SPIRITT×
Claude Sonnet 5.5Fix the bug. Refine the design. Finish the deck. Claude Sonnet 5.5 focuses on well-scoped everyday work, with Anthropic reporting 30%+ faster output generation and up to 30% lower task costs than Sonnet 5.
Explore interactive builds, coding demos, and early evaluations of Sonnet 5.5, alongside Anthropic’s official launch.
Introducing Claude Sonnet 5.5, the second model in the Claude 5.5 family. It’s a clear upgrade over Sonnet 5, runs more than 30% faster, and costs up to 30% less for most work.
— Claude (@claudeai) 2026-09-28
i’ve been testing claude sonnet 5.5 over the past week. here are some experiments i tried .. project 1: a forest made of javascript i wanted to recreate the feeling of sitting alone in the sun under a canopy of leaves. this scene was made with 3000 lines of js / webgl shaders
— AA (@measure_plan) 2026-09-28
I got early access to Claude Sonnet 5.5. One test was an F-Zero-style racing game in Unreal. This is STARLANE. I also tested an RTS, a paint shooter, animation, and Sonnet 5.5 vs 5 on the same briefs. Demos below.
— Future Brian (@ForwardEditor) 2026-09-28
i tested sonnet 5.5 on "code-to-painting". every pixel is generated by python the model wrote. it's shown a photo, writes a brush engine, renders, looks, and revises. you can see the large step-up from sonnet 5 in terms of coding + visual reasoning.
— Lance Martin (@RLanceMartin) 2026-09-28
Sonnet 5.5 fixing a bug with Claude Code.
— Boris Cherny (@bcherny) 2026-09-28
One of my early access tests with Sonnet 5.5 was to make a horror game that would genuinely scare me... I do not do well with horror games... Let's see how it goes
— Alex (@The_Alex) 2026-09-28
Claude Sonnet 5.5 drew every frame of this animation in JavaScript. The history of dinosaurs in under one minute.
— Kevin Ngo (@kevin_t_ngo) 2026-09-28
Sonnet 5.5 is priced the same as Sonnet 5, but it typically needs far fewer tokens to do the same work. In our testing, it costs up to 30% less per task than its predecessor. It also generates output more than 30% faster, making it our fastest Sonnet model to date.
— Claude (@claudeai) 2026-09-28
Sonnet 5.5 is out today. I asked it for a murmuration of 400 starlings, and this is what it built- one HTML file, no libraries 🙀
— Dheepan Ratnam (@Dheepanratnam) 2026-09-28
Claude sonnet 5.5 design this in one single prompt Thought? 🤔
— Tushar Pandey (@TUSHARPAND0848) 2026-09-28
asked sonnet 5.5 to make something cool And it made A lighthouse on a floating island, with aurora, a waterfall, and a moon. All written in code.
— agim𓍯 (@aGim_asf) 2026-09-28
Built this using Sonnet 5.5. Daym, the output is really good and very close to Opus 5.5, especially the music sync. The stamps, paper flips, tears and transitions all land really nicely on the beat. It needs 2-3 more iterations to make this perfect.
— Pankaj Kumar (@pankajkumar_dev) 2026-09-28
Here it rakes a yard of leaves using tool calls. 24s faster than Sonnet 5, with 6K fewer tokens. It leaves her extra time to read :)
— cat (@_catwu) 2026-09-28
Coming soon to Harvey: Sonnet 5.5. Early testing found gains over Sonnet 5 in document-heavy legal work like funds and asset management, emerging companies and venture capital, and data privacy and cybersecurity.
— Harvey (@harvey) 2026-09-28
Made with claude sonnet 5.5 in 5 prompts and 2 hours
— Lax Acharya (@laxachx) 2026-09-28
Holy macaroni! Claude Sonnet 5.5 is INSANE but see it yourself. It built a Video to introduce himself! I love the creativity
— Justin (@JustinGorya) 2026-09-28
Sonnet 5.5 improves on Sonnet 5 across benchmarks, in some cases dramatically. It’s a faster, lower-cost complement to Claude Opus 5.5, strongest at well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets.
