A folkfox special report, 28 September 2026
Sonnet 5.5 vs Opus 5.5: the gap, measured.
Claude Sonnet 5.5 is here: a faster model with a striking jump in coding and knowledge work. Twelve small instruments and a wall of same-prompt experiments show what changed, what it costs, and where the claims need care.
By the folkfox newsroom, An interactive release edition
Sonnet 5.5 vs Opus 5.5: Anthropic's newer mid-tier model costs half as much per token and posts scores close to Opus 5.5's. Some launches ask for a new leaderboard; this one asks for a new working assumption.
Anthropic released Claude Sonnet 5.5 on . Claude Sonnet is the middle tier of the Claude family, placed by Anthropic between Opus and Haiku. Its headline is unusual: the listed API token rates stay the same as Sonnet 5, while Anthropic says the model generates output more than 30 per cent faster and uses fewer tokens per task. The charts below put Claude Sonnet 5.5 vs Opus 5.5 side by side wherever the published figures allow, with vendor claims labelled as such. Each chapter stands on its own if you leave the controls untouched.
Reuters reported it as the second model in the Claude 5.5 family. Like a fox at the treeline, this report watches first, weighs every claim against its source, and only then pounces on the numbers worth keeping.
- 70.6%Terminal-Bench 4.0, against 66.4% for Opus 5.5 at Xhigh and 10.3% for Sonnet 5. Vendor-reported.
- Halfthe token price of Opus 5.5: $2 / $10 against $4 / $20 per million. Confirmed.
- 30%+faster output generation than Sonnet 5. Vendor-reported.
- $0.20per million tokens for cache reads, unchanged. Confirmed.
- 30%lower cost per task, at most, in Anthropic's tests. Vendor-reported.


Chapter one. The arrival.
The second instrument arrives
Same price, new pace.
Opus 5.5 arrived first, on 22 September, as Anthropic's announcement records. Six days later, Anthropic released the model most people will be tempted to use for ordinary work: a quicker, calmer companion for the daily trail.
Sonnet 5.5 is available across Claude platforms and major cloud providers; Amazon Web Services lists it on Amazon Bedrock and on Claude Platform on AWS in North America. The API model name is claude-sonnet-5-5. Anthropic positions it between Opus 5.5, for open-ended work requiring sustained judgement, and a forthcoming Haiku 5.5 for high-volume work. Sonnet 5.5 vs Opus 5.5 is the first question of the week, and that positioning matters more than a single chart win: a model you can put into the daily loop changes how often work gets attempted.
The release ledger
Choose a row to read its scope| Release fact | Published value | Reporting note |
|---|---|---|
| Release date | 28 September 2026 | |
| Output generation | 30%+ faster | |
| API price per million tokens | $2 input / $10 output |
The first delight is simple: the name in the model picker has changed, and the billable token rates have not.
Sonnet 5.5 vs Opus 5.5: the shape of the release
Toggle a model to compare

Chapter two. The terminal.
A coding leap you can see
Big number. Read the small print.
Anthropic reports 70.6 per cent for Claude Sonnet 5.5 on Terminal-Bench 4.0, against 10.3 per cent for Sonnet 5.
That is a dramatic change within Anthropic's published comparison, and a reason to try the new model on real code. In Sonnet 5.5 vs Opus 5.5 terms it is 70.6 against 66.4, with Opus at Xhigh effort. It is still an evaluation result, not a promise that seven out of ten tickets in your repository will land cleanly. Terminal-Bench is one part of a larger coding story; CursorBench and FrontierCode test different kinds of agent work. The Decoder notes that independent testing still needs to confirm Anthropic's claims, so treat the number as a scent, not a verdict: follow it, then verify.
Three coding lenses
Switch benchmark; hover or focus rows for the exact scoreSonnet 5.5 vs Opus 5.5, with Sonnet 5 as the baseline
Opus 5.5 is the ring

