Claude Opus vs Sonnet: The 2x Premium Buys You Open-Ended Work
Claude Opus 5.5 costs $4 and $20 per million tokens, double Sonnet 5.5's $2 and $10. On Anthropic's own coding benchmarks Sonnet is within a few points, and ahead on Terminal-Bench 4.0. The gap opens on long, open-ended jobs.

Key takeaways
- Claude Opus 5.5 (released September 22, 2026) lists at $4 per million input tokens and $20 per million output tokens, and Claude Sonnet 5.5 (released September 28, 2026) lists at $2 and $10, so Opus costs exactly twice as much per token.
- On Anthropic's own benchmarks Sonnet 5.5 scores 70.6% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, and trails Opus by 2.3 points on CursorBench 4.0 (55.5% versus 57.8%), but the gap widens to 8.2 points on FrontierCode 1.1 (46.2% versus 54.4%).
- Anthropic says Opus 5.5 remains the stronger model for complex, open-ended work requiring sustained judgment, while Sonnet 5.5 is strongest at well-scoped everyday tasks.
- Opus's price premium over Sonnet was 5x when Claude Opus 4.1 ($15 and $75) sat beside Claude Sonnet 4 ($3 and $15), and it has been between 1.67x and 2.5x since Opus 4.5.
- Claude Code defaults to Opus 5.5 on Pro, Max, Ultra, Team, Enterprise and the API, and its opusplan alias plans with Opus and executes with Sonnet.
Claude Opus 5.5 is Anthropic's stronger model for long, open-ended agentic work, priced at $4 per million input tokens and $20 per million output tokens. Claude Sonnet 5.5 is the faster everyday model at $2 and $10, and on several of Anthropic's own coding benchmarks it lands within a few points of Opus, or ahead.[1,5]
So here is the short answer: start with Sonnet 5.5 for well-scoped tasks, and reach for Opus 5.5 when the job needs sustained judgment across many steps.
Both models share a 1M-token context window, a 128K-token output limit and a June 2026 knowledge cutoff, so the decision is about price, speed and reasoning, not capacity.[1] Every figure below was checked against Anthropic's pages on November 18, 2026, and the benchmarks are Anthropic's own numbers, not independent tests.
Opus and Sonnet are close on tasks with a clear finish line
I expected Opus to win every row of the benchmark table. It doesn't.
Anthropic's Sonnet 5.5 launch page puts the two side by side, and on Terminal-Bench 4.0, a test of agentic coding in a terminal, Sonnet scores 70.6% against 66.4% for Opus, with Anthropic noting the Opus figure was run at xhigh effort.[5]
On CursorBench 4.0, which scores agentic coding tasks, Opus leads by 2.3 points, 57.8% to 55.5%.[4,5]
Sonnet 5.5
Opus 5.5
- Terminal-Bench 4.0 (agentic coding)70.6%66.4%
- CursorBench 4.0 (agentic coding)55.5%57.8%
- OSWorld 2.1 (computer use, partial)80.1%81.8%
- Humanity’s Last Exam (with tools)64.5%67.7%
- FrontierCode 1.1 (Sonnet at Max effort)46.2%54.4%
Takeaway
Sonnet matches or beats Opus on a terminal coding test and trails by 1.7 to 3.2 points on three others. The one large gap, 8.2 points on FrontierCode 1.1, is where the problems get long and underspecified.
Read that FrontierCode row with a caveat. Anthropic notes that Sonnet scores lower at Max effort than at Xhigh, and the page is ambiguous about what Sonnet scores at Xhigh, so the 8.2-point gap may be smaller than it looks.[5]
Anthropic also warns on the Opus 5.5 page that benchmark margins are “a less reliable guide to real-world differences.”[4] That is a sensible warning from a company publishing the margins.
Anthropic's own summary of where the real difference sits is plain. Sonnet 5.5 “is strongest at well-scoped everyday tasks,” and Opus 5.5 remains stronger at “complex, open-ended work requiring sustained judgment.”[5]
“Complex, open-ended work requiring sustained judgment.”
That matches what the model pages list as typical uses. Opus 5.5 is for “multihour autonomous coding agents, large-scale refactoring, complex systems engineering, vision-heavy workflows, and computer use.” Sonnet 5.5 is for code generation, data analysis, content creation, visual understanding and agentic tool use.[2,3]
The Opus premium shrank from 5x to 2x
The part of this story people miss is the price. Opus used to cost five times Sonnet, and the lineup now makes the premium a modest doubling. Here is every Opus and Sonnet pairing on Anthropic's models page, with the multiple worked out.
| Pairing | Opus (USD per million input / output tokens) | Sonnet (USD per million input / output tokens) | Opus premium |
|---|---|---|---|
| Opus 4.1 and Sonnet 4 | $15 / $75 | $3 / $15 | 5x |
| Opus 4.5 and Sonnet 4.5 | $5 / $25 | $3 / $15 | 1.67x |
| Opus 4.6 and Sonnet 4.6 | $5 / $25 | $3 / $15 | 1.67x |
| Opus 5 and Sonnet 5 | $5 / $25 | $2 / $10 | 2.5x |
| Opus 5.5 and Sonnet 5.5 | $4 / $20 | $2 / $10 | 2x |
Source: Anthropic, Models overview, “All models” table, November 18, 2026. Multiples are my arithmetic. Opus 4.1 and Sonnet 4 are listed as retired, with exceptions on two cloud platforms.
