Anthropic has released another Claude, because apparently choosing what to ask AI was not enough—we now have to choose which AI deserves the question. Sonnet 5.5 arrived September 28, 2026, as a faster everyday option alongside Opus 5.5, not a reason to use one model for every job.
For most clearly defined work—turning supplied notes into a client email, drafting a presentation, or fixing a reproducible bug—I would start with Sonnet 5.5. Keep Opus 5.5 for the messy assignment where the requirements conflict, the stakes are higher, or the first answer still needs too much rescuing.
The short answer: choose by the work, then by the result
Start here
Sonnet 5.5
Use it for a clear brief with a checkable result: everyday writing, document or slide drafts, summaries, and well-defined coding fixes.
If it misses a material constraint after a fair correction, move the task up rather than asking for ten increasingly desperate rewrites.
Escalate when needed
Opus 5.5
Try it for competing requirements, long-running work, difficult architecture, or judgment calls where the cost of a plausible but wrong answer is high.
Its higher API list rate and potentially heavier usage mean “use Opus for everything” is an expensive default without a measured benefit.
Ready to try the default? Open Claude and check the model label beside the message box; available choices can depend on your plan or administrator.
Which model fits the work in front of you?
Anthropic describes Sonnet 5.5 as a fast choice for everyday tasks and continues to position Opus 5.5 for long-running coding and knowledge work. Here is how I would apply that distinction to actual assignments:
- Client email from approved notes: start with Sonnet. Give it the audience, tone, facts that must survive, and what it must not promise. Check the final email yourself before sending.
- Sales deck or policy summary: start with Sonnet when the source material and decision are clear. Move to Opus if several sources disagree and you need a defensible reconciliation rather than a neat-looking summary.
- Bug fix versus system redesign: give Sonnet a reproducible bug and an acceptance test. An architecture choice that crosses services, security boundaries, and team ownership is a better candidate for Opus—and still needs human review.
The key is to judge the finished work, not the model name. A polished wrong answer is still wrong. For sensitive or high-stakes work, verify source facts, calculations, permissions, and final decisions outside the chat whichever model you select.
Lower effort can stretch your Claude allowance
Claude’s model menu also lets you change effort. Anthropic’s current help page says higher effort uses more tokens and can reach usage limits faster; Low or Medium can stretch usage on routine tasks. So a simple draft does not need the most intense setting just because the dial goes that far.
Anthropic also documents app model-switch behavior. If Claude changes models during a conversation, look at the model label before judging an answer. That app behavior is separate from choosing a model ID for an API integration.
The API price is clear. The completed-task bill is not.
If your team uses the Claude API—Claude built into software and billed by usage, rather than your chat subscription—the published standard list rates are $2 per million input tokens and $10 per million output tokens for Sonnet 5.5, versus $4 and $20 for Opus 5.5. Tokens are pieces of text the model processes or generates. Sonnet is half the list rate per token in this comparison. It does not make every successful task half the cost.
That distinction matters because an independent Artificial Analysis evaluation found strong Sonnet 5.5 benchmark performance at maximum effort but unusually high output-token use and task cost in that setting. Those tests used a pre-release deployment with a structured-output bug that Anthropic says it fixed before launch. One benchmark is not a verdict on your prompts. Compare completed work at the effort level you would actually use rather than multiply a token price by wishful thinking.
A quick test before you change your default
- Pick one representative task with a known brief, source material, and a result you can check. Keep confidential material within your organization’s approved AI workflow.
- Try Sonnet 5.5 first. Note whether it met every required fact, constraint, tone or acceptance test without multiple rescues.
- If it fails a consequential requirement, clarify the brief once. Then try