— Claude (@claudeai) 2026-09-28
On several benchmarks, Sonnet 5.5 at Low or Medium effort beats Sonnet 5’s best score for about a tenth of the cost.
— Claude (@claudeai) 2026-09-28
Sonnet 5.5 is built for focused coding, fast iteration, and polished documents, slides, and spreadsheets. Keep Opus 5.5 for the most complex work that demands sustained judgment.
Anthropic introduced Claude Sonnet 5.5 on September 28, 2026. It supports text and image input, text output, a one-million-token context window, and up to 128,000 output tokens in standard requests. Adaptive reasoning offers five effort levels, from low to max.
Standard API rates are $2 per million input tokens, $10 per million output tokens, and $0.20 per million cache-read tokens. These are unchanged from Sonnet 5. Anthropic’s claimed savings come from using fewer tokens on typical work, not a token-price cut.
At max effort, Sonnet 5.5 scores 56 on Artificial Analysis’s Intelligence Index v4.3.2, close to Opus 5.5’s 58. But max is not always the economical setting: AA estimates $7.60 per Index task for Sonnet versus $5.98 for Opus at max. Match the effort to the job.
Anthropic reports major coding gains and stronger visual understanding, while noting that Opus 5.5 remains better at complex, open-ended work. Model access in SPIRITT depends on the models and connections enabled for your workspace.
Index v4.3.2, September 28, 2026. Selected configurations; rounded index points, not percentages. Claude uses adaptive reasoning; Sonnet 5.5, Opus 5.5 and Fable 5.1 use Default Fallback. Sonnet max and Opus xhigh both round to 56; their tiny unrounded gap does not establish a meaningful difference.
Selected coding, knowledge-work, visual-reasoning, and automation results, plus the cost of reaching them. Independent Artificial Analysis runs stay separate from Anthropic’s reported evaluations.
Anthropic’s Claude Code bare-mode run: 66 tasks, cached dependencies, no internet egress. Sonnet and Opus 5.5 used five trials per task with safeguards; fallback affected 1.2% and 2.5% of requests respectively. Sonnet is max; Opus is xhigh. Other comparators’ repeat counts differ. Separate from AA’s run.
Artificial Analysis: mini-swe-agent, 66 tasks, three repeats, pass@1, 500-step cap. All rows use max effort; Claude 5.5 configurations and Fable use Default Fallback. These results must not be substituted for Anthropic’s Claude Code run.
Cognition’s 150-task evaluation, reported by Anthropic, using Claude Code and Codex CLI. Composite scores, not simple test pass rates. Sonnet scores lower at max than xhigh. In two cases Cognition examined, extra code-review agents caused a timeout or out-of-scope edits.
Cursor’s production-agent evaluation, reported by Anthropic. Max-effort scores shown. Sonnet’s reported task costs were estimated from Cursor’s token counts; no uncertainty intervals are supplied.
Artificial Analysis, Stirrup, 220 tasks. Rounded Elo, not percent correct; Index values may differ from later live leaderboards. Max effort; Claude 5.5 and Fable use Default Fallback. Anthropic reports a since-fixed structured-output bug in the prerelease Sonnet deployment used here; these are not confirmed post-fix reruns.
Artificial Analysis: 91 professional-work tasks using Stirrup. Rounded Elo combines rubric, analysis and presentation quality, not completion rate. Max effort; Claude 5.5 and Fable use Default Fallback. The since-fixed prerelease Sonnet structured-output bug also affected this run; a fix does not establish a rerun.
Anthropic’s 2,500-question evaluation, five trials at max effort, with web search, fetch and code execution. This is the tool-assisted version, not Artificial Analysis’s separately configured HLE.
