Claude Sonnet 5.5 kept Sonnet 5's $2 and $10. Of its five effort settings, one is ever the cheapest way to its score, and it is the one the API picks when you leave effort unset.
Anthropic shipped Sonnet 5.5 yesterday with an unchanged rate card and a promise that it "costs up to 30% less per task." Artificial Analysis had already run it at every effort level. We put those fifteen runs, Sonnet 5.5 plus Opus 5.5 and GPT-6 Sol, on one ladder and asked which ones you would ever buy.
Photo by Pieter Pienaar on Unsplash
Two things to take away
- Leave effort on the API default (high) and Sonnet 5.5 scores 46.7 for $1,176 on AA's index. Nothing from Anthropic or OpenAI gets there for less. Every other setting is matched or beaten for less money, by GPT-6 Sol below it or Opus 5.5 above it.
- Don't run it on max. It costs $269 more than Opus 5.5 on max and scores 1.6 points lower.
Nothing on the card moved
Sonnet 5.5 is $2.00 in and $10.00 out, with a $0.20 cache hit. That is Sonnet 5 to the cent, and it is also GPT-6 Sol's card, so the two mid-tier flagships now sell tokens at the same price. Opus 5.5 is double on every line except the cache hit, where Anthropic gave it a 0.05x multiplier and it lands on $0.20 too.
Two smaller numbers did change. The minimum prompt you can cache dropped from 1,024 tokens to 512, and the tool-use system prompt shrank from 354 tokens to 286. Neither will show up on an invoice unless you send a lot of short, tool-heavy requests. There is still no Fast mode on Sonnet.
| Per 1M tokens | Sonnet 5.5 | Sonnet 5 | Opus 5.5 | GPT-6 Sol |
|---|---|---|---|---|
| Input | $2.00 | $2.00 | $4.00 | $2.00 |
| Cache write | $2.50 (5 min), $4.00 (1 hr) | $2.50 / $4.00 | $5.00 / $8.00 | $2.50 |
| Cache hit | $0.20 | $0.20 | $0.20 | $0.20 |
| Output | $10.00 | $10.00 | $20.00 | $10.00 |
| Batch in / out | $1.00 / $5.00 | $1.00 / $5.00 | $2.00 / $10.00 | $1.00 / $5.00 |
| Fast mode | none | none | $8.00 / $40.00 | $4.00 / $20.00 |
| Over 272K input | no change | no change | no change | 2x in, 1.5x out |
| API default effort | high | high | medium | medium |
Anthropic and OpenAI pricing pages, read September 29, 2026. Regional endpoints add 10%: Vertex AI US and EU list Sonnet 5.5 at $2.20 / $0.22 / $11.00, and inference_geo: "us" on the Claude API is 1.1x. Sonnet 5.5 uses the same tokenizer as Sonnet 5, so the same text is the same number of tokens.
Against Sonnet 5, "up to 30%" undersells it until max
Anthropic hasn't published how it measured the 30%. VentureBeat asked and had no answer by the time it ran its story. Artificial Analysis gives us something we can check: what it cost to run its whole Intelligence Index (v4.3.2) through each model at each setting.
| Effort | Sonnet 5.5 | Sonnet 5 | Run cost change | Score change |
|---|---|---|---|---|
| low | 35.8 / $544 | 24.3 / $653 | -17% | +11.6 |
| medium | 40.7 / $701 | 28.1 / $1,218 | -42% | +12.7 |
| high | 46.7 / $1,176 | 31.7 / $2,086 | -44% | +15.1 |
| xhigh | 51.9 / $2,738 | 34.4 / $3,255 | -16% | +17.5 |
| max | 56.0 / $8,977 | 38.2 / $6,998 | +28% | +17.8 |
Score / cost to run the full index at list price. Artificial Analysis Intelligence Index v4.3.2, read September 29, 2026.
On medium and high, the two settings most people will actually run, the new model is more than 40% cheaper and scores 13 to 15 points higher. That's better than the headline. Anthropic's other claim, that Sonnet 5.5 on low or medium beats Sonnet 5's best for about a tenth of the cost, holds up on this index almost exactly: 40.7 for $701 on medium against 38.2 for $6,998 on max is a tenth, give or take a percent.
Max breaks the pattern. Sonnet 5.5 spent about 193,000 output tokens per task there, which AA calls "the highest token use we have measured," around 60% above Opus 5.5 on max. Per task, AA puts it at $7.60 against Sonnet 5's $5.09. If you pinned effort: "max" on Sonnet 5 and just swap the model ID, your bill goes up.
Fifteen settings, ten worth buying
Comparing Sonnet 5.5 with its predecessor tells you whether to upgrade. It doesn't tell you whether to use it. For that, sort every setting of the three models you would realistically pick between by score, and cross out any setting where something else scores at least as much for no more money. What's left is the frontier.
- 57.6Opus 5.5 max$8,708
- 56.0Opus 5.5 xhigh$4,057
- 56.0Sonnet 5.5 max$8,977
- 53.6Opus 5.5 high$2,172
- 51.9Sonnet 5.5 xhigh$2,738
- 51.2Opus 5.5 medium$1,627
- 47.5Sol max$1,546
- 46.7Sonnet 5.5 high$1,176
- 44.1Sol xhigh$860
- 42.8Sol high$610
- 42.3Opus 5.5 low$860
- 40.7Sonnet 5.5 medium$701
- 39.8Sol medium$416
- 35.8Sonnet 5.5 low$544
- 33.9Sol low$268
Bars show the cost to run AA's index. Accent is Sonnet 5.5, mid grey is Opus 5.5, light grey is GPT-6 Sol. Struck-through rows are beaten on both score and cost by another row.
