Claude Sonnet 5.5: The New Default for Most Teams

Claude Sonnet 5.5 keeps Sonnet 5 pricing, runs faster and comes within a few points of Opus 5.5. The case for making it your default model.

Disclosure: this article draws on Anthropic’s announcement, Artificial Analysis evaluations and published third-party testing, not on a months-long deployment of our own. Figures are attributed where they appear.

Claude Sonnet 5.5 launched on September 28, 2026, six days after Opus 5.5, and it quietly made the more expensive model harder to justify. Same price as Sonnet 5, about 30% faster, and within a few points of Opus 5.5 on most coding benchmarks. On one terminal coding test it actually comes out ahead.

The honest verdict: for most teams, Claude Sonnet 5.5 should now be the default model, with Opus kept as an escalation path for the small share of tasks where Sonnet fails. That is not a hot take. It is roughly what Anthropic itself implies when it says Opus remains clearly stronger only at complex, open-ended work that needs sustained judgment.

Table of Contents

Claude Sonnet 5.5 keyboard closeup for daily coding work
Claude Sonnet 5.5 did not get cheaper per token. It got cheaper per task, which is the number that shows up on the invoice.

What Claude Sonnet 5.5 Changed

Claude Sonnet 5.5 is the second model in the Claude 5.5 family and the direct replacement for Sonnet 5 as Anthropic’s mid-tier model. The API ID is claude-sonnet-5-5, with anthropic.claude-sonnet-5-5 on Amazon Bedrock. It is also available on Google Cloud, Microsoft Foundry and Claude Platform on AWS.

Anthropic’s pitch has three parts. It runs more than 30% faster than Sonnet 5. It costs up to 30% less per task, because it uses fewer tokens and fewer tool calls to reach the same result. And it writes more clearly, picking up the more natural voice Anthropic introduced with Opus 5.5.

The context window is one million tokens, billed at the standard rate across the whole window. That last detail is easy to skip, but it matters: several competitors charge double once a prompt passes 272,000 tokens, applied to the whole request. Long-document work on Claude Sonnet 5.5 has no such cliff.

Anthropic also added protections against large-scale distillation, meaning attempts to extract the model’s reasoning by harvesting outputs at volume. For ordinary users this is invisible. For anyone running unusual high-volume scraping of reasoning traces, it may not be.

Claude Sonnet 5.5 Pricing and Real Cost per Task

List prices did not move. Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens. There is a backstory here: Sonnet 5 launched in June at $2 and $10 as an introductory rate, with a planned rise to $3 and $15 on September 1. Anthropic made the lower price permanent in August instead, and Sonnet 5.5 inherited it.

Item Claude Sonnet 5.5 Opus 5.5
Input per 1M tokens $2 $4
Output per 1M tokens $10 $20
Cache reads per 1M $0.10 (cut from $0.20 on October 7) $0.20
Context 1M at standard rate 1M at standard rate

The cache cut is the sleeper story. On October 7, alongside the Haiku 5.5 launch, Anthropic halved Sonnet cache reads from $0.20 to $0.10 per million. Anthropic estimated roughly 20% lower cost for most agentic work as a result. Agents re-send their history on every turn, so the cached share of input is large, and halving its price moves the bill.

Artificial Analysis found prices varying up to 18 times across effort settings. At low effort, the cheapest configuration came in at $0.41 per task on its intelligence index suite. That spread is the real pricing lesson: the effort dial matters more than the list price.

Anthropic also claims that on several benchmarks, Claude Sonnet 5.5 at low or medium effort beats Sonnet 5’s best score for about a tenth of the cost per task. If that holds on your workload, upgrading is close to free money.

Claude Sonnet 5.5 Benchmarks in Context

The number everyone quoted was Terminal-Bench 4.0. Sonnet 5 scored 10.3%. Claude Sonnet 5.5 scores 70.6%. A sixty-point jump in a single release is unusual enough that it deserves scepticism, and Labellerr’s write-up rightly called it a different model rather than an incremental update.

The rest of the coding picture is less dramatic. On FrontierCode 1.1, Anthropic lists 46.2% at max effort and 52.1% at xhigh, against 49.3% for GPT-6 Sol on the same benchmark. Good, competitive, not a blowout.

Artificial Analysis scored it at 56 on its Intelligence Index at max effort, which put it in the 99th percentile of tracked models at the time. But the same analysis noted that at $2 and $10, matching GPT-6 Sol’s price, Claude Sonnet 5.5 sits off the intelligence-versus-cost Pareto frontier. Translation: it is very good, but not the cheapest way to buy that level of intelligence.

