Claude Opus vs Sonnet, Tested: Which Model for Which Task

Which Claude model do you get on each plan?

Pro includes Opus, Sonnet and Haiku; Free includes Sonnet and Haiku only. The verified plan facts name model families, so they do not establish access to every specific version or confirm which models every higher-priced plan includes.

Prices and plans checked 2026-10-19. USD list prices exclude tax.

For Team’s annual options, the figures are per user per month with annual billing. Readers in the UK, Australia, India and elsewhere should check the official pricing page and applicable local rules: local prices, taxes and rules differ.

The current lineup is distinct from the models in our comparison. Anthropic recommends Opus 5.5 as its starting model; Sonnet 5.5 occupies the balanced tier.

Model lineup and API prices checked 2026-10-19. API charges below are separate from subscription prices.

Claude Mythos 5.1 is restricted to vetted cyber-defense and life-science organizations, not generally available to consumers.

We have not verified the vendor’s data retention, training or business-protection terms for this page. Read your plan’s data/privacy terms and professional confidentiality rules before pasting client, patient or employee material. For broader context, see our guide to Claude for professionals.

How did Opus and Sonnet score in our October 2026 rounds?

In our October 2026 round 1, Opus 5.5 and Sonnet 5 had similar aggregate results; in round 2, Opus had the higher aggregate. Both were graded by gpt-6-astra, but judges differ in strictness; these were API-harness evaluations, not consumer-app evaluations, and app results may differ.

Opus 5.5, Sonnet 5 and GPT-6 Astra were tested on 2026-10-19. Round 1 used 40 prompts, comprising 30 English and 10 Spanish prompts across 10 professions. Each had 5 prewritten rubric criteria and 3 runs per model per prompt, totaling 360 runs across those models on that date.

Round 2 used 40 prompts, comprising 20 French and 20 Brazilian Portuguese prompts, with the same design and 360 runs across Opus 5.5, Sonnet 5 and GPT-6 Astra on 2026-10-19. Its inputs included deliberate traps, such as a wrong total and a wrong unit conversion. Scores were lower for every model. Do not average the rounds: both language and input difficulty changed.

The models tested on 2026-10-19 were Claude Opus 5.5, Claude Sonnet 5 and GPT-6 Astra. This article compares the Claude pair. Sonnet 5 is an older version; current Sonnet 5.5 was not part of these rounds. Claude Fable 5.1 and Gemini 4 Argon were added on 2026-10-20 on all 80 prompts with the same rubrics and judge gpt-6-astra: Fable passed 90% of criteria in round 1 and 78% in round 2; Gemini passed 91% and 80%, respectively. Gemini ran through Google’s Antigravity CLI with a tool-less agent rather than the command-line API harness used for the other models; none ran through consumer apps, so app results may differ.

October 2026 results, test date 2026-10-19. Both Claude models shared judge gpt-6-astra; judges are not equally strict. Models ran through a command-line API harness with tools, web and memory off, so consumer-app results may differ.

Profession-level results show why the aggregate is insufficient for choosing a model.

October 2026 results, test date 2026-10-19. Values are shares of rubric criteria passed. Same judge: gpt-6-astra; unequal judge strictness remains a limitation. API harness, tools/web/memory off; consumer-app results may differ.

Calibration illustrates the judge limitation. In the October 2026 API-harness evaluation, regrading the same Claude outputs changed the scores: across 30 sampled round 1 runs, gpt-6-astra passed 93% of criteria versus 98% for swap judge claude-opus-5-5. Across 10 sampled round 2 runs, the figures were 74% and 86%. Neither judge establishes ground truth, and these scores do not predict consumer-app performance.

An editor agent, not a human, independently reread 20 round 1 verdicts and agreed with 18; it reread 10 round 2 verdicts and agreed with 9. These checks are not professional validation.

Most lost criteria involved added content: a warning, date, benefit or legal assumption absent from the input. Both Claude models also lost points on word limits.

When is Opus worth paying for instead of Sonnet?

Consider paying for Pro when you need access to Opus alongside Sonnet, then compare their outputs on representative, nonconfidential work. If you already have Pro, the verified plan includes both model families.

