sol (medium) — 13 refusals, 3.5%
• GPT-6.1 Sol (max) — 15 refusals, 4.0%
• GPT-6.2 Astra (max) — 8 refusals, 2.2%
• Claude Opus 5.5 — 52 refusals, none of them hard stops, 14.0%
The Opus line is the interesting one and it cuts against a lazy reading. Opus 5.5 refuses more often than any model in this table — and the harness recovered all 52, so none of them cost the run. Sol refuses less often and the harness could not recover a single one of its 25. A refusal that ends an agent's turn is a task a human now has to pick up; a refusal that gets recovered is a retry. On the evaluation that most resembles "answer questions about this repository", the model the tweet is complaining about fails in the mode that produces unfinished work rather than wrong work.
Read alongside the 13.7-point comprehension fall, this is a coherent story rather than two unrelated numbers: Sol is less reliable at the retrieve-and-report part of coding, and when it declines, it declines in a way that leaves the task open.
Pricing the cleanup pass
Now the third claim, the economic one, which is where intuition most often beats arithmetic. Use the Intelligence Index cost per task, because that suite is the same harness for every model on it and so the numbers are directly comparable: GPT-6.2 Sol $1.01, Claude Opus 5.5 $10.04.

Take a task Sol completes and Opus then repairs. Let p be the share of tasks that need repair and r be the cost of a repair as a fraction of a full Opus pass. Sol-then-repair costs $1.01 + p × r × $10.04. Opus alone costs $10.04. Sol-then-repair is cheaper whenever p × r is below 0.899.
That is a high bar to clear. If a repair costs half a full pass, you can repair every single task and still be 78% cheaper than going straight to Opus. Only when repairs approach the cost of a full redo on nearly every task does the bill invert: at p = 1 and r = 1, you pay $11.05 for what Opus would have done for $10.04. On a thousand tasks, that is $11,050 against $10,040 — a 10% penalty, not the rout the tweet implies.
So on money alone, the rework complaint does not overturn the price case unless the rework is close to total. But money is not the constraint most teams hit. Time is:
• Time per Intelligence Index task — GPT-6.2 Sol 9.6 min vs Claude Opus 5.5 25.7 min
• Time when a task needs a full Opus redo — 9.6 + 25.7 = 35.3 min against 25.7 min
• Token spend on the same suite — 60.9M vs 218M output tokens
Here the rework thesis holds with no caveats. Any task Sol attempts and then hands back is a task that took 37% longer than it would have if you had gone to Opus first, and you cannot buy that time back with a cheaper token. If your constraint is a deadline rather than a budget — which is the situation almost every developer is actually in when they post this complaint — @dotey is right, and the argument is about throughput, not price.
What would change the verdict
Two things, and neither is available yet. The first is a published rework rate: what fraction of Sol's complex-task diffs a stronger model actually has to modify, measured on a fixed task set. Nobody publishes this, because it requires running two models on the same tickets and diffing the results, and it is the single number that would settle the argument. The second is a GPT-6.2 Sol figure on an Anthropic-harness coding run, or Opus 5.5 on Codex, which would turn the cross-harness comparison above into a real head-to-head.
Against that, note what is already settled. The regression is real and it is inside one harness. The effort sweep does not fix it. The comprehension evaluation — the one that predicts rework — moved the wrong way by 13.7 points. A reader who wants a reason to keep complex coding on Opus 5.5 has one that does not depend on anyone's impression of a launch week.
Running the test on your own tasks
The rework rate is measurable on your own backlog in an afternoon, and it is the only version of this number that matters to you. Take twenty tasks your team actually shipped last month, run each through both models at the same effort setting, and diff the resulting patches against what a human eventually merged. The fraction that needed substantive rework is your p; the token cost of the repair pass is your r.
OrcaRouter makes that a routing decision rather than a procurement one: anthropic/claude-opus-5.5 and openai/gpt-6.1-sol are both behind a single key, at the vendor's own list price with 0% markup passed through, so the numbers above are the numbers you will be billed. That matters more than usual here, because the model the tweet is defending against — GPT-6.2 Sol — is not in our catalogue yet; we do not serve it, and no amount of routing changes that. What we do serve is the model Sol replaced, which is the right control for an experiment: run 6.2 Sol wherever you can reach it, run GPT-6.1 Sol on the same tasks through us, and the 13.7-point comprehension fall becomes something you can confirm or refute on your own code rather than on a synthetic suite.
For the deployment shape the data suggests, the routing DSL is the honest answer: send planning and codebase-comprehension steps to Opus 5.5, send execution and terminal loops to the cheaper model, and let automatic failover cover the case where a provider stalls mid-run. That is a configuration, not an architecture — and it is what "use Opus to clean up after Sol" looks like when you write it down as a pipeline instead of as a complaint.
The one thing worth taking from the tweet regardless of which model you pick: the bill for a coding task is not the token count. It is the tokens plus the human minutes plus the elapsed wall clock between "done" and "actually done". Sol wins the first line of that ledger decisively, Opus 5.5 wins the second and third whenever a task comes back, and no vendor publishes the rate at which tasks come back. Until someone does, the only trustworthy number is the one you measure on your own diffs.
Compared in this article1
Detected from this article · Benchmarks: Artificial Analysis · updated daily