Chapter three. The effort dial.
Sonnet 5.5 effort settings: more thought is not always more score
Thinking longer is not thinking better.
FrontierCode makes the awkward detail visible. Claude Sonnet 5.5 scores 52.1 per cent at Xhigh effort, then 46.2 per cent at Max.
Anthropic's footnote says Max more often invoked a code-review skill, splitting review across subagents. In two cases examined by Cognition, this led to a timeout or extra edits outside the task's scope. FrontierCode asks whether a code change could be merged without human edits, so helpful changes beyond the brief can still count against a model. The lesson is practical: Claude effort levels are a setting to test, not a dial to leave at maximum. In Sonnet 5.5 vs Opus 5.5 terms the Max run trails Opus by 8.2 points on FrontierCode, 46.2 against 54.4. For the wider field, folkfox covered the GPT-6 Sol and Luna price cut on 22 September.
The effort switch
Choose a run

Chapter four. The desk.
The model at the working desk
Where it earns its keep.
The appeal reaches beyond code. Anthropic reports Claude Sonnet 5.5 close to Opus 5.5 on two knowledge-work evaluations.
Sonnet 5.5 vs Opus 5.5 comes to 1,844 against 1,846 on GDPval-AA and 1,811 against 1,822 on AA-Briefcase. GDPval-AA v2.1 covers tasks across occupations. AA-Briefcase v1.1 tests longer-horizon knowledge work. These are scores on their own scales, not percentages, so their absolute values should not be compared with the coding bars above. Anthropic says Artificial Analysis ran the evaluations on a pre-release deployment that had a structured-output bug. It expects any effect to be small, but that is its assessment.
Sonnet 5.5 vs Opus 5.5 on two working-desk scores
Switch evaluationSonnet 5.5 vs Opus 5.5 on the two desk scores
Same axis, four models

Chapter five. The eyes.
Eyes, hands, and the page
It reads the page as you do.
The new Claude Sonnet looks at screens and charts with a different level of fluency from Sonnet 5, according to Anthropic's figures.
On OSWorld 2.1 (partial), Sonnet 5.5 is reported at 80.1 per cent versus Sonnet 5 at 57.0. On Chartography without tools, it is 61.6 per cent versus 15.6. Opus 5.5 remains ahead on both: in Sonnet 5.5 vs Opus 5.5 terms, 81.8 against 80.1 on OSWorld and 64.4 against 61.6 on Chartography. These are valuable leads for teams working with graphical interfaces, but a partial computer-use benchmark and a chart-recognition test cannot substitute for a trial in your own application.
Two ways of seeing
Choose a task type

Chapter six. The price.
Claude Sonnet API pricing: a task, not a token
A task costs more than its tokens.
The published Claude Sonnet API pricing is $2 per million input tokens and $10 per million output tokens, the same as Sonnet 5, as Reuters also reported. The Decoder adds $0.20 per million tokens for cache reads, and Anthropic's pricing documentation lists the current rates.
On Claude Sonnet 5 pricing, Anthropic's developer account says nothing changes: its rate post on X.
In Sonnet 5.5 vs Opus 5.5 terms, the token price is half: $2 and $10 against $4 and $20 per million. Anthropic says the new model typically needs fewer tokens to finish comparable work, and costs up to 30 per cent less per task in its testing. “Up to” is doing work in that sentence. Try the calculator with your own expected token mix; it calculates list-price input and output charges only, before caching, tools, platform fees, and changes in token use. It does not turn Anthropic's measured claim into a universal discount. The real bill hides in the thicket of tokens, tools and retries, so price the whole trail, not the first step.
A small cost calculator
Move either sliderWhere the money goes
Follows the calculator above

Chapter seven. The speed.
A faster creative loop
Faster loops, braver drafts.
Anthropic reports that Claude Sonnet 5.5 generates output more than 30 per cent faster than Sonnet 5. That will be felt most when a person is waiting and revising.
Fast generation is only part of elapsed task time. Tools, network calls, human review and retry loops can dominate. Early testers describe fewer steps and quicker iteration, but those observations come from selected partner reports on Anthropic's launch page. Use the slider to see what a change in generation speed alone would mean for a hypothetical task.
The waiting-time thought experiment
Adjust the share spent generatingTime saved, drawn out
Follows the slider above