A footnote worth knowing: Artificial Analysis ran its initial evaluations on a pre-release build that had a bug degrading structured outputs. Anthropic fixed it for the public release, and the testers expected minimal change or slightly understated performance. So the published scores may run a little low.

Claude Sonnet 5.5 vs Opus 5.5

Here is how the two compare where it counts.

Dimension Claude Sonnet 5.5 Opus 5.5
Price Half of Opus on every line $4 / $20
Terminal coding Slightly ahead on one benchmark Close behind
Factual recall (AA-Omniscience) 54% 66%
Open-ended judgment Good Clearly stronger, per Anthropic
100-turn agent session About $4 About $8

The session estimate comes from Claude Fast’s arithmetic at current cache rates. The factual-recall gap is the one I would weigh most heavily. Twelve points on a knowledge test shows up in research tasks, where the model needs to know things rather than just reason about the text in front of it.

So the split is fairly clean. Use Claude Sonnet 5.5 for coding, tool-heavy agents and anything high-volume. Escalate to Opus for long-horizon judgment and knowledge-heavy analysis.

Breaking Changes to Check

Claude Fast counted five breaking changes against Sonnet 5. The one most teams hit first: setting temperature, top_p or top_k to a non-default value now returns a 400 error. If your code sets temperature to 0 for determinism, it will fail until you remove that parameter.

Effort levels replace manual thinking budgets. Configurations tuned for Sonnet 5 will not map one to one. Plan to retest rather than copy settings across.

Structured-output behaviour deserves a test pass too, given the pre-release bug, even though Anthropic says it is fixed. If your pipeline depends on strict JSON, run your validation suite against the production model before cutting over.

Picking an Effort Level

The effort dial is where most of the savings live, so it is worth a simple rule of thumb.

  • Low: extraction, classification, short answers, quick edits. Fast and cheap.
  • Medium: the sensible default for coding assistance and everyday agent steps.
  • High or xhigh: multi-file changes, tricky debugging, planning steps in an agent.
  • Max: the hardest tasks only, where a failure costs more than the extra tokens.

A good pattern is to start each task type at medium, log success rates for a week, and move up only where failures cluster. Teams that set everything to max by default end up paying Opus-level bills for Sonnet-level work.

Where Claude Sonnet 5.5 Still Falls Short

Three limits are worth stating plainly. The first is knowledge. The recall gap against Opus is real, so research-heavy work benefits from retrieval, either through search tools or your own document store.

The second is price efficiency at the top end. Artificial Analysis’s Pareto observation means there are cheaper routes to similar intelligence, especially among open-weight models such as DeepSeek V4.1 Flash. If cost per task is your only metric, Claude Sonnet 5.5 is not the winner.

The third is the effort tax. Max effort can be slow and expensive. Interactive use needs low or medium settings, and that trade-off is something you have to manage rather than something the model manages for you.

There is also the question of lock-in. Anthropic ships through four clouds, which helps, but a codebase full of model-specific prompt tricks and parameters is harder to move than one that treats the model as a swappable dependency. Given how often frontier releases now land, building for portability is cheap insurance. Keep a thin wrapper around every model call, keep evaluation tasks in version control next to the code, and you can move to whatever wins next quarter without a rewrite.

Claude Sonnet 5.5 for Coding Agents

Coding agents are where the release makes the biggest practical difference, because agents magnify both speed and token efficiency. An agent that needs eight tool calls instead of twelve to fix a bug is not just cheaper. It is faster to review, and it has fewer chances to wander off course.

Anthropic’s claim of fewer tool calls per task lines up with what early testers described. Computingforgeeks ran the same three DevOps tasks it had used on Opus 5.5 and GPT-6 Sol, grading the output with real linters and a live cluster. Its conclusion was that both effects Anthropic advertised, speed and token savings, showed up, but how much you get depends on the workload and the effort level. That is the most honest summary available.

For a typical setup, Claude Sonnet 5.5 works well as the main worker in an agent loop: reading files, proposing edits, running tests and iterating. Planning steps that decide the overall approach can go to Opus if your budget allows, while Haiku handles cheap sub-tasks such as summarising logs or classifying test failures. That three-tier arrangement keeps the expensive model where it adds value and nowhere else.

One caution. Agents that run unattended for long periods still need guardrails regardless of model: limits on file system access, a cap on total spend per run, and a human review before anything merges. Better models reduce failure rates. They do not remove the need to plan for failure.