Opus 5.5 also undercuts its own predecessor. Anthropic says it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5,” and its list price fell from $5 and $25 to $4 and $20.[4]
Sonnet 5.5 kept Sonnet 5's price but runs 30% or more faster, and Anthropic says it typically needs fewer tokens per task, for up to 30% lower cost per task.[5]
Sticker price is not task price
A 2x per-token premium is only a 2x task premium if both models use the same tokens. They don't have to. Thinking depth, output length and how often the agent re-reads files all change the count. Anthropic says so directly: “Tuning effort is often a better lever than switching models.”[6]
For a feel of the raw spread, take an illustrative job I chose: 10 million input tokens and 1 million output tokens, no caching. At list price that is $60 on Opus 5.5 and $30 on Sonnet 5.5.
Prompt caching narrows the input side, since a cache read costs $0.20 per million tokens on Opus and $0.10 on Sonnet, and the Batch API takes 50% off both.[2,3] Those levers move the bill more than the model choice does on repetitive agent work.
Heads up
high for Sonnet 5.5 and medium for Opus 5.5, while Claude Code's configuration page says both default to medium.[1,7] Set effort explicitly if you are comparing the two, or you may be benchmarking the default, not the model.| Spec | Opus 5.5 | Sonnet 5.5 |
|---|---|---|
| API model ID | claude-opus-5-5 | claude-sonnet-5-5 |
| Released | September 22, 2026 | September 28, 2026 |
| Latency (Anthropic's label) | Moderate | Fast |
| Thinking | Adaptive, always on | Adaptive |
| Cache write, 5 minutes (USD per million tokens) | $5 | $2.50 |
| Cache read (USD per million tokens) | $0.20 | $0.10 |
| Retirement commitment | Not sooner than September 22, 2027 | Not sooner than September 28, 2027 |
Sources: Anthropic model pages for Opus 5.5 and Sonnet 5.5, November 18, 2026. Both: 1M-token context, 128K-token output, June 2026 knowledge cutoff.
Anthropic's own defaults lean toward Opus
Here is the detail that made me stop. Anthropic's model-choice guide says “Most workloads start with Claude Opus 5.5,” and Claude Code uses Opus 5.5 as its default model on Pro, Max, Ultra, Team, Enterprise and the API.[6,7]
That is a sensible default for a vendor selling a more capable model, and it is also the more expensive path for anyone who never changes it.
My read is that the default is right for open-ended agent sessions and wasteful for the daily small stuff.
If you are renaming a function, writing a test for a pure function or summarizing a file, you are paying double per token for judgment you won't use. Claude Code makes the switch cheap: /model sonnet mid-session, or claude --model sonnet at startup.[7]
There is also a built-in hybrid. The opusplan alias uses Opus in plan mode for reasoning and architecture decisions, then switches to Sonnet to write the code.[7]
It matches Anthropic's general advice to pair a cheaper model with a frontier one so most tokens are billed at the lower rate, either as an executor that escalates hard decisions or an orchestrator that delegates bulk work.[6]
If you build agents yourself, the same split is a single model field per subagent, which we cover in the Claude Agent SDK explainer.
A rule for choosing, and a rule for checking it
A starting point, not a law
Ask how much judgment the job needs, then start there.
What the job looks like
- Clear finish lineBug fix, test, document, data pull
- Long and underspecifiedRefactor, migration, multihour agent run
- Volume and speed pressureBulk jobs, tight latency
How much judgment does it need?
Then test on your own tasks
Where to start
- Sonnet 5.5, $2 / $10Everyday coding, agents, documents
- Opus 5.5, $4 / $20Long-running agentic coding, computer use
Sources: Anthropic's choosing-a-model guide and model pages. Fable 5.2 ($8 / $40) sits above Opus for the hardest work.
Now the checking rule, which matters more than the starting rule. Model quality can shift under you, and the Opus 4.7 regression postmortem is the reminder that even a flagship can feel worse for a stretch.
Anthropic's guidance is to build benchmark tests specific to your use case and run your real prompts through each model, comparing accuracy, quality and edge cases.[6] Ten of your own failing tasks will teach you more than the table above.