Chapter eight. The field.
Reports from the field
What people said, and who said it.
The most interesting launch evidence comes from people who put Claude Sonnet 5.5 into a real workflow before release.
Anthropic quotes partner tests from Slack, Zendesk, Balyasny and others. They are useful directional reports, each with a different task, baseline and measurement. Select a workflow below. The table retains the source, denominator where one is given, and the limit that stops a figure becoming a universal claim. OfficeChai singled out one internal test in which two reviewers judged a 10-slide earnings review ready to send without changes, a useful hint for content teams, though not a benchmark.
Early tester ledger
Filter by workflow| Tester | What they reported | Scope |
|---|
What early testers reported
Three tester figures, side by sideSonnet 5.5 vs Opus 5.5, same prompt: what people posted on launch day
Within hours of release, people began running one prompt through two models and posting both results, and Sonnet 5.5 vs Opus 5.5 was one of the most common pairings. Claude's own accounts started the pattern with Sonnet 5 against Sonnet 5.5, and the community extended it to Opus 5.5 and GPT-6. These are single runs on chosen prompts, so read them as field notes: none is a controlled benchmark, and every time, cost and token figure is the poster's own. Each post is quoted in full, links are kept, and its video or photo loads from X only when you ask for it. Text and times were read on 28 September 2026.
From Anthropic's accounts
A thread of early experiments with Claude Sonnet 5.5.
A fall foliage simulator by @_re_pete, made with Sonnet 5 vs Sonnet 5.5.
Bouncing-ball physics tests by @forwardeditor from the same prompt, Sonnet 5 vs Sonnet 5.5.
Price is the same as Sonnet 5: $2/$10 per million tokens with cache reads at $0.20.
Thinking is always on but you can use `between_tools` to turn off upfront thinking.
Here’s our Sonnet 5.5 migration guide: platform.claude.com/docs/en/models….
Sonnet 5.5 vs Opus 5.5, one prompt
opus 5.5 vs sonnet 5.5
same prompt, same skill: "create a short 10 second ad for markstudios"
opus took 17 minutes. sonnet took under 4
every frame is code, no ai video model
which one would you post?
A Sonnet 5.5 vs Opus 5.5 run where Sonnet finished first, on the poster's own clock.
Opus 5.5 vs Sonnet 5.5: candle benchmark
same prompt, one shot each. a lit candle in a dark room, 100% code, no assets. flame, melted wax, smoke, light flicker, 12s seamless loop
left is Opus, right is Sonnet. who wins?
A Sonnet 5.5 vs Opus 5.5 visual test; the poster leaves the verdict open.
Same prompt
Opus 5.5 version: claude.ai/artifact/Jugv9…
Sonnet 5.5 version:
claude.ai/artifact/1B7q3…
The Opus version allows the user to rotate images by direct manipulation. The Sonnet version does not.
A Sonnet 5.5 vs Opus 5.5 difference in interactivity, not speed.
same prompt. different models
i ran the exact same prompt on:
- sonnet 5
- sonnet 5.5
- opus 5
- opus 5.5
no changes to the prompt
there’s an obvious upgrade with the 5.5 models
testing more complex prompts now
let’s see how it holds up
what do you guys think?
Sonnet 5.5 vs Opus 5.5 next to their predecessors: the poster reports an obvious upgrade.
Sonnet 5.5 against GPT-6, one prompt
sonnet 5.5 vs gpt 6 astra
tested both models with same prompt at xhigh reasoning
but results came out really different
> astra took about 15 minutes and cost $16
> sonnet took about 25 minutes and cost $3
if you check the artifact, you'd see clear difference in each site,
i gave sonnet 5.5 and gpt 6 sol the same prompt: build minecraft
the results are staggering
sol: 27 min, 4.3M tokens
sonnet 5.5: 1h 11m, ~64M tokens, 40% of my 5 hour limit on pro
sonnet's is a full game: infinite terrain, survival, crafting, mobs, music
which one would you
Reported run times, same prompt
Three posters, three different prompts

