Claude Opus 5.5 vs Sonnet 5.5: Which Model for Which Job?

What separates Claude Opus 5.5 and Sonnet 5.5?

Claude Opus 5.5 released on 22 September 2026, and Claude Sonnet 5.5 on 28 September 2026 are Anthropic’s current models for long-running agentic coding. Both offer a 1 million token context window and 128,000 token output. Opus 5.5 costs $4 and $20 per million tokens; Sonnet 5.5 costs half that, at near-identical benchmark results.

What’s new in Claude Opus 5.5 and Sonnet 5.5?

Anthropic released Claude Opus 5.5 on 22 September 2026, with Claude Sonnet 5.5 following six days later on 28 September 2026. Both are built for long-running agentic coding and knowledge work tasks that run across many steps and tool calls.

The core specifications are identical: A 1 million token context window, a maximum of 128,000 output tokens per response, and a knowledge cutoff of June 2026. On the Message Batches API, the output-300k-2026-03-24 beta header raises the output ceiling to 300,000 tokens.

The most consequential change concerns thinking itself: Adaptive thinking is always on in both models and cannot be switched off in Opus 5.5. Depth is controlled through the effort parameter — and the defaults differ: Opus 5.5 starts at medium, Sonnet 5.5 at high. Anyone comparing the two without setting effort explicitly is comparing two different configurations.

Both are available on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and the Claude Platform on AWS. For Opus 5.5, Anthropic has committed to no retirement before 22 September 2027, a dependable planning horizon for production integrations. On the previous generation: how Claude Opus 4.6 handles long coding tasks.

Benchmarks head-to-head

Opus 5.5 leads the SWE-Bench family clearly — 89.9% against 81.3% on SWE-Bench Pro. Sonnet 5.5 is more than 30% faster than Sonnet 5 and, per Anthropic, costs up to 30% less per task because it uses tokens more efficiently.

The effort trap in benchmark tables

One number in that table deserves a warning. Sonnet 5.5’s 70.6% on Terminal-Bench 4.0 was measured at effort level max; Opus 5.5’s 66.4% at xhigh — one tier lower. Anthropic labels the Opus figure explicitly as the model’s highest score at xhigh, while the Sonnet figure carries no effort label in the table.

That is not a like-for-like comparison. The lesson matters more than the scores: without a stated effort level, a model comparison is worthless. Check every vendor table for the setting it was measured at. That applies across the industry — including Google’s cyber-defense model, Gemini 4 Carbon, whose security benchmarks were run without standard guardrails in place.

What do Claude Opus 5.5 and Sonnet 5.5 cost?

Opus 5.5 is cheaper than its predecessor. Opus 5 cost $5 and $25; Opus 5.5 at $4 and $20 is exactly 20% below that. Anthropic puts the savings on typical workloads at default settings at around 40%, because the model also consumes fewer tokens. Cache reads fell from $0.50 to $0.20 per million tokens — a 60% cut.

Two further line items reduce the bill: cache reads cost 5% of the base input price on both models, and the Batch API takes 50% off input and output. If you need maximum speed, Opus 5.5 runs in Fast Mode at up to 2.5× output speed for $8 and $40.

Budget in cost per completed task, not in token prices. A model that costs half as much but needs two attempts is not the cheaper model.

Which model for which job?

The decision reduces to three questions.

Is a failed attempt expensive? → Opus 5.5. Migrations spanning many files, architectural decisions, cross-disciplinary analysis, and long autonomous agent runs with no human checkpoint. Its SWE-Bench Pro lead shows up precisely there.

Interactive or high volume? → Sonnet 5.5. Code review, tests, mid-sized refactors, documentation, ticket triage — anything that benefits from a fast response. For the bulk of engineering work, Sonnet 5.5 is the more economical choice.

Classification or extraction? → Haiku 5.5. High-volume routine work from $0.10 per million input tokens.

A tiered setup — Haiku pre-filters, Sonnet does the work, Opus escalates the hard cases — is almost always cheaper in practice than one model for everything. For the competitive picture, our analyses of GPT-6 Astra and GPT-5.6 cover the other side of the frontier race.

Breaking changes when migrating

The checklist differs by model

Opus 5 → Opus 5.5 — four breaking changes:

1. Thinking cannot be disabled. Adaptive thinking is permanently on; control depth with effort.

2. Forced tool use returns an error. Pinning tool choice to a specific tool returns an error instead of a response.

3. Thinking blocks are tied to the model and conversation that produced them. Blocks from another model or conversation cannot be replayed.

4. The computer_20251124 computer use tool is no longer accepted, affecting the Claude API and Google-managed Claude models on Google Cloud.

Sonnet 5 → Sonnet 5.5 — five breaking changes: 

1. Points 2 to 4 apply here too, plus two Sonnet-specific changes. 

2. Up-front thinking can be turned off with the between tools setting (at effort high or below), and the advisor tool no longer accepts four models as advisors — Claude Opus 4.8, Opus 4.7, Sonnet 5, and Haiku 5.5. 

3. Sonnet 5.5 also returns a 400 error if you set temperature, top_p, or top_k. Those parameters appear in a great many existing integrations.

4. There is also a change that fails no request but can silence your interface: text between tool calls now comes back inside thinking blocks, whose text is empty at the default display setting. 

5. Applications that stream that text as progress updates go quiet between tool calls until a display value is set that returns it. This is the defect that passes testing and surfaces in production.

What should companies do now?

1. Set effort explicitly on every call rather than relying on defaults. Otherwise, you are comparing configurations and paying unpredictably.

2. Measure cost per task, not token prices. Instrument the logging before you migrate, not after.

3. Work through every migration point, including the progress-display change and the 400 error, first. Our managed cybersecurity services from TechNow also audit which data your agents actually send to the model.

4. Tier your models instead of standardising on one. Haiku pre-filters, Sonnet works, Opus escalates.

5. Implement prompt caching if your agents reuse large or stable context. At 5% of the input price, this is the single biggest cost lever available.

For day-to-day operations, we run IT support for mid-market companies. If you want your model strategy costed out properly, book an AI adoption assessment.

FAQ

Which is better, Claude Opus 5.5 or Sonnet 5.5?
Opus 5.5 leads most benchmarks, 89.9% against 81.3% on SWE-Bench Pro. Sonnet 5.5 costs half as much and responds over 30% faster. For the bulk of engineering work, Sonnet 5.5 is more economical; for long autonomous runs, Opus 5.5.

When were Claude Opus 5.5 and Sonnet 5.5 released?
Anthropic released Claude Opus 5.5 on 22 September 2026 and Claude Sonnet 5.5 on 28 September 2026.

How large is the context window in Claude Opus 5.5 and Sonnet 5.5?
Both process 1 million tokens of context and output a maximum of 128,000 tokens per response.

What do Claude Opus 5.5 and Sonnet 5.5 cost?
Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 costs $2 and $10. The Batch API applies a 50% discount.

What are the breaking changes when migrating from Opus 5?
There are four stages: thinking cannot be disabled, forced tool use returns an error, thinking blocks are tied to the model and conversation that produced them, and computer_20251124 is no longer accepted.

Sources

• Anthropic: Claude Opus 5.5 model documentation

• Anthropic: Claude Sonnet 5.5 model documentation

• The Decoder: Sonnet 5.5 benchmarked against Opus 5.5