Claude Sonnet 5.5 Complete Guide 2026: Pricing, Benchmarks, API

Anthropic shipped Claude Sonnet 5.5 on September 28, 2026, six days after Claude Opus 5.5 opened the Claude 5.5 family. The release notes describe it as "a faster, lower-cost complement to Claude Opus 5.5." Pricing did not move from Claude Sonnet 5: still $2 in and $10 out per million tokens. The benchmark numbers moved a lot, though. Terminal-Bench 4.0 went from 10.3% to 70.6%, and on several agentic and office-work evals Sonnet 5.5 is now close to Opus 5.5.

If you searched "sonnet 5.5", you probably want three things: what it costs, how it compares to Opus 5.5 and GPT-6 Sol, and whether switching your code from claude-sonnet-5 will break anything. This guide covers all three. It has the full spec table, both Anthropic's benchmarks and Artificial Analysis's independent numbers, a working Python quickstart, the five breaking API changes, and a straight recommendation on when to use Sonnet 5.5 and when to pay for Opus.

What is Claude Sonnet 5.5?

Claude Sonnet 5.5 is the Sonnet-tier model in Anthropic's Claude 5.5 family. Anthropic's model docs describe it as "the best combination of speed and intelligence" in the current lineup. It sits between Claude Opus 5.5 ($4/$20) and Claude Haiku 4.5 ($1/$5), and its latency is rated "Fast". Opus 5.5 is rated "Moderate".

Anthropic's product page says it is aimed at "well-scoped everyday tasks, fixing bugs, and creating polished documents, slides, and spreadsheets," and it also calls out long-horizon work, image understanding, and agentic coding. The system card goes further. It says Sonnet 5.5 significantly outperforms Sonnet 5 across many domains and "in a few areas, it rivals or exceeds Claude Opus 5.5." The same card says it is "broadly less capable than Opus 5.5 across domains" and does not cross any new Responsible Scaling Policy thresholds.

Key facts at a glance

• Release date: September 28, 2026 (Anthropic release notes, model docs, and system card all agree.)
• Family: second model in Claude 5.5, after Opus 5.5 (September 22, 2026)
• Knowledge cutoff: June 2026 (reliable and training-data cutoff)
• Thinking: adaptive thinking, on by default, steered by the effort parameter (default high on the API)
• Retirement: not sooner than September 28, 2027

Claude Sonnet 5.5 specs

All values below come from Anthropic's Sonnet 5.5 model page and models overview.

Two details worth knowing. The 300K batch output needs the output-300k-2026-03-24 beta header. The minimum cacheable prompt drops to 512 tokens, from 1,024 on Sonnet 5, which makes prompt caching useful on shorter system prompts. The tokenizer is the same as Sonnet 5's, so identical text produces identical token counts.

How much does Claude Sonnet 5.5 cost?

Sonnet 5.5 costs exactly what Sonnet 5 costs. Anthropic's pricing page lists these rates:

The full 1M context is billed at the standard rate, so a 900K-token request costs the same per token as a 9K-token one. Pinning inference to the US with inference_geo: "us" adds a 1.1x multiplier. Bedrock and Google Cloud regional endpoints carry a 10% premium over global endpoints. Batch and caching discounts stack.

One interesting quirk: cache reads cost $0.20 per million on both Sonnet 5.5 and Opus 5.5, because Opus 5.5 uses a 0.05x cache multiplier instead of the usual 0.1x. For cache-heavy agent loops, the gap between the two models is mostly on uncached input and output.

Is Sonnet 5.5 actually cheaper per task?

This is where the sources disagree, and it matters for your bill.

• Anthropic says Sonnet 5.5 generates output more than 30% faster than Sonnet 5 and is "up to 30% cheaper" per task, because it needs fewer tokens and batches tool calls better.
• Artificial Analysis ran its Intelligence Index at max effort. It measured about 193K output tokens per task, the highest it has recorded and roughly 60% more than Opus 5.5 or Sonnet 5 at max. Its cost to run was about $7.60 per task, around 50% more than Sonnet 5.

Both can be true. Anthropic's claim is about typical settings. AA's number is about max effort, where Sonnet 5.5 thinks a lot. So set effort on purpose and don't leave it on max. Anthropic's own guidance is to start at medium for well-specified agentic coding and at medium or low for chat.

Claude Sonnet 5.5 benchmarks

The first table is Anthropic's launch table, from the Sonnet 5.5 product page and the system card (Table 8.1.A). These are vendor-reported numbers. The system card says the Claude results use adaptive thinking at max effort, averaged over 5 trials, and that competitor figures come from those developers' published system cards or leaderboards.

*Anthropic footnotes the Opus 5.5 Terminal-Bench score as its best result, at xhigh effort. Anthropic also notes that Sonnet 5.5's FrontierCode score is at max effort and is lower than its xhigh score. "n/a" means the vendor did not publish a comparable number.

