Claude Sonnet 5.5 Launch: Coding Performance on Par with Opus 5.5 at Half the Price

Anthropic has released Claude Sonnet 5.5: an upgrade to Sonnet 5 at the same price, and a faster, less expensive counterpart to Opus 5.5. It addresses a practical problem. A model may be able to fix a bug, edit several files, or prepare a presentation, but repeated attempts, tool calls, and lengthy intermediate output increase both the wait and the total cost of the task. Sonnet 5.5 aims to complete well-defined everyday work faster, while leaving open-ended tasks that demand sustained judgment to Opus.

Four features are worth noting first:

• Finish the same task faster. Sonnet 5.5 has the same API token prices as Sonnet 5. Anthropic says its output is more than 30% faster, it typically uses fewer tokens, and in its tests it costs up to 30% less per task.
• Handle a continuous coding workflow better. The gains extend beyond code completion to terminal operations, edits across multiple files, tool use, and iteration.
• Produce work closer to a deliverable. It can use source materials and templates to create documents, slides, and spreadsheets. On some knowledge-work evaluations, its scores are close to Opus 5.5.
• Match effort to the task. Effort settings let you trade off speed, cost, and quality. Simple tasks need not always run at Max.

Here is how those four points translate into speed, suitable work, and practical settings.

Same token prices: why might the whole task cost less?

Two kinds of price are easy to confuse. Sonnet 5.5 costs $2 per million input tokens, $10 per million output tokens, $0.20 for cache reads, and $2.50 for cache writes; Sonnet 5 has the same prices. Compared with Opus 5.5, its input, output, and cache-write prices are half as high, while cache reads cost the same $0.20.

But token prices are only half the story. If two models complete the same fix and one searches repeatedly, rewrites its work, and calls more tools, that model can still cost more overall. Anthropic's cost-versus-score charts therefore put actual cost per task on the horizontal axis and score on the vertical axis. Low, Medium, High, Xhigh, and Max effort appear as separate points for a given model. On a particular evaluation, a point farther up and left means a higher score for less money.

On several evaluations, Sonnet 5.5 at Low or Medium effort exceeded Sonnet 5's best score at roughly one-tenth of the cost per task. At higher effort, its scores sometimes approached Opus 5.5, but its costs approached Opus too. The upgrade is not merely a smarter model: routine tasks may reach usable quality with less work. The one-tenth figure comes from those evaluations, not a universal discount on every bill.

Coding: beyond completing a line, toward finishing the workflow

The most striking number for Sonnet 5.5 is 70.6% on Terminal-Bench 4.0, versus 10.3% for Sonnet 5. This benchmark asks agents to perform multistep professional tasks in a terminal. It measures a chain of reading the environment, using tools, making changes, running checks, and finishing—not simply completing one line of code. The release table gives Opus 5.5 a score of 66.4%; its footnote specifies Xhigh effort for that Opus result. Sonnet's lead on this benchmark does not mean it is stronger than Opus on every complex coding project.

Two other coding benchmarks cover different work. CursorBench 4.0 draws on ambiguous, multi-file tasks from real Cursor sessions: Sonnet 5.5 scores 55.5%, Sonnet 5 scores 34.1%, and Opus 5.5 scores 57.8%. FrontierCode measures whether a code change can be merged without human editing. At Xhigh, Sonnet 5.5 scores 52.1%, versus 54.4% for Opus 5.5. Together, these tests show substantial progress across understanding a project, using tools, changing files, and delivering a code change.

Product tests give a more concrete picture of how that ability might be used. Across 118 app builds, Base44 reported that Sonnet 5.5's apps scored on par with Opus 5's while requiring an average of 3.6 iterations rather than 7.7. Unity's test reopened projects and checked results at runtime; it reported that Sonnet 5.5 completed 90% of its multistep editor and coding tasks. Epic Games described game-system architecture and tasks lasting hours, while CodeRabbit plans to migrate simpler and medium-difficulty code reviews first. These are each company's own tests or plans, not a single shared leaderboard.

A short ClaudeDevs video offers a smaller example. One coordinator sends four agents from different starting points to solve a puzzle in parallel. The clip labels the figures as the median of 10 runs: Sonnet 5.5 finishes in 10 seconds using about 23,000 tokens and $0.11; Sonnet 5 takes 15 seconds, about 28,000 tokens, and $0.15. The difference is not just how quickly the model answers, but how many steps the whole tool-using process takes.

Documents, slides, and spreadsheets: from source materials to a finished file

Anthropic also emphasizes knowledge work. The goal is not merely to answer “What does this earnings report say?” but to read the materials, organize their main points, and produce a usable file in the requested format. GDPval-AA covers tasks across 44 occupations and nine industries. Sonnet 5.5 scores 1844, Opus 5.5 scores 1846, and Sonnet 5 scores 1449. On the longer-horizon AA-Briefcase v1.1 evaluation, their respective scores are 1811, 1822, and 1359. On these tasks, Sonnet 5.5 is very close to Opus 5.5.

The release page gives a more tangible example: the model received a public company's quarterly earnings materials, a conference-call transcript, and a slide template, then was asked to make a 10-slide operating review. Two experts Anthropic consulted thought the first draft was ready to send. The point is not merely that it can “write slides”; it can connect research, synthesis, and adherence to a template in one task. Claude's official thread also shows a side-by-side sand-dune webpage demonstration, making its visual output easier to inspect.