Chapter nine. The guardrails.
The guardrails travel with it
They travel with the model.
This is the first Claude Sonnet launch with the cyber safeguards and fallbacks Anthropic uses for its more capable models.
Anthropic says Sonnet 5.5 matched or improved on Sonnet 5 across most measures of an automated behavioural audit of roughly 1,850 scenarios. It also says Opus 5.5 was slightly better overall and that no evaluation catches every failure. More capable cyber work can trigger a visible fallback to Sonnet 5; biology safeguards are the same as before. Those conditions belong in a deployment decision, not in a footnote nobody reads.
The Next Web called it the first Sonnet to ship with frontier-style cyber safeguards and with classifiers that block reasoning extraction. Guardrails work like a hedgerow: they keep a fast animal on the path. Security teams weighing the cost of a false alarm can compare notes in folkfox's AI security incident response brief and on the cybersecurity marketing page.
Safeguard map
Open each area| Area | Model behaviour | Decision point |
|---|---|---|
| Cybersecurity | Higher-risk requests may fall back to Sonnet 5 | Check the fallback against your legitimate workflow |


Chapter ten. The fit.
Claude Sonnet vs Opus: where Sonnet belongs
Not every job wants Sonnet.
The Sonnet 5.5 vs Opus 5.5 question is a routing question, not a crown. Any Claude model comparison should start with Anthropic's own recommendation: Sonnet 5.5 for well-scoped daily tasks, Opus 5.5 for complex work that needs sustained judgement.
The Sonnet 5.5 vs Opus 5.5 choice should be tested against the work, not made by label alone. Select a task to see a starting model and the reason. This is an editorial routing suggestion based on Anthropic's positioning, not a measured ranking of every model. Review a representative sample for quality, speed, actual token use and failure cost before changing a production workflow. Teams choosing which model to standardise on can talk it through with folkfox's AI consultancy, and teams turning documents and decks into a steady rhythm can take the results to content marketing services.
The task selector
Choose a work shape| Work shape | Starting model | Why |
|---|---|---|
| Bounded bug fix | Sonnet 5.5 | Clear success condition and fast iteration |


Chapter eleven. The migration.
The small print in a real migration
Read this before you switch.
For Claude API users, moving to Claude Sonnet 5.5 is more than swapping the model string.
The model ID is claude-sonnet-5-5. Anthropic's migration guide says adaptive thinking runs by default. Code that used thinking: {"type":"disabled"} must change: between_tools is the lowest thinking setting. Responses may begin with thinking blocks, so clients should read blocks by type rather than assuming the first block is text. The guide also calls for a fresh effort sweep and cost baseline.
A migration rehearsal
Tick items as you verify them locally| Migration check | Local status |
|---|---|
| To check | |
| To check | |
| To check | |
| To check | |
| To check |
How rehearsed are you?
Follows the checklist above