About SWE-Bench Pro: third-party coverage quoted 81.3%, and it is not in the product-page table. It is in Anthropic's official system card, both in the capability summary table and in section 8.2 ("Sonnet 5.5 achieved 81.3%"). So it is a primary-source number. It just wasn't headlined.

What independent testing says (Artificial Analysis)

AA's numbers mostly back up Anthropic's story. Sonnet 5.5 comes in +18 Index points over Sonnet 5 and just 2 points behind Opus 5.5. AA's own Terminal-Bench run gives 64%, lower than Anthropic's 70.6%, but it still puts Sonnet 5.5 ahead of Opus 5.5 on that test. The trade-off is factual recall: Sonnet 5.5 answers fewer knowledge questions correctly than Opus (54% vs 66%), though it hallucinates less often when it doesn't know (47% vs 59%).

Claude Sonnet 5.5 vs Opus 5.5: which should you use?

On agentic and office-work evals, Sonnet 5.5 is within about 2 points on OSWorld, GDPval-AA, and CursorBench and ahead on Terminal-Bench, at half the per-token price. On harder software engineering (SWE-Bench Pro 81.3 vs 89.9, FrontierCode 46.2 at max effort or 52.1 at xhigh vs 54.4) and raw knowledge (HLE without tools 56.9 vs 64.4), Opus 5.5 is clearly ahead. We break this down in more detail in Claude Sonnet 5.5 vs Opus 5.5. For how the previous generation split, see Claude Opus 5 vs Sonnet 5.

And vs GPT-6 Sol?

Anthropic's table only has shared numbers on four evals. Sonnet 5.5 leads GPT-6 Sol on GDPval-AA (1844 vs 1487), AA-Briefcase (1811 vs 1483), and Chartography (61.6% vs 53.6%), and trails on FrontierCode (46.2% vs 49.3%). Anthropic footnotes that an OpenAI bug fix may not be reflected in some GPT-6 Sol scores. Those are vendor numbers from a competitor, so read them with that in mind. OpenAI replaced GPT-6 Sol with GPT-6.1 Sol on September 29; on Artificial Analysis's independent index Sonnet 5.5 (max) scores 56 vs 52 for GPT-6.1 Sol (max), but costs $7.60 per task vs $0.72. Our Sonnet 5.5 vs GPT-6.1 Sol comparison goes into this, and the GPT-6 Sol guide covers OpenAI's side.

Where can you use Claude Sonnet 5.5?

• Claude apps (claude.ai, desktop, mobile): available from launch day, per Anthropic's release notes. In the apps the default effort is medium, according to the product page.
• Claude API: all customers, model claude-sonnet-5-5.
• Cloud platforms: Amazon Bedrock (anthropic.claude-sonnet-5-5), Claude Platform on AWS, Google Cloud, and Microsoft Foundry (all claude-sonnet-5-5).
• Claude Code: the 2.1.284 changelog (September 28) made Sonnet 5.5 the default Sonnet model.
• GitHub Copilot: generally available for Copilot Pro, Pro+, Max, Business, and Enterprise, in VS Code, JetBrains, Copilot CLI, the coding agent, and more (GitHub changelog, September 28).
• Cursor: Cursor announced Sonnet 5.5 support, saying it performs "on par with Opus in many tasks."

How to use the Claude Sonnet 5.5 API (Python quickstart)

Install the SDK and set your key:

pip install anthropic
export ANTHROPIC_API_KEY="sk-ant-..."

A basic Messages API call, with effort set explicitly. The output_config.effort field follows Anthropic's effort docs:

import anthropic

client = anthropic.Anthropic()  # reads ANTHROPIC_API_KEY

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=16000,
    messages=[
        {"role": "user", "content": "Review this function for edge cases: def div(a, b): return a / b"}
    ],
    output_config={"effort": "medium"},  # low | medium | high (default) | xhigh | max
)

for block in response.content:
    if block.type == "text":
        print(block.text)

Thinking counts toward max_tokens even when the thinking text is not returned, so leave room. For agentic coding, Anthropic recommends max_tokens=128000 with streaming. Also note that setting temperature, top_p, or top_k to a non-default value returns a 400 error on this model.

To turn off up-front thinking for fast, simple calls, use between_tools. disabled is rejected on Sonnet 5.5, and between_tools only works at high effort or below:

response = client.messages.create(
    model="claude-sonnet-5-5",
    max_tokens=4096,
    thinking={"type": "between_tools"},
    output_config={"effort": "low"},
    messages=[{"role": "user", "content": "Classify this ticket: 'Login page 500s on Safari'"}],
)

What breaks when you migrate from Sonnet 5 to Sonnet 5.5?