Teams also care about how many steps it takes to reach a comparable result. Slack says that, without changing its prompts, Sonnet 5.5 performed better on nearly all of its offline Slackbot tests while producing about 14% fewer output tokens. Zendesk reports 20% faster ticket handling across hundreds of reply and escalation cases. In a private set of 2,441 finance tasks, Balyasny Asset Management reports about 121,000 tokens per answer, versus roughly 497,000 for Sonnet 5. Lovable's coding evaluation recorded about one-third fewer tool calls and roughly half as many shell runs. Box reports 2.4-times processing speed and 12% fewer total tokens; Atlassian expects up to 30% faster Rovo Agents. These numbers come from different companies and tasks. They suggest measuring cost per completed item in your own workflow; they cannot be added together or treated as mutual replications.

Other tasks improve too. Sonnet 5.5 scores 64.5% on Humanity's Last Exam with tools, 80.1% on a subset of OSWorld 2.1, and 61.6% on Chartography without tools. The corresponding Sonnet 5 scores are 54.9%, 57.0%, and 15.6%. These results point to broader reasoning, computer-use, and chart-reading abilities, but they are different tests and cannot be combined into one overall score.

Effort: start with the task, not with Max by habit

Effort determines how much reasoning and checking a model devotes to a task. Claude Code and the Claude app default to Medium; Claude Platform defaults to High. Lower settings are usually faster and cheaper. Higher settings can spend more time checking difficult work, but can also raise the total cost per task. The four official cost-versus-score charts illustrate this: the blue Sonnet 5.5 points beat many older-model results at lower settings, while both score and cost move closer to Opus as effort rises.

FrontierCode supplies a useful exception. Sonnet 5.5 scores 52.1% at Xhigh but falls to 46.2% at Max. Anthropic explains that Max more often triggers multi-agent code review; in two examined cases, that led to timeouts or changes outside the task's scope. For work with clear acceptance criteria, such as fixing a bug or editing files, it makes sense to test at Medium or below first and increase effort only if the result requires it. That better matches the cost advantage shown in this release than starting every task at Max.

See the four official cost-versus-score charts: knowledge work, CursorBench, FrontierCode, and Terminal-Bench

What changed in two generations of webpages built from the same prompts?

On the release page, Sonnet 5 and 5.5 each build three visual projects as a single HTML file. The static images show where each was at the captured moment; the original interactive programs below show the code-generation and running process.

A flock of 400 starlings

The Sonnet 5 screenshot still shows code being written at 4,520 tokens. After outputting 4,158 tokens, Sonnet 5.5 has a running flock on screen.

Dunes shaped by wind

The older model's screenshot still shows code being written at 2,966 tokens. The newer model has produced layered dune lines and a red sun after outputting 2,720 tokens.

A large clock made of 24 small clocks

The difference is especially visible here: Sonnet 5 is still displaying code at 27,581 tokens, while Sonnet 5.5 has output 14,386 tokens and the hands of its small clocks form large digits on screen.

Expand the six original interactive programs to see the full process. A screenshot taken while the older model was writing code does not mean it could never finish the project; these images capture the progress of the official demos at specific moments.

Other settings and limits to check before using it

Sonnet 5.5 also changes some safety handling. Across roughly 1,850 automated behavior scenarios, Anthropic reports that it matched or improved on Sonnet 5 in most alignment, honesty, and abuse-resistance measures; Opus 5.5's overall results were still slightly better. This indicates that the new version passed a more demanding release review, but real workflows still need ordinary permissions and acceptance checks.

Three changes may matter in practice. Ordinary bug finding and fixing is not affected by the cybersecurity fallback, while higher-risk cybersecurity requests may be routed to Sonnet 5. Biological safeguards remain in place, and a small number of legitimate requests may be blocked in error. When a conversation moves across accounts, expanded preserved thinking requires the reasoning content to be regenerated under the new account. The first two cases may rarely affect ordinary coding or document work; security researchers and teams moving conversations between organizations should test them in advance.

Which tasks should try Sonnet 5.5 first?

Start with tasks that have clear inputs and acceptance criteria: fixing a reproducible bug, preparing an operating review from a template, or processing customer-support tickets in batches. Run a set of real samples with Sonnet 5.5 at Medium or lower effort. Record completion rate, rework time, tool calls, elapsed time, and final cost, then compare the results with Sonnet 5 or Opus. If the goal keeps changing and the work demands long-term trade-offs, Opus 5.5 remains Anthropic's recommended starting point.

Developers can use model ID claude-sonnet-5-5 on Claude Platform, as well as through AWS, Google Cloud, and Microsoft Azure; Anthropic says it supports zero data retention. Teams that previously disabled thinking for Sonnet also need to use the between_tools setting when migrating to 5.5, following the official migration guide. The release page says Haiku 5.5, intended for high-throughput, cost-sensitive applications, will join the family later; there are not yet results here on which to base a choice.

Sonnet 5.5's value is not that it wins every leaderboard. It is that more everyday coding and office tasks may reach a deliverable result without calling on Opus. The most useful test is not a single benchmark score, but how much waiting, rework, and money your own work actually saves.