- Start on Sonnet 5.5 for bug fixes, tests, documents, data analysis and anything you could describe in one clear paragraph.
- Start on Opus 5.5 for multihour autonomous runs, large refactors, systems work and computer use. If Opus at high effort still falls short, Anthropic points to Fable 5.2.[2,6]
- Mind your plan. Anthropic's pricing page lists Sonnet and Haiku on the Free plan with Opus unavailable, and Opus on Pro and above.[8]
If you want last spring's version of this argument, with GPT in the ring, read GPT-5.5 vs Claude Opus 4.7. For how the tools around these models bill you, see Claude Code vs Cursor, and for automating either model on a schedule, the Claude Code routines guide.
My verdict: Opus is not twice as good as Sonnet, and Sonnet is not a cut-down Opus. They are two tools priced a factor of two apart, and the skill is knowing which kind of job you are holding. Pick Sonnet until it fails you, then pay for Opus on purpose.
Primary sources
- 1.PrimaryAnthropic, "Models overview". Checked November 18, 2026. Current lineup, prices, context window, output limits, knowledge cutoff, default effort and the full list of older models with their prices.
- 2.PrimaryAnthropic, "Claude Opus 5.5" model page. Checked November 18, 2026. Release date September 22, 2026, pricing including cache rates, typical uses, retirement commitment.
- 3.PrimaryAnthropic, "Claude Sonnet 5.5" model page. Checked November 18, 2026. Release date September 28, 2026, pricing including cache rates, typical uses, retirement commitment.
- 4.PrimaryAnthropic, "Claude Opus 5.5" announcement. September 22, 2026. Benchmark table versus Opus 5, price versus Opus 5, the 40% lower cost to run claim and the warning on benchmark margins.
- 5.PrimaryAnthropic, "Claude Sonnet 5.5" announcement. September 28, 2026. Sonnet 5.5 versus Opus 5.5 benchmark table with footnotes, speed and cost claims, and the positioning quotes.
- 6.PrimaryAnthropic, "Choosing the right model". Checked November 18, 2026. Starting-point guidance, the effort-as-a-lever note, multimodel strategies and the advice to build your own benchmarks.
- 7.PrimaryAnthropic, "Claude Code model configuration". Checked November 18, 2026. Default model by plan, model aliases, switching with /model and --model, the opusplan alias and effort defaults.
- 8.PrimaryAnthropic, "Claude pricing". Checked November 18, 2026. Which models each consumer plan includes and the plan prices.
Frequently asked questions
- What is the difference between Claude Opus and Sonnet?
- Opus is Anthropic's stronger model for long-running agentic coding and knowledge work, and Sonnet is the faster, cheaper model for everyday coding, agent and enterprise workloads. As of November 18, 2026, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, and Sonnet 5.5 costs $2 and $10. Both have a 1M-token context window and 128K-token maximum output.
- Is Claude Opus better than Sonnet?
- Opus is better at open-ended, multi-step work, and Sonnet is close or ahead on well-scoped tasks. Anthropic's benchmarks show Sonnet 5.5 at 70.6% on Terminal-Bench 4.0 against Opus 5.5's 66.4%, but Opus leads 54.4% to 46.2% on FrontierCode 1.1. Anthropic itself says Opus is the stronger model for complex work requiring sustained judgment.
- How much does Claude Opus cost compared to Sonnet?
- Opus 5.5 costs twice as much per token as Sonnet 5.5: $4 versus $2 per million input tokens, and $20 versus $10 per million output tokens. Cache reads cost $0.20 versus $0.10 per million tokens, and the Batch API takes 50% off both. Real cost per task also depends on how many tokens each model uses, so the sticker price is only half of the bill.
- Which Claude model should I use for coding?
- Start with Sonnet 5.5 for well-scoped coding such as bug fixes, tests and single-feature changes, and move to Opus 5.5 for multihour autonomous agents, large refactors and complex systems work. Claude Code defaults to Opus 5.5, and its opusplan alias uses Opus in plan mode and Sonnet for execution. Test both on your own tasks before you standardize.
- Can you use Claude Opus on the free plan?
- No, Anthropic's pricing page lists Sonnet and Haiku on the Free plan and shows Opus as unavailable. Opus is available on Pro ($20 per month billed monthly, $17 billed annually), Max ($100), Ultra ($200), Team and Enterprise plans. Claude Code is also not included in the Free plan.
- What is Claude Fable, and is it better than Opus?
- Claude Fable 5.2 is the tier above Opus, built for demanding reasoning and long-horizon agentic work, and it costs $8 per million input tokens and $40 per million output tokens. Anthropic's guidance is to try Opus 5.5 first and move to Fable 5.2 if your evaluations at high effort still fall short.
Written by
Tech Talk News Editorial
Computer engineering background. Writes about software, AI, markets, and real estate, and the places where the three meet.
More about the author