Anthropic: 100 visual-reasoning tasks, five runs at max effort, image cropping and code tools, Gemini 3.5 Flash grader. These are point estimates; the close Sonnet/Opus scores do not establish a decisive lead.
Anthropic: 108 computer-use tasks, five runs, max effort, 500-action limit and September 10 task/assets versions. Partial credit weights completed checkpoints. An 80.1% score does not mean 80.1% of tasks were fully completed.
Same Anthropic OSWorld 2.1 setup as the partial-credit chart. Strict pass requires every checkpoint to be satisfied. Keeping the two scores separate shows the gap between partial progress and a fully completed task.
Zapier v1.0.6, reported by Anthropic. Max-effort task pass rates require all assertions to pass. Sonnet 5.5 used default fallbacks; Opus’s 42.5% includes selective fallback-enabled reruns of refusal tasks (40.0% before reruns). Other rows’ fallback settings are not established here. Not a uniform rerun or AA’s objective-completion metric.
Artificial Analysis, dataset v1.0.6: 657 private tasks, 50-turn cap. This measures the mean share of guardrail-adjusted objectives completed, not fully completed tasks. Max effort; Sonnet and Opus use Default Fallback.
Anthropic’s internal harness, June 2026 final release: 108 tasks, three trials, adaptive max effort. Pass³ requires all three trials to succeed. Sonnet trails Opus and Fable. Classifiers were on for those three models; two blocked Sonnet trials count as failures. Not the public leaderboard setup.
Artificial Analysis Intelligence Index v4.3.2, September 28. Rounded index points across five adaptive-reasoning settings, all with Default Fallback. Max raises the score, but the cost chart shows why it is not an automatic everyday default.
Artificial Analysis’s weighted mean Index task-cost estimate, based on token usage, list prices and representative cache-hit rates. Same effort configurations as the score chart. These are evaluation costs, not a price for every real-world task.
Artificial Analysis, all at max effort; Claude 5.5 and Fable use Default Fallback. Sonnet’s weighted mean is $7.60 per Index task versus Opus’s $5.98 despite lower token rates. These averages include unsuccessful attempts, not costs per successful completion. Lower is better.
Anthropic standard USD rates, September 28, 2026. Sonnet 5.5 keeps Sonnet 5’s token prices; task savings come from usage efficiency. Batch discounts, caching, service modifiers and tool charges are separate. Not SPIRITT subscription prices.
Anthropic standard USD rates, September 28, 2026. Sonnet 5.5 keeps Sonnet 5’s token prices; task savings come from usage efficiency. Batch discounts, caching, service modifiers and tool charges are separate. Not SPIRITT subscription prices.
Sources: Anthropic’s September 28 launch, system card and API documentation; Artificial Analysis model data and Intelligence Index v4.3.2, checked September 28, 2026. Benchmark versions, reasoning settings, fallbacks and scoring differ. Percentages are rounded to one decimal; Elo and Index values are rounded to integers; cost estimates to cents. Close scores do not establish statistically meaningful superiority. AA marks CritPt, 10% of this Index, under review; these are its published scores, not a recalculated index. Effort names do not equalize compute across models. API list prices and evaluation task costs are not SPIRITT plan prices or universal savings guarantees.
Bring a clear goal, use the model access available in your workspace, and judge the finished result.
Start with a bug to fix, a page to improve, or a document to finish. Give SPIRITT the relevant files and a clear definition of what a good result looks like.
Check whether Sonnet 5.5 is available through your enabled models or connections. Access varies by workspace. Where it is enabled, start with a focused task and compare the result against your usual model.
Try a representative task before scaling up. Compare quality, total cost and completion time, then raise the effort where the result benefits. For complex open-ended judgment, compare against Opus 5.5.
Bring your goal into SPIRITT and work with the models available in your workspace. Try Sonnet 5.5 on the tasks that matter to you wherever access is enabled.