EXPLAINER · WEDNESDAY, SEPTEMBER 23, 2026

Claude Opus 5.5 vs GPT-6 Sol explained: token price versus cost per task, the benchmark rows each launch left out, and the safeguards that change which model answers you

Anthropic and OpenAI shipped 90 minutes apart on September 22 and both led with price. Neither launch compares against the other, because both charts compare against Opus 5. This page puts both vendors' own numbers on shared rows, works out what a task actually costs on each, explains why max effort can cost more for a worse answer, and spells out the Opus 5.5 fallback that sends most security work to Opus 4.8.

This explains reporting by Anthropic, September 22, 'Introducing Claude Opus 5.5'.
Read the original first: https://www.anthropic.com/claude-opus-5-5

In one minute

On September 22, Anthropic shipped Claude Opus 5.5 at $4 per million input tokens and $20 per million output, down from $5 and $25. OpenAI shipped GPT-6 Sol at $2 and $10 and GPT-6 Luna at 10 cents and 50 cents, half the GPT-5.6 prices, about 90 minutes later.
The two price cuts are different in kind. Artificial Analysis scores Opus 5.5 at 58, the highest score it has recorded, and GPT-6 Sol at 48, one point above GPT-5.6 Sol. Sol is roughly the same model at half price. Opus 5.5 is a stronger model at a lower price.
Per token, Sol is half the price of Opus 5.5. Per task on Artificial Analysis's index, both at max effort, Sol cost $1.06 and Opus 5.5 cost $5.98, because Opus 5.5 at max spends far more tokens thinking.
Both vendors' charts compare against Opus 5, and on at least one shared row, OSWorld 2.0, they report very different Opus 5 scores: 74.0 from Anthropic, 60.3 from OpenAI. Treat cross-vendor charts as each company's own harness, not a common test.
Opus 5.5 routes most cybersecurity requests to Opus 4.8 and some biology requests to other models, without refusing. Security teams testing it may be testing Opus 4.8.

What shipped, exactly

Anthropic released Claude Opus 5.5 on Tuesday, September 22. The API name is claude-opus-5-5, and it is on the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure. The launch page claims it is 30 percent faster at writing output than Opus 5, and that its writing is clearer, which The Decoder framed as Anthropic promising to fix 'Claudish' prose.

About 90 minutes later, TechCrunch reports, OpenAI released GPT-6 Sol and GPT-6 Luna. These are the middle and small models of the GPT-6 family. The big one, GPT-6 Astra, came out earlier this month and is unchanged. Sol is for coding, review, debugging and analysis. Luna is for high-volume, narrow jobs like summaries and extraction. Both are in ChatGPT Work and Codex for paid plans and in the API as gpt-6-sol and gpt-6-luna. Free and Go users get Luna in the desktop app only.

OpenAI told VentureBeat the new rates are 'permanent prices, not promotional or introductory pricing.' That matters because GPT-5.6 launched with promotional rates.

Two price cuts that are not the same thing

A token is a chunk of text, roughly three quarters of a word. You pay for input tokens, what you send, and output tokens, what the model writes back. Thinking counts as output. That last part is the one that surprises people, and it is why the cost of a task is not the same as the price of a token.

Here is the price table from the two launches, per million tokens.

Model                Input   Output   Cached input
Opus 5 (old)         $5.00   $25.00   $0.50
Opus 5.5             $4.00   $20.00   $0.20
GPT-5.6 Sol (old)    $4.00   $20.00   not listed
GPT-6 Sol            $2.00   $10.00   90% off input
GPT-5.6 Luna (old)   $0.20   $1.20    not listed
GPT-6 Luna           $0.10   $0.50    90% off input

Anthropic cut the per-token price 20 percent and says the model uses fewer tokens per task, which together 'nets out to a 40% drop in costs' against Opus 5. OpenAI cut the per-token price 50 percent and says the savings come from 'improvements in caching and inference.'