Claude Sonnet 5.5 and Your Existing Prompt Library

If you have a library of prompts tuned for Sonnet 5, expect most of them to work, and a few to work worse than before. Newer models tend to follow instructions more literally, so prompts stuffed with emphatic warnings written to correct an older model’s habits can make the new one overly cautious or rigid.

The fix is usually subtraction. Strip out repeated warnings, remove examples that only existed to patch old failure modes, and state the task plainly. Then test. Claude Sonnet 5.5 is better at inferring intent from a short, clear instruction than its predecessor, and long prompts also cost more on every single call.

Put stable content at the top of each prompt so it caches. With cache reads at $0.10 per million, a well-structured prompt that reuses a large fixed preamble becomes very cheap to call repeatedly. That ordering change alone often saves more than any amount of wording tweaks.

Making the Switch in a Real Team

Most teams do not switch models in one go, and they should not. A staged rollout catches the surprises before they reach customers.

Start with a shadow week. Send a copy of real traffic to the new model, keep the old one in production, and compare outputs side by side. You are looking for three things: tasks where quality drops, tasks where cost rises, and any change in output format that breaks downstream parsing. The temperature error described above usually surfaces in the first hour, which is exactly why you want it surfacing in a shadow run.

Then move one workload at a time, starting with the lowest-risk one. Internal tools and code review assistants are good first candidates, because the people using them will tell you quickly if something feels off. Customer-facing features go last.

Keep the old model reachable for a month. Rolling back should be one configuration change, not a code deploy. And write down what you measured, because in eight weeks there will probably be another model to evaluate, and having last month’s baseline turns that evaluation from a debate into a comparison.

One more habit pays off: tag each request with its task type. When you later look at the bill, you want to see that summarisation costs pennies and refactoring costs dollars, not one undifferentiated total. That breakdown is what tells you where a cheaper model could take over.

Who Should Use What

If you are Start with Why Link
A dev team using AI daily Claude Sonnet 5.5 at medium Near-Opus coding at half the price See Sonnet 5.5
Running long research agents Opus 5.5 Stronger factual recall and judgment Compare models
Processing millions of short tasks Haiku 5.5 A fraction of the price per token See pricing
Chasing the lowest cost per task An open-weight Flash model Sits closer to the cost frontier Read the specs

Related on this site: for more context, see Claude Opus 5.5, Claude Haiku 5.5 and Claude Code vs Codex vs Cursor.

Frequently Asked Questions About Claude Sonnet 5.5

When did Claude Sonnet 5.5 come out?
Anthropic released it on September 28, 2026, six days after Opus 5.5. Haiku 5.5 followed on October 7, completing the 5.5 family apart from Fable.

What does Claude Sonnet 5.5 cost?
$2 per million input tokens and $10 per million output tokens, the same as Sonnet 5. Cache reads dropped to $0.10 per million on October 7, which Anthropic says lowers most agentic costs by about 20%.

Is it better than Opus 5.5?
On one terminal coding benchmark it edges ahead, and it trails by a few points elsewhere. Opus remains stronger on factual recall and complex open-ended work, so the answer depends on the task.

Why does my temperature setting throw an error?
Claude Sonnet 5.5 rejects non-default temperature, top_p and top_k values with a 400 error. Remove those parameters and use effort levels to control behaviour instead.

How big is the context window?
One million tokens, billed at standard rates across the whole window. There is no long-context surcharge, which makes it attractive for large document and codebase work.

Is the Terminal-Bench jump real?
The 10.3% to 70.6% figure is Anthropic’s own. Independent testing broadly confirms a large coding improvement, but treat any single benchmark jump that size as something to verify on your own tasks.

Verdict

Make Claude Sonnet 5.5 your default model at medium effort. For coding, agents and everyday knowledge work, it delivers most of what Opus 5.5 does at half the price, and the October cache cut made long agent sessions noticeably cheaper.

The caveat is that it is not the most cost-efficient model on the market, and its factual recall trails Opus by a meaningful margin. Most published benchmark numbers are vendor-reported, and the headline coding jump is the kind of figure that deserves your own verification.

Next step: remove any temperature settings from your code, switch one low-risk workload over, and compare a week of cost and success rates against your previous model before moving everything.

Sources and Further Reading

On this site: see the top tier in our Claude Fable 5.1 guide, choose tooling with AI coding assistants in 2026, trim spend with LLM API cost optimization, and compare vendors in AI models compared.