Chapter twelve. The verdict.
A release worth opening
Open it. Then test it yourself.
Claude Sonnet 5.5 is exciting because its case is ordinary work done better: more capable coding, stronger visual and knowledge tasks, faster output, and the same headline API token rates.
The honest reading is narrower than a victory lap. Anthropic's comparisons are vendor-published; effort levels and agent settings matter; partner reports are selected; independent post-launch evidence will take time. Yet the release is real and available now. The sensible next move is not to crown a champion. It is to put Sonnet 5.5 through the jobs Sonnet 5 already does for you and see which ones come back cleaner and quicker.
Launch-day posts on X kept returning to Sonnet 5.5 vs Opus 5.5, so that is the fair last word: the cheaper model is close, and the difference is yours to measure. Move like a fox: prowl the edges, test in the undergrowth, then pounce on the jobs that come back clean. Attribution is the discipline here, and folkfox's AI agent rumour story shows how to speak while a claim is reported but not confirmed.
The evidence ledger
Filter by what you need to know| Claim | Evidence status | What to do |
|---|
A good release changes the default question from “can it do this?” to “which part should we give it first?” Claude Sonnet 5.5 has earned that test, and Sonnet 5.5 vs Opus 5.5 is the test to run first.
How much of this is verified
The five claims in the ledger, by evidence
What should you know about Claude Sonnet 5.5?
Is Claude Sonnet 5.5 available now?
Yes. Anthropic announced it on 28 September 2026 for Claude platforms, its API and major cloud providers. The API model ID is claude-sonnet-5-5.
Did the per-token API price go down?
No. Anthropic lists the same $2 per million input tokens and $10 per million output tokens as Sonnet 5. It says fewer tokens per comparable task can lower total cost, up to 30 per cent in its tests.
Is Sonnet 5.5 better than Opus 5.5?
In Sonnet 5.5 vs Opus 5.5 terms, not across the board. Several published Sonnet 5.5 scores are close, but Anthropic says Opus 5.5 remains stronger on complex, open-ended work that requires sustained judgement.
What should a team test first?
Run a representative set of bounded code, document and visual tasks at the effort levels you actually use. Compare quality, elapsed time, token use and the work left for a person to correct.
What is Claude Sonnet 5.5 good for?
Anthropic positions it for focused coding and knowledge work at speed: bounded bug fixes, polished documents, slide decks, and reading screens or charts. Amazon Web Services describes it the same way. For open-ended work that needs sustained judgement, Anthropic still points to Opus 5.5.
How much does Claude Sonnet 5.5 cost?
The list price is $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20 per million, the same as Sonnet 5 according to Reuters and The Decoder. Anthropic says fewer tokens per task can lower the total bill by up to 30 per cent in its tests.
Sonnet 5.5 vs Opus 5.5: which should you use?
Start with Sonnet 5.5 for well-scoped daily work such as bounded bug fixes and polished documents, and with Opus 5.5 for open-ended tasks that need sustained judgement. Reuters puts Opus 5.5 at $4 per million input tokens and $20 per million output tokens, twice the Sonnet rate. Test both on your own tasks.
Which Sonnet 5.5 effort settings should you start with?
Test Xhigh before Max. In Anthropic's table Sonnet 5.5 scores 52.1 per cent on FrontierCode at Xhigh and 46.2 per cent at Max, and Anthropic says Max more often triggered a code-review step that timed out or edited outside the task. Compare lower settings on your own work too.
Keep reading.
Read more on this topic

Sources and reporting notes
All model scores in this edition come from Anthropic's release table unless otherwise stated. The linked announcement is the cited source. Posts quoted from X were read through the X API and X's public post data on 28 September 2026, which returned each post's text, time, links and media. A benchmark score is evidence about its test setting. It is not a forecast for every deployment.
- Anthropic, Introducing Claude Sonnet 5.5: release, performance table, prices, partner statements, safeguards and footnotes.
- Claude Sonnet 5.5 System Card: Anthropic's evaluation detail.
- Anthropic, migration guide: model ID, thinking and API changes.
- CursorBench 4.0: evaluation description and public leaderboard. Its leaderboard may update independently of Anthropic's launch table.
- Reuters, Anthropic rolls out second Claude 5.5 model: release date, unchanged list prices, and the Opus 5.5 prices.
- The Decoder, Sonnet 5.5 nearly matches Opus 5.5: cache-read price, the 30 per cent claims and the note that independent testing is pending.
- Amazon Web Services, Claude Sonnet 5.5 on AWS: Amazon Bedrock and Claude Platform on AWS availability.
- The Next Web, Sonnet 5.5 and the cyber limits: safeguards and reasoning-extraction classifiers.
- OfficeChai, Sonnet 5.5 benchmarks: the slide-deck test reported from Anthropic's partners.
- Anthropic, Introducing Claude Opus 5.5: the sibling launch of 22 September.
- folkfox, What did Dario Amodei see?: structural and editorial reference for this special report.
Three launch-table qualifications matter: the published Opus 5.5 Terminal-Bench score uses Xhigh effort; Sonnet 5.5's FrontierCode Xhigh score exceeds its Max score for reasons Anthropic discusses; Artificial Analysis ran two knowledge-work evaluations on a pre-release deployment with a structured-output bug that Anthropic says it has fixed. Anthropic also notes a recent GPT-6 Sol image-understanding fix that may not yet be reflected in some comparison scores.