The difference that matters is what you get for the money. Artificial Analysis, an independent benchmarking firm, runs every model through the same set of tests and publishes one combined score. Opus 5.5 scored 58, first of 212 models it tracks. GPT-6 Sol at max effort scored 48, 18th. The Decoder, reading the same data, reports Sol moved from 47 to 48 and Luna stayed at 37. So Sol is the model you already had, at half the price. Opus 5.5 is a better model at 80 percent of the old price.

Cost per task, worked out

Take a plain agent task: 50,000 tokens of input, say a ticket plus some files, and 10,000 tokens of output including thinking. At list prices with no caching:

Opus 5.5    50,000 x $4/M  = $0.20   10,000 x $20/M = $0.20   total $0.40
GPT-6 Sol   50,000 x $2/M  = $0.10   10,000 x $10/M = $0.10   total $0.20
GPT-6 Luna  50,000 x $0.10/M = $0.005  10,000 x $0.50/M = $0.005  total $0.01

That arithmetic assumes each model writes the same number of tokens, and they do not. Artificial Analysis publishes what it actually cost to run its test suite on each model. At max effort, Opus 5.5 cost $5.98 per task and GPT-6 Sol cost $1.06. That is a gap of about 5.6 times, not 2 times, because Opus 5.5 at max thinks for much longer.

Agents change the math again. A coding agent re-sends the same long context on every step, so most of its input is cached. Anthropic's cached read price fell from 50 cents to 20 cents. OpenAI discounts cached input 90 percent. If your workload is mostly cached input, compare the cached column, not the headline column.

The honest summary: per token, Sol is half the price of Opus 5.5. Per task, it depends almost entirely on the effort setting and how much of your input is cached, and the only reliable number is the one you measure on your own work.

Effort levels, and why max failed on launch day

Both companies now sell the same model at several effort levels. Anthropic's page names low, medium, high, xhigh and max, and uses medium as the default for its comparisons. Effort is a budget for thinking. Higher effort lets the model write more reasoning tokens before it commits to an answer. Those tokens are billed as output.

More thinking is not always better. On launch day Simon Willison ran his standing test, an SVG of a pelican riding a bicycle, on Opus 5.5 at max. The model kept reasoning until it hit its 128,000-token output limit, produced no image at all, and cost $2.56 for nothing. His verdict was that max 'over-thinks to breaking point.' Factory, quoted on Anthropic's page, called Opus 5.5 'the first model we'd default to at medium effort.'

OpenAI has the same problem from the other side. The Decoder notes that Luna at max effort matches Sol at xhigh while costing 78 percent less. When the small model at high effort ties the big model at medium, the choice of effort matters as much as the choice of model.

Start at medium on Opus 5.5. - Raise effort only for the tasks that fail at medium. - For Sol and Luna, test whether Luna at a higher effort already does the job.

Reading the benchmark rows the two launches share

Both companies compare against Claude Opus 5, which gives us a common reference. It also exposes a problem. Here are the rows where both vendors publish a number, with each vendor's own figure for Opus 5.

Benchmark          Opus 5.5  GPT-6 Sol  Opus 5 per Anthropic  Opus 5 per OpenAI
AutomationBench    40.0      33.2       26.9                  26.9
OSWorld 2.0        81.8      60.5       74.0                  60.3

On AutomationBench the two companies agree on Opus 5, so the rows are comparable: Opus 5.5 at 40.0 beats Sol at 33.2, and GPT-6 Astra is at 41.4 on Anthropic's table. On OSWorld 2.0, a test of using a real computer, the two companies report Opus 5 at 74.0 and 60.3. Same model, same test name, 14 points apart. The likely reason is a different harness or effort setting, but neither launch says. Either way, Sol's 60.5 and Opus 5.5's 81.8 were not measured the same way, and you should not subtract one from the other.

OpenAI's DeepSWE number for Sol is 68.8, against 69.9 for Claude Fable 5. The Decoder reports Opus 5 at 73.7 on the same test at max effort. Anthropic did not publish a DeepSWE row for Opus 5.5.

Anthropic's Terminal-Bench 4.0 row is the biggest jump on its table: Opus 5.5 at 66.4, ahead of Fable 5.1 at 55.8 and GPT-6 Astra at 57.9. OpenAI did not publish Terminal-Bench 4.0 for Sol, which The Decoder counts as one of the tests OpenAI left out.