Sol holds everything up to 44. Sonnet 5.5 on low and medium both lose to it: Sol on high scores 42.8 for $610, more than Sonnet on medium gets for $701. Opus 5.5 holds everything from 51 up. Sonnet on xhigh scores 51.9 for $2,738, and Opus on high scores 53.6 for $2,172, so the model with twice the token price is $566 cheaper and 1.7 points better.
Sonnet 5.5 on high is the one gap. 46.7 for $1,176 sits between Sol on xhigh (44.1, $860) and Sol on max (47.5, $1,546). AA says the same thing in its write-up: high is "the most competitive" setting, "very narrowly behind GPT-6 Sol" on intelligence at about the same cost per task. If the score you need falls between 44.1 and 46.7, Sonnet 5.5 on high is the cheapest way to get it. Anywhere else on the ladder, it isn't.
On max, the cheap Claude isn't the cheap one
Sonnet 5.5 on max cost $8,977 to run the index. Opus 5.5 on max cost $8,708 and scored 57.6 against 56.0. Opus on xhigh matches Sonnet's 56.0 exactly for $4,057, which makes Sonnet on max 2.2x the price of Opus for the same score. Half the token price doesn't help when you write 60% more tokens and still come out lower.
Anthropic's own launch charts show the same thing on three of four benchmarks that report cost per task. On FrontierCode 1.1, Sonnet 5.5 on max cost $20.78 a task and scored 46.2%, while Opus 5.5 on max cost $6.19 and scored 54.4%. Anthropic's footnote says Sonnet on max more often launched a multi-subagent code-review skill, which explains the bill but doesn't change it. On AA-Briefcase it is $29.19 against $21.05, and on Terminal-Bench 4.0 $12.54 against $11.24. CursorBench is the exception, where Sonnet on max ($9.67) undercuts Opus on max ($13.43).
Terminal-Bench 4.0 is the one place max earns its money. Sonnet 5.5 on max scored 70.6%, the top number on the chart, above Opus 5.5's best of 66.4%. It cost $12.54 an attempt against $7.35 for that Opus run. If terminal agents are your workload, that 4.2 points is worth testing. AA's independent run of the same benchmark puts the two at 63.6% and 59.6%, a similar gap at lower absolute scores.
One more reason to keep max away from anything a user is waiting on. Simon Willison's max-effort test thought for the full 128,000 output tokens, billed $1.28, and returned nothing. The same prompt on xhigh finished in 41 seconds for under six cents. AA measured 854 seconds per task on max against 216 on high.
The default is the good setting, if you're on the API
Opus 5.5 defaults to medium on the API. Sonnet 5.5 defaults to high there, and to medium in Claude Code and the Claude apps. That's lucky for API callers who never set the parameter: they land on the one Sonnet setting that is on the frontier. Claude Code users land on medium, which Sol on high beats.
Anthropic's docs say the effort levels were recalibrated and tell you to "re-run your effort sweep rather than carrying a setting over." The AA numbers show why. Sonnet 5 on xhigh scored 34.4. Sonnet 5.5 on low already scores 35.8, for a sixth of the money ($544 against $3,255).
Two changes will break old Sonnet 5 code before they affect cost. thinking: disabled now returns a 400, and so does forced tool use. If you had thinking switched off to save output tokens, the replacement is between_tools, allowed on high effort or below.
Should you move off Sonnet 5 today?
Yes, unless you run it on max. Anthropic keeps Sonnet 5 alive until at least June 30, 2027, so nobody is forcing the move. But a team running Sonnet 5 on high that spends $1,000 a month would, on AA's ratio, spend about $560 for the same number of tasks after swapping the model ID, and get answers that score 15 points higher. We can't think of a reason to wait.
Picking a model for something new is a different question. Sonnet 5.5 on high makes sense when you need a score in the mid 40s. Below that, GPT-6 Sol on medium or high gets the same result for less. Above 50, go straight to Opus 5.5 on medium or high rather than pushing Sonnet to xhigh or max.
Terminal agents are the exception worth a trial run. Sonnet 5.5 on max tops Terminal-Bench 4.0 on both Anthropic's and AA's numbers, but at $12.54 an attempt we would cap output tokens and measure on your own tasks first. The calculator will price a run at any token count you plug in.
Sources
- Anthropic: Pricing - Sonnet 5.5, Sonnet 5 and Opus 5.5 rows, Batch, Fast mode, the 1.1x inference_geo multiplier, tool-use system prompt tokens. Read September 29, 2026
- Anthropic: Claude Sonnet 5.5 - September 28, 2026. The "up to 30% less" and "about a tenth of the cost" wording, default effort by surface, benchmark table and cost-per-task charts
- Anthropic: What's new in Sonnet 5.5 - same tokenizer as Sonnet 5, recalibrated effort, 512-token cache minimum, breaking changes
- Anthropic: Model deprecations - Sonnet 5 retirement not sooner than June 30, 2027
- Artificial Analysis: Claude Sonnet 5.5 and launch write-up - Intelligence Index v4.3.2 scores and cost to run for all five effort levels, per-task cost and output tokens, time per task, Terminal-Bench 4.0 run. Read September 29, 2026
- Artificial Analysis: GPT-6 Sol vs Claude Opus 5.5 - scores and cost to run for the Opus 5.5 and Sol rows
- OpenAI: API pricing and GPT-6 Sol model page - GPT-6 Sol Standard, Batch, Fast and long-context rows; medium as the default effort
- Google Cloud: Vertex AI pricing - Sonnet 5.5 global and US/EU multi-region rows
- VentureBeat: Anthropic launches Claude Sonnet 5.5 - no methodology for the 30% claim at publication
- Simon Willison: Claude Sonnet 5.5 - the $1.28 max-effort run that hit 128K output, and the xhigh run