Claude Sonnet 5.5 vs Sonnet 5: Same price, big jump

Most model upgrades involve a tradeoff: better results for a higher price, or lower cost for weaker output. Claude Sonnet 5.5 is unusual. It has the same API price as Claude Sonnet 5, and it scores dramatically higher on nearly every measure in our benchmark.

In our Ready for Enterprise (RFE) LLM Benchmark, Sonnet 5.5 ranks #2 overall with a score of 63.2. Sonnet 5 ranks #31 with a score of 41.5 and falls below the benchmark's capability threshold. On Terminal-Bench 4.0, which measures agentic work in a terminal, Sonnet 5.5 scores 61.8% compared with 12.4% for Sonnet 5.

The short answer to Sonnet 5.5 vs Sonnet 5: upgrade. Anthropic calls the new model a clear upgrade, and our data agrees. It's one of the easiest same-price upgrades in our benchmark, though there are a few migration details to handle first.

What is Claude Sonnet 5.5?

Anthropic released Claude Sonnet 5.5 on September 28, 2026, as the second model in the Claude 5.5 family, six days after Claude Opus 5.5. Anthropic describes it as a faster, lower-cost complement to Opus 5.5. Where Claude Opus is built for complex work that requires careful judgment, the Sonnet model is strongest at well-scoped everyday tasks: fixing bugs and creating polished documents, slides, and spreadsheets. At launch, Anthropic said Claude Haiku 5.5, built for high-volume and cost-sensitive applications, would join the family in the coming weeks, which was proven correct. 

According to Anthropic, Sonnet 5.5 generates output more than 30% faster than Sonnet 5, making it the fastest Sonnet model to date, and costs up to 30% less for most work. The API price per token didn't change. The savings come from the model needing far fewer tokens to do the same work. Anthropic also says the model writes more clearly than the previous generation of Claude models, and its speed suits fast iteration on less complex tasks.

Sonnet 5.5 is available on the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry, as well as in the Claude apps and Claude Code. It has a 1 million token context window, up to 128,000 output tokens, and adaptive thinking on by default. Like other recent Claude models, it's available with zero data retention.

Sonnet 5.5 vs Sonnet 5 at a glance

Source: FullStack RFE LLM Benchmark, data snapshot used for this article. Rankings can change as new models and results are added.

Sonnet 5.5 leads on every benchmark and speed measure in the table, at the same price. Our RFE score combines six public benchmarks covering agentic work, factual accuracy, human preference, web development, graduate-level science, and advanced math, then adds price and speed. The capability score reflects benchmark performance alone.

What improved in Sonnet 5.5

Agentic coding

The biggest change is in agentic work. Sonnet 5.5 scores 61.8% on Terminal-Bench 4.0 in our snapshot, five times Sonnet 5's 12.4%. Anthropic's own launch results show a similar gap: it reports 70.6% for Sonnet 5.5 versus 10.3% for Sonnet 5. The numbers differ because Anthropic tests in its own setup, while our benchmark uses independent leaderboard results.

Anthropic also reports that at medium effort, the default in the Claude apps, Sonnet 5.5 far exceeds Sonnet 5's best Terminal-Bench score for less than a tenth of the cost per task. On FrontierCode, which measures whether an agent's code change would be merged, Sonnet 5.5 at high effort scores 10 points higher than Sonnet 5 at about one fifteenth of the cost per task.

Early testers saw the same pattern in real builds. Anthropic notes that Sonnet 5.5 batched tool calls together more often than Sonnet 5, leading to fewer steps and lower costs. Base44 reported that across 118 app builds, Sonnet 5.5 matched Opus 5's quality in an average of 3.6 iterations per build, compared with 7.7 for Opus 5.

Web development

On Arena WebDev, a head-to-head rating of web development output, Sonnet 5.5 scores 1774 Elo versus 1540 for Sonnet 5. That puts it close to Claude Opus 5.5's 1811 in the same benchmark.

Knowledge work, reasoning, and math

On GDPval-AA, a test of real-world professional work across 44 occupations that Artificial Analysis ran for the launch, Anthropic reports that Sonnet 5.5 scores 1844 Elo versus 1449 for Sonnet 5, nearly level with Opus 5.5's 1846. Slack reported that without changing any prompts, Sonnet 5.5 did better than Sonnet 5 on almost all of its offline Slackbot evaluations, with fewer steps and about 14% fewer output tokens, delivering better outcomes for users.

In our benchmark, Sonnet 5.5 also improved on advanced math, from 65.6% to 88.8% on FrontierMath Tiers 1–3, and on graduate-level science, from 90.5% to 95.6% on GPQA Diamond. Its GPQA Diamond score is higher than Opus 5.5's 90.6%.

Speed

In our measurements, Sonnet 5.5 generated 88 tokens per second versus 68 for Sonnet 5, about 29% faster. That lines up with Anthropic's claim of more than 30% faster output. Sonnet 5.5 also returned its first token sooner, at 1.49 seconds versus 2.20, so it wins on speed both ways.

Alignment and safety

On Anthropic's automated behavioral audit, which tests Claude across roughly 1,850 scenarios, Sonnet 5.5 improves on or matches Sonnet 5 on most measures of alignment, resistance to misuse, and honesty.

Where Sonnet 5.5 still has limits

Factual accuracy

Sonnet 5.5's weakest score is SimpleQA Verified at 46.5%. That's an improvement over Sonnet 5's 33.7%, but it's well below Opus 5.5's 72.2%. For workflows that depend on recalling facts without retrieval, ground the model with search or retrieval-augmented generation (RAG), or route those requests to a stronger model.