Our October 2026 results justify examining Opus on tasks resembling round 2, particularly in healthcare, construction and engineering. They do not establish that Opus is better for every numerical task or every non-English language. Both Claude models shared gpt-6-astra as judge, judges differ in strictness, and the API harness does not reproduce consumer-app use.

Use a short comparison workflow:

• Select a representative task and remove confidential information.
• Define what a satisfactory answer must include, exclude and verify before submitting it.
• Give each model the same source material and instructions.
• Check unsupported additions, missing requirements, calculations and length.
• Choose based on the output you can verify and the work needed to correct it.

Do not treat the language difference as a proven cause of the round 2 gap. French and Brazilian Portuguese appeared alongside deliberate traps, so the evaluation cannot isolate those effects. Nor does it establish performance in untested languages or on spreadsheets generally. In a separate round of 10 Excel and Microsoft 365 prompts on 2026-10-20, Opus 5.5 passed 93% of criteria and Sonnet 5 passed 83%, both judged by gpt-6-astra. Chat models answered pasted tasks in text, with formulas judged on logic and stated result but not executed in Excel; this was not Copilot in Excel or Microsoft 365, and Copilot remains untested.

For accounting workflows, our AI for accountants guide provides context for deciding what needs independent verification. For legal work, consult AI for lawyers before assigning a drafting or review task.

The following suggested templates target the observed failure modes; they are not separately validated prompts. Use placeholders or nonconfidential examples.

Review a draft [DOCUMENT TYPE] against its source.

Source material, the only permitted factual basis:
[PASTE NONCONFIDENTIAL SOURCE]

Draft:
[PASTE NONCONFIDENTIAL DRAFT]

List statements unsupported by the source. Quote each sentence
and label it "not in source."

List figures, dates and names that differ from the source.
Show the draft value and source value.

Do not add warnings, recommendations or factual context.
Keep the response under [WORD LIMIT] words.
If the material is insufficient, identify what is missing.

For a table containing totals or conversions:

Check this nonconfidential table from [DOCUMENT DESCRIPTION]:

[PASTE TABLE]

Recompute stated subtotals and totals.
Check stated unit conversions.
Report discrepancies with your calculation and corrected value.
Do not infer missing values.
If a check cannot be completed, name the missing information.
Do not interpret what the figures mean.
Keep the response within [LINE LIMIT] lines.

Independently check the calculations in the response. For other legal-task examples, browse our prompt library for lawyers.

Where does Claude Fable 5.1 fit?

Claude Fable 5.1 is Anthropic’s flagship, listed for demanding reasoning and long-horizon agentic work. Added to our rounds on 2026-10-20 with the same prompts, rubrics and judge gpt-6-astra, it passed 90% of criteria in round 1 (67 of 120 runs met every criterion) and 78% in round 2 (28 of 120). Judges differ in strictness: on 10 sampled round 1 Fable runs, gpt-6-astra passed 88% of criteria versus 100% for swap judge claude-opus-5-5, so these results do not establish an overall winner.

Released 2026-09-01, it has a listed API price of $10 / $50 per MTok for input/output. On the verified subscription facts, Free and Pro exclude Fable; Max 5x includes it at 50% of weekly limits, from $100 per month. Those are different purchasing routes, not a single entry price.

For a non-developer choosing between Opus and Sonnet, Fable’s flagship label does not answer whether its output would improve their work. Define the task and verification requirements before considering a different plan.

Limits: what must a licensed professional check?

A licensed professional in the reader’s jurisdiction must check legal, clinical or tax conclusions before relying on them. This article is not legal, medical or tax advice; “tested” here means at least 2 runs completed and graded, not professional sign-off.

• US healthcare: HIPAA Privacy Rule 45 CFR 164.502(e) permits a covered entity to disclose PHI to a business associate only after obtaining satisfactory assurance, through a business associate agreement, that it will safeguard the information. An AI vendor receiving PHI is a business associate. Verify the agreement before disclosure; do not assume a subscription supplies it.
• Confidentiality: read the specific plan’s data/privacy terms and applicable professional rules. We have not verified retention, training or business protections.
• Output accuracy: check every consequential figure, date, citation and assumption against authoritative source material. A plausible answer is insufficient.
• Local requirements: check jurisdiction-specific rules rather than applying the US discussion elsewhere.

Our guide to AI for doctors and nurses provides further context for professional review.

FAQ