Changing claude-sonnet-5 to claude-sonnet-5-5 is the easy part. Anthropic's "What's new" page lists five breaking changes:

1. thinking: {"type": "disabled"} returns a 400. Use {"type": "between_tools"} instead, at low, medium, or high effort.
2. Forced tool use is gone. tool_choice of any or tool returns a 400. Use auto plus strict: true tool definitions, or structured outputs.
3. Thinking blocks are tied to the model and the conversation. For accounts created on or after August 31, 2026, replaying a Sonnet 5.5 thinking block after editing the system prompt, tools, or earlier messages returns a 400. Keep history append-only.
4. Computer use needs the new toolset (computer_toolset_20260801) on the Claude API and Google Cloud. computer_20251124 is rejected there. Bedrock still accepts it.
5. Advisor tool pairings: Opus 4.8, Opus 4.7, and Sonnet 5 can no longer be advisors for a Sonnet 5.5 executor.

There's also one silent change. Longer text between tool calls now comes back as thinking blocks, and at the default display setting those blocks are empty. A UI that streams those notes will go quiet with no error. Effort levels were also recalibrated, so re-run your effort sweep instead of reusing the Sonnet 5 setting. The migration pattern is similar to the one in our Opus 5.5 migration guide.

Who should use Claude Sonnet 5.5?

• Default for most production agents and coding assistants. Within about 2 points of Opus 5.5 on OSWorld, GDPval-AA, and CursorBench, ahead on Terminal-Bench, at half the token price. Start here and only move up when your evals say so.
• Teams on Sonnet 5 today: upgrade. Same price, large gains on every benchmark Anthropic published. Budget an afternoon for the breaking changes, especially tool_choice and disabled thinking.
• Hard, multi-file software engineering: test Opus 5.5 too. The SWE-Bench Pro gap (81.3 vs 89.9) is real. The FrontierCode gap is 8.2 points at Sonnet's max effort (46.2 vs 54.4) but only 2.3 at its better xhigh score (52.1).
• Knowledge-heavy Q&A without retrieval: Opus 5.5 recalls more facts (AA-Omniscience 66% vs 54%). If you use RAG, this matters less, and Sonnet's lower hallucination rate helps.
• High-volume, simple tasks: run Sonnet 5.5 at low effort with between_tools, or look at Haiku 4.5 at $1/$5.
• Cost-sensitive workloads: avoid max effort unless you've measured a gain. AA's 193K-tokens-per-task figure shows how fast it adds up.

FAQ

When was Claude Sonnet 5.5 released?

September 28, 2026, according to Anthropic's release notes, model docs, and system card. It came six days after Claude Opus 5.5 (September 22, 2026).

How much does Claude Sonnet 5.5 cost?

$2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Cache reads are $0.20, 5-minute cache writes $2.50, 1-hour cache writes $4, and the Batch API halves prices to $1/$5.

What is the Claude Sonnet 5.5 API model ID?

claude-sonnet-5-5 on the Claude API, Google Cloud, Microsoft Foundry, and Claude Platform on AWS. On Amazon Bedrock it is anthropic.claude-sonnet-5-5.

What is Claude Sonnet 5.5's context window?

1M tokens of input, with up to 128K output tokens on the synchronous Messages API and up to 300K on the Batch API with the output-300k-2026-03-24 beta header.

Is Claude Sonnet 5.5 better than Opus 5.5?

Not overall. Artificial Analysis scores it 56 on its Intelligence Index, behind Opus 5.5 at 58, and Opus leads on SWE-Bench Pro, FrontierCode, and HLE. Sonnet 5.5 does beat Opus on Terminal-Bench 4.0, AutomationBench, and HealthBench Professional in Anthropic's numbers, and it costs half as much.

Can I turn off thinking on Sonnet 5.5?

Not with disabled, which returns a 400 error. The lowest setting is thinking: {"type": "between_tools"}, which turns off up-front thinking and works at low, medium, or high effort.

Is Sonnet 5.5 available in Claude Code, Copilot, and Cursor?

Yes. Claude Code 2.1.284 made it the default Sonnet model, GitHub Copilot made it generally available on paid plans, and Cursor added it on launch day.

Sources

• Anthropic: Claude Sonnet 5.5 product page (benchmarks, pricing, speed claims)
• Claude Docs: Claude Sonnet 5.5 model page
• Claude Docs: What's new in Claude Sonnet 5.5
• Claude Docs: Models overview
• Claude Docs: Pricing
• Claude Docs: Effort
• Anthropic: Claude Sonnet 5.5 System Card
• Claude Help Center: Release notes
• Artificial Analysis: Claude Sonnet 5.5 analysis
• GitHub Changelog: Claude Sonnet 5.5 in GitHub Copilot
• Claude Code 2.1.284 changelog
• Cursor announcement
• ComputingForGeeks: Claude Sonnet 5.5 released

If your team is building on these models, Codersera can help you hire vetted remote developers.