Cybersecurity requests

Because its cyber capabilities are comparable to Opus 5's, Sonnet 5.5 is the first Sonnet model to launch with safeguards similar to those on Anthropic's most capable models. Higher-risk cybersecurity tasks visibly fall back to Sonnet 5, while routine software development, like finding and fixing bugs, isn't affected. Security teams should test their real workloads before switching.

Sonnet 5.5 vs Opus 5.5

Sonnet 5.5's results raise a practical question: when do you still need Claude Opus 5.5? In our benchmark, Sonnet 5.5's RFE score (63.2) sits just behind Opus 5.5's (65.3), at half the blended price ($4.00 versus $8.00). It's also faster, with 88 tokens per second versus 79 and a much quicker first token.

In our benchmark, Opus 5.5 still leads on agentic coding (64.8% versus 61.8% on Terminal-Bench 4.0), web development, and factual accuracy. Anthropic’s own launch results rank them the other way on Terminal-Bench (70.6% for Sonnet 5.5 versus 66.4% for Opus 5.5), because the two use different test setups.

Anthropic also says that in its own testing and that of external testers, Opus 5.5 remains clearly stronger at complex, open-ended work that requires sustained judgment. For the hardest multistep work, large codebases, and fact-heavy tasks, Opus remains the safer choice. For well-scoped coding and everyday knowledge work, Sonnet 5.5 is now the better value. 

^{For a deeper look at how the two tiers split the work, see our comparison of Sonnet 5 vs. Opus: which model wins which job.}

Pricing: same rates, lower cost per task

Both models cost $2 per million input tokens and $10 per million output tokens, for a FullStack blended price of $4.00 per million tokens. Sonnet 5.5 does lower one rate: cache reads cost $0.10 per million tokens, half of Sonnet 5's $0.20, which helps workflows that reuse long system prompts. Check Anthropic's pricing page for current rates, including batch discounts.

Equal token rates don't mean equal bills. Because Sonnet 5.5 typically uses fewer tokens per task, it's less expensive to run. Some early testers reported large drops in token use: Balyasny Asset Management found Sonnet 5.5 used about 121,000 tokens per answer on its finance tasks, compared with 497,000 for Sonnet 5. Your actual savings depend on your prompts, effort settings, and how often the model needs retries or human correction. Measure cost per completed task, not cost per token.

Should you upgrade?

Yes, for nearly every team on Sonnet 5. At the same price, Sonnet 5.5 is faster, scores much higher on agentic coding and web development, and improves on every benchmark in our comparison. The only cases that call for extra testing are fact-heavy workflows without retrieval and security work that may hit the new safeguards.

Teams most likely to benefit include those running:

• Coding agents and Claude Code workflows
• Web app and front-end generation
• Document, slide, and spreadsheet work
• Computer use and long-horizon agent tasks
• High-volume tasks where faster responses improve throughput

How to migrate from Sonnet 5

Sonnet 5.5 isn’t a pure drop-in swap. Anthropic’s migration guide flags several changes that return errors for code already running on Sonnet 5. Check your requests for each of these:

1. Turning thinking off works differently. If you run Sonnet with thinking off, switch to the new “between tools” setting, which turns off up-front thinking. It works at high effort or below, so it can’t be combined with xhigh or max effort.
2. Thinking budgets are rejected. Requests that set budget_tokens return an error. Remove the setting and control reasoning with the effort level instead.
3. Sampling parameters are rejected. Requests that set temperature, top_p, or top_k return an error. Remove them from your requests.
4. Assistant prefill returns an error. Requests that pre-fill the start of the model’s reply aren’t supported. Put that guidance in the prompt or system message instead.
5. Forced tool use returns an error. Requests that force the model to call a specific tool aren’t supported. Let the model choose tools instead.
6. Thinking blocks are tied to the model and the conversation. They also can’t be separated from the account that created them, which matters if you move conversations between accounts.
7. The older computer use tool isn’t accepted on the Claude API and Google Cloud. Move to the newer computer use toolset.
8. The advisor tool has new pairings. It rejects Claude Opus 4.8, Claude Opus 4.7, Claude Sonnet 5, and Claude Haiku 5.5 as advisors.

One more change doesn't cause errors but alters what users see: text that appears between tool calls now comes back inside thinking blocks. Apps that stream that text may go quiet between tool calls until they're updated.

Effort defaults also differ by surface. Sonnet 5.5 defaults to medium effort in Claude Code and the Claude apps, and high effort on the Claude Platform. At lower settings, Claude answers faster and uses fewer tokens, which suits routine work. At higher settings, it reasons longer and checks its work more thoroughly. Set the effort level explicitly in API requests so your comparisons are like-for-like.

A practical migration plan:

1. Capture a Sonnet 5 baseline: prompts, tools, effort settings, and acceptance criteria.
2. Fix any request settings Sonnet 5.5 no longer supports.
3. Test both models with identical prompts, tools, and evaluation sets, and compare completion rate, tokens, latency, and cost per accepted result.
4. Route a small share of traffic to Sonnet 5.5, then expand once results meet your thresholds.

How FullStack can help

FullStack is an AI-native engineering partner and is Anthropic Certified through Anthropic's Claude Partner Network. We build on Claude every day, from our own engineering workflows to the systems we ship for clients. Our Claude implementation services help enterprises evaluate Claude models on their real workloads, migrate between model versions without surprises, and build multi-model routing that sends each task to the model that handles it best.

^{Compare these models in the RFE benchmark.}