Sonnet 5.5 vs GPT-6.1 Sol is a same-price matchup with different bills: both list at $2 input and $10 output per million tokens, but at max effort Artificial Analysis measured $7.67 per task for Sonnet (index 56) and $0.72 for Sol (index 52). Compared at similar scores, the gap narrows to roughly 3.8 times. Sol is cheaper per task; Sonnet answers faster at high effort and reaches higher scores.
Anthropic’s launch page for Sonnet 5.5 sets it against GPT-6 Sol. A day later OpenAI replaced that model with GPT-6.1 Sol, so the vendor comparison already describes a rival that no longer exists. This article uses independent measurements of the current pair. It also closes a gap in our GPT-6.1 Sol review, which listed Sonnet’s cost per task as unknown.
The ledger: what the sticker price leaves out
| Claude Sonnet 5.5 | GPT-6.1 Sol | |
|---|---|---|
| Released | September 28, 2026 | September 29, 2026 |
| Input / output per 1M tokens | $2 / $10 | $2 / $10 |
| Cached input read per 1M | $0.20 | $0.10 |
| Context window | About 1M tokens | About 1M tokens |
| Surcharge on very long prompts | None listed by trackers | Trackers report 2x input and 1.5x output on the whole request above 272K input tokens |
| Default effort | High on the API; Medium in Claude apps and Claude Code | Medium, per one tracker’s reading of OpenAI’s docs |
| Where you get it | Claude API, AWS, Google Cloud, Azure; zero data retention available | API, ChatGPT Work and Codex (see our review for the chat-app caveat) |
Anthropic’s own launch page says Sonnet 5.5 costs up to 30 percent less per task. That compares it with Sonnet 5, not with OpenAI, and the saving comes from fewer tokens, not a lower token price.
Cost per task by effort: where the 10.7x comes from
Artificial Analysis (AA) runs every model through the same ten-test Intelligence Index, version 4.3.2, and reports what a task costs on average. The max-effort and high-effort rows below are from AA’s model pages for Sonnet 5.5 and GPT-6.1 Sol, which we read on October 7. The two xhigh rows come from a tracker’s copy of AA’s pages and we have not checked them ourselves.
| Model and effort | Index score | Cost per task | Output tokens, whole index | Output speed | Time to first answer token |
|---|---|---|---|---|---|
| Sonnet 5.5, max | 56 | $7.67 | 420M | 127 tok/s | 442 s |
| Sonnet 5.5, xhigh (tracker) | 51.9 | $2.75 | 100M | 99 tok/s | 36 s |
| Sonnet 5.5, high | 47 | $1.12 | 52M | 107 tok/s | 12 s |
| GPT-6.1 Sol, max | 52 | $0.72 | 67M | 59 tok/s | 282 s |
| GPT-6.1 Sol, xhigh (tracker) | 51.0 | $0.39 | 36M | 54 tok/s | 161 s |
| GPT-6.1 Sol, high | 50 | $0.32 | 25M | 53 tok/s | 68 s |
AA’s figures drift from day to day. The tracker’s copy, taken two days earlier, shows Sonnet’s max run at 410M tokens and a slightly longer wait, so treat every number here as dated October 7.
At max effort, Sonnet produced about 6.3 times as many output tokens as Sol across the same tasks (420M against 67M). The per-token price is identical, so the bill follows the token count: $7.67 divided by $0.72 is about 10.7 times. That ratio is real, but it compares the highest setting of one model with the highest setting of the other.
Compare at matched scores, not matched labels
An effort label means different things to each vendor. Matching by score is fairer:
- Around 52 points: Sol at max costs $0.72. Sonnet needs xhigh to reach about the same score and costs $2.75, roughly 3.8 times more.
- At the “high” label: Sol scores 50 for $0.32, and Sonnet scores 47 for $1.12. Sol is both higher and about 3.5 times cheaper.
- The last four points: Only Sonnet has a setting that reaches 56. Moving from xhigh to max adds about 4.1 index points for $4.92 more per task, roughly $1.20 per point.
More effort doesn’t always help, either. Anthropic’s launch footnotes show Sonnet 5.5 scoring 46.2 percent on FrontierCode at max effort but 52.1 percent at xhigh, because at max it ran extra review steps that made out-of-scope edits. Run your own sweep before you default to the top setting. Our guide to reasoning token pricing explains why hidden thinking is billed at the output rate.
Speed: the trade runs the other way
Sol is cheaper, but you wait longer for it. At high effort, AA measured about 12 seconds to Sonnet’s first answer token against about 68 seconds for Sol, and Sonnet streamed roughly twice as fast. At max, both models think for minutes before answering (about 7.4 minutes for Sonnet, about 4.7 for Sol). If a person is waiting in a chat window, Sonnet at high or medium feels faster. If a script runs overnight, nobody waits and the per-task bill matters more.
Who wins which test
Sonnet’s GDPval-AA and AA-Briefcase scores (1844 and 1811 Elo) come from Anthropic’s page. The other rows below are a tracker’s reading of AA’s comparison page, max effort on both sides.
| Test (AA harness) | Sonnet 5.5 | GPT-6.1 Sol | Leader |
|---|---|---|---|
| Intelligence Index v4.3.2 | 56 | 52 | Sonnet |
| Terminal-Bench 4.0 (terminal coding) | 64% | 56% | Sonnet |
| GDPval-AA v2.1 (work tasks, Elo) | 1844 | 1575 | Sonnet |
| AA-Briefcase v1.1 (knowledge work, Elo) | 1811 | 1564 | Sonnet |
| AutomationBench-AA | 71% | 65% | Sonnet |
| SciCode | 61% | 54% | Sonnet |
| GDP.pdf (document reasoning) | 26% | 31% | Sol |
| AA-Omniscience Index (knowledge reliability) | 32 | 42 | Sol |
| AA-LCR v1.1 (long context) | 83% | 83% | Tie |
The pattern: Sonnet leads on agentic and office-style work, and Sol is more reliable when it has to answer from knowledge alone.
Why the scoreboards disagree
- Different composites, different winners. AA puts Sonnet ahead, 56 to 52, but BenchLeader’s index puts Sol slightly ahead, 69.4 to 68.3. Each index weights its tests differently. The same trap shows up in our model rankings piece.
- Vendor numbers run higher. Anthropic reports 70.6 percent for Sonnet 5.5 on Terminal-Bench 4.0, while AA measured 64 percent on its own harness. OpenAI doesn’t publish a Terminal-Bench 4.0 score for Sol at all.
- Different test versions. The two vendors cite different versions of the computer-use test OSWorld (2.1 and 2.0), so those scores can’t be compared.
- Launch-week bugs on both sides. Anthropic’s footnotes say AA scored two tests on a pre-release Sonnet deployment with a structured-output bug, which likely understated Sonnet. They also say OpenAI fixed an image-understanding bug in GPT-6 Sol, so older Sol scores may be stale.
- Rankings shift. AA listed Sonnet at #2 of 225 when we looked; a tracker two days earlier saw #3 of 224. Quote a score with its date.
Long prompts and cache reads: the list-price arithmetic
This is our own arithmetic from the list prices above, assuming Sol’s surcharge works as the trackers describe. It holds output equal, which the table above shows is the wrong assumption at high effort, so read it as a lower bound on the difference.
| One agent turn | Sonnet 5.5 | GPT-6.1 Sol |
|---|---|---|
| 150K cached + 20K fresh input, 15K output (under 272K) | $0.220 | $0.205 |
| 380K cached + 20K fresh input, 15K output (over 272K) | $0.266 | $0.381 |
Under the threshold, Sol’s cheaper cache wins by about 7 percent. Over it, the surcharge makes Sol about 43 percent more expensive. For related pricing patterns, see our API vs. Plus vs. Business guide.
Which to pick, by job
| Your job | Lean toward | Why |
|---|---|---|
| Chat or an assistant a person waits on | Sonnet 5.5 at high or medium | First answer in about 12 s at high, against about 68 s for Sol at high |
| High-volume background automation, cache-heavy, under 272K tokens | GPT-6.1 Sol | $0.32 to $0.72 per index task and half-price cache reads |
| Prompts regularly over 272K tokens | Sonnet 5.5 | No long-prompt surcharge listed, but check tokens used |
| Documents, slides and spreadsheets | Sonnet 5.5 | Leads GDPval-AA and AA-Briefcase |
| Answering from knowledge without lookups | GPT-6.1 Sol | Leads AA-Omniscience, 42 to 32 |
| Coding agents | Run your own test | AA favors Sonnet; DeepSWE favors Sol, but the Sonnet figure there is secondhand. Harness choice also moves cost, as in our harness-versus-model guide |
If two models look close for your work, our guide to comparing AI models on your own tasks turns this into an afternoon’s test. For the tier above, see our Opus 5.5 and Gemini 4 Argon comparison.
What we couldn’t confirm
- Independent cost and score rows for Sonnet at medium or low effort, and for Sol at medium, which tracker docs list as its API default. Anthropic says Sonnet’s low and medium settings beat Sonnet 5’s best score at about a tenth of the cost, but that compares against Sonnet 5, not Sol.
- The xhigh rows and the 3.8-times figure, which rest on a tracker’s copy of AA’s pages.
- Sol’s long-prompt surcharge, which we took from trackers rather than OpenAI’s pricing page.
- Whether AA re-ran Sonnet’s GDPval-AA and AA-Briefcase scores after Anthropic’s bug fix.
- How either model performs on your tasks, which no benchmark can tell you.
Frequently Asked Questions
Is Claude Sonnet 5.5 better than GPT-6.1 Sol?
On Artificial Analysis’s composite at max effort, Sonnet 5.5 scores higher, 56 to 52, and wins most of the agentic and knowledge-work tests. Another tracker’s composite puts Sol slightly ahead, and Sol is cheaper per task, so “better” depends on whether you value score or cost.
Do Sonnet 5.5 and GPT-6.1 Sol cost the same?
Per token, nearly: both are $2 per million input tokens and $10 per million output tokens, though Sol’s cached input is half the price. Per task they differ: Artificial Analysis measured $7.67 for Sonnet and $0.72 for Sol at max effort.
Why does Sonnet 5.5 cost more per task at max effort?
It writes far more tokens, mostly hidden thinking. Across the same index, Artificial Analysis counted 420M output tokens for Sonnet at max against 67M for Sol. Dropping to high effort cuts Sonnet’s cost to $1.12 per task.
Which is cheaper at the same quality?
Sol. At an index score of about 52, a tracker’s copy of Artificial Analysis’s data puts Sonnet at xhigh effort at $2.75 per task and Sol at max effort at $0.72, roughly 3.8 times cheaper.
Which is faster?
Sonnet. Artificial Analysis measured about 127 tokens per second at max for Sonnet and 59 for Sol, and at high effort Sonnet’s first answer token arrived in about 12 seconds against about 68.
Which should a small business use?
Test both on your own tasks. Claude’s apps default to medium effort, so the max-effort rows in this article aren’t what you will experience there. Our own-task guide shows how to score outputs and cost per usable result.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
If the small-model end of the price ladder matters to your workload, our Claude Haiku 5.5 review puts the $0.10-per-million-token model through its paces and compares its rate card against GPT-6 Luna.
