Claude Opus 5.5 lists at exactly twice GPT-6 Sol. Run both flat out and Opus costs 5.6x as much; run Opus at its own default and it outscores Sol's best setting for 5% more money, and past 272K tokens it is the cheaper bill.

Two cards, one shared line

Opus 5.5 is $4.00 in and $20.00 out, 20% under Opus 5. GPT-6 Sol is $2.00 and $10.00, half of what GPT-5.6 Sol charges on its promotional rate. Put those two sentences side by side and you notice that Opus 5.5 now costs exactly what last month's Sol cost on input and output. Anthropic's new flagship landed on OpenAI's old price, and OpenAI moved.

The line nobody mentioned is the cache hit. Anthropic dropped Opus 5.5 to a 0.05x multiplier, half the 0.1x most Claude models use, which puts it at $0.20. Sol's cached input is also $0.20. On an agent loop that rereads the same context every turn, that is the line doing most of the work, and on that line there is no gap at all.

Anthropic and OpenAI pricing pages, read September 28, 2026. Both add 10% for regional or data-residency endpoints. Sol's 1.05M context is 922K of input plus 128K of output.

Same effort, 3.2x to 5.6x

Per-token prices don't tell you what a job costs, because the two models don't spend the same number of tokens. They don't even count tokens the same way. Artificial Analysis sidesteps both problems by running its whole Intelligence Index (v4.3.2) through each model at each effort level and reporting what the run actually billed.

Artificial Analysis Intelligence Index v4.3.2, cost to run the full index at list price, read September 28, 2026.

The ratio never drops to the 2x on the price list. At every setting Opus 5.5 writes more, and at max it writes a lot more: 119,000 output tokens per task against Sol's 31,000. That is 3.8x the tokens at 2x the price. The bill lands at 5.6x rather than 7.7x mostly because both models pay the same $0.20 for the cached input the index rereads.

Max effort on Opus 5.5 is also slow. AA measured 682.71 seconds to first token and 798 seconds per task. Simon Willison hit the 128K output ceiling twice on max while the model was still thinking, and paid $2.56 for each failed attempt. We would not put max behind anything a user waits on.

Line them up by score instead

Comparing max to max assumes you would run both models the same way. You wouldn't. You pick a score you need and buy it as cheaply as you can. Here are all eleven configurations AA published, sorted by score, with the bar showing what the index run cost.

• 58Opus 5.5 max$8,708
• 56Opus 5.5 xhigh$4,057
• 54Opus 5.5 high$2,172
• 51Opus 5.5 medium (default)$1,627
• 48Sol max$1,550
• 44Sol xhigh$865
• 43Sol high$610
• 42Opus 5.5 low$860
• 40Sol medium (default)$417
• 34Sol low$268
• 28Sol, reasoning off$448

Accent bars are Opus 5.5, grey bars are GPT-6 Sol. Score is AA Intelligence Index v4.3.2.

Two things jump out. Below 48, Sol wins every time. Opus 5.5 on low scores 42 for $860, and Sol on xhigh scores 44 for $865, so for the same money Sol does better. If your work is fine around the low 40s, stop reading and use Sol.

Above 48 there is no Sol row at all, and the first Opus row is its own default. Opus 5.5 on medium scores 51 for $1,627. Sol on max scores 48 for $1,550. That is three points more for $77, a 5% premium. It is also faster: 203 seconds per task and a 13-second first token, against Sol-at-max's 351 seconds and 118. For this pair, that's the comparison we'd actually use, and it reads nothing like "twice the price."

One Sol oddity worth knowing: with reasoning switched off it scored 28 and cost $448, more than the $268 it cost on low, where it scored 34. Off is not the cheap setting. Low is.

At 272,001 tokens the order flips

AA's index is mostly short prompts. Coding agents are not. OpenAI's model page for Sol says that over 272K input tokens it bills "2x input and cache rates and 1.5x output for the full request," which makes Sol $4.00 in, $0.40 cached and $15.00 out. Opus 5.5 has no long-context tier and bills its whole 1M window at the standard rate. So past the threshold Sol's input equals Opus 5.5's, and its cache read is double.

We priced one agent turn at different context sizes, with 90% of the prompt coming from cache and 2,000 tokens of output. We left cache writes out, since a long-running loop pays them once and then reads for many turns.

One request: 90% cache hits, 10% fresh input, 2,000 output tokens, Standard tier, no cache writes. Assumes the same token count on both models, which is not quite true because the tokenizers differ.

At 272K Opus 5.5 costs 1.60x as much. One token later, it costs 0.83x as much. Cache hits matter less than you'd guess: with no caching at all, a 400K turn is $1.640 on Opus and $1.630 on Sol, a one-cent difference. At that size, anything above a 12.5% cache hit rate makes Opus the cheaper of the two (the break-even is lower for bigger prompts, higher for smaller ones), and agent loops run far above that.

The tokenizer is the soft spot in this table. Anthropic says its tokenizer on 4.7 and later "produces approximately 30% more tokens for the same text" than the older Claude one, and nobody has published a clean Claude-vs-OpenAI ratio. If the same codebase and its output both come out 20% longer on Opus, the 0.81x at 400K becomes about 0.98x, still not worse than Sol. We would test your own repository in the token counter before trusting either number.

What Anthropic's 40% is measuring

The launch post says Opus 5.5 "costs 40% less to run than Opus 5," and one line later scopes it to "default settings" and "typical workloads." The rate card covers part of it: 20% off input and output, 60% off cache reads, which works out to about 37% on the cached 400K turn in the table above. The rest comes from a change most people will miss: Opus 5 defaulted to high effort, and Opus 5.5 defaults to medium. Leave the parameter unset and you get a cheaper setting as well as cheaper tokens.

Force max effort and the saving disappears. AA puts Opus 5.5 max at $5.98 a task and Opus 5 max at $5.86. The new model spent 260 million output tokens on the index against 140 million, 1.86x as many, which more than eats a 20% cut. If your code pins effort: "max" from the Opus 5 days, the upgrade costs you 2% more per task, not 40% less.

Compared at the same score, the story gets better for Anthropic, not worse. Opus 5 at max scores 51 on the current index. So does Opus 5.5 at medium, for $1.34 a task against $5.86. That is 77% less for the same number. The 40% claim is conservative, provided you let effort drop.

How we would split the traffic

Short prompts, high volume, and a quality bar in the low 40s: Sol at low or high. Nothing from Anthropic competes there, and Claude Sonnet 5 sits on the same $2 / $0.20 / $10 card as Sol if you want to test a second vendor at the same price.

Short prompts where you need to score above 48: Opus 5.5 on medium, not Sol on max. Five percent more money, three points more score, and it finishes sooner.

Agent loops that carry more than 272K of context: Opus 5.5, unless your own token counts show Claude's tokenizer running more than about 23% longer on your code, which is where the 400K example breaks even.

Anything that can wait a day: Opus 5.5 on the Batch API. It is $2.00 and $10.00, which is Sol's real-time card. Sol's own Batch and Flex price is $1.00 and $5.00, so Sol is still cheaper there. But if you already know you want Opus for a job, Batch is where the 2x disappears.

Where every number came from

• Anthropic: Pricing - Opus 5.5 at $4 / $5 / $8 / $0.20 / $20, the 0.05x cache-hit footnote, Batch $2 / $10, Fast $8 / $40; Opus 5 and Sonnet 5 rows; 1M context at standard pricing; the 1.1x inference_geo multiplier; the tokenizer note. Read September 28, 2026
• Anthropic: Introducing Claude Opus 5.5 - September 22, 2026. The 40% and 20% wording, "default settings," the 60% cache-read cut
• Anthropic: Models overview - 1M context, 128K max output, medium as the default effort
• Anthropic: Opus 5 overview - high as Opus 5's default effort
• OpenAI: GPT-6 Sol model page - $2 / $0.20 / $2.50 / $10; the 272K rule for the full request; 1,050,000 context, 922,000 max input, 128,000 max output; effort levels
• OpenAI: API pricing - Sol long-context rows including the $5.00 cache write, Batch, Flex and Fast rows; GPT-5.6 Sol at $4 / $20. Read September 28, 2026
• OpenAI Developer Community: GPT-6 Sol and GPT-6 Luna announcement - "50% lower API prices" against GPT-5.6 promotional pricing
• Artificial Analysis: GPT-6 Sol vs Claude Opus 5.5 - Intelligence Index v4.3.2 scores, cost to run, output tokens, TTFT and time per task for all eleven configurations. Read September 28, 2026
• Artificial Analysis: Claude Opus 5 - index 51 at max, $5.86 per task, 140M output tokens
• Simon Willison: Opus 5.5, Sol and Luna - two max-effort runs that hit the 128K output limit at $2.56 each
• Hacker News: Claude Opus 5.5 and GPT-6 Sol and Luna - submission times used for the release order