The safeguards that change which model answers you

This is the part that is not on the price table. Opus 5.5 is the first Opus model to ship with the same class of safeguards as Claude Fable 5.1 for cybersecurity, biology and distillation. Anthropic says all of these 'fall back to another model transparently.'

Transparent fallback means that when a request trips a safeguard, you do not get a refusal. You get an answer from a different, older model. For security work, Anthropic's page says routine bug finding and fixing in your own code stays on Opus 5.5, 'but most cybersecurity tasks will be re-routed to Opus 4.8.' Biology work that trips the safeguards falls back the same way. The paths back to full access are the Cyber Verification Program, which adds Opus 5.5 in the coming weeks with three tiers, and a new Life Sciences Verification Program for labs, startups and drug companies.

Distillation is different. Distillation means sending a model huge numbers of questions and training your own model on the answers. Classifiers that detect it block on Opus 5.5 with no fallback. Opus 5.5 also ships with preserved thinking, first introduced with Fable 5.1. It stops API users from editing Claude's earlier turns to pull out its reasoning, and it applies to API accounts created after August 31, 2026. Thinking cannot be turned off on Opus 5.5.

The practical effect for an MSP or a security team: if you benchmark Opus 5.5 on triage, pentest planning or exploit analysis, some of those answers may come from Opus 4.8. The price you pay and the model you think you are testing may not match what actually ran.

Who is affected

Case	Status
Teams on Claude Opus 5 through the API	Moving to claude-opus-5-5 is 20 percent cheaper per token and 60 percent cheaper on cached reads. Anthropic claims about 40 percent lower cost per task. Test at medium effort first.
Teams on GPT-5.6 Sol or Luna	Switching the model name to gpt-6-sol or gpt-6-luna halves the price. Artificial Analysis shows about the same intelligence, and a drop of about 100 Elo on office-work tasks for Sol.
Security teams and MSPs using Claude for cyber work	Most cybersecurity requests on Opus 5.5 go to Opus 4.8 unless your organization joins the Cyber Verification Program.
Biology and life sciences R&D	Some requests fall back to other models. Apply to the Life Sciences Verification Program for full access.
API users who edit prior turns or strip thinking	Preserved thinking applies to API accounts created after August 31, 2026, and thinking cannot be disabled on Opus 5.5.
Claude Code users	Fast mode on Opus 5.5 runs up to 2.5 times faster at $8 in and $40 out.

What to do

Find out which model your tools default to. A config that says gpt-5.6-sol is now paying twice what gpt-6-sol costs.
Run one real weekly task on Opus 5.5 at medium effort and on GPT-6 Sol. Record cost, time and whether the answer was right.
Do not use max effort as a default. Raise effort only for tasks that fail at medium.
If most of your agent's input is repeated context, compare the cached-input prices, not the headline prices.
If you do security work on Claude, check whether answers are coming from Opus 4.8, and apply to the Cyber Verification Program if you need Opus 5.5 itself.
When you read a cross-vendor chart, check whether both vendors report the same score for the reference model before you compare anything else.

What is still unknown

OpenAI's launch post returns 403 to automated readers, so every GPT-6 figure on this page comes from VentureBeat, TechCrunch and The Decoder's reading of it, plus Artificial Analysis. We did not read OpenAI's page directly.
Why the two companies report different OSWorld 2.0 scores for the same Opus 5. Neither launch explains its harness in enough detail to say.
Whether an API response tells you when a safeguard has routed your request to Opus 4.8. The launch page says the fallback is transparent; we did not find a description of how it is labeled.
Opus 5.5's context window. The launch page we read does not state it.
There is no same-harness, head-to-head public result for Opus 5.5 and GPT-6 Sol yet. The cost per task figures are from Artificial Analysis's own test suite at max effort, which may not match your workload.
Anthropic's 40 percent cost claim is its own measurement on 'typical workloads.'

Sources

Today's full edition: AI News Report · every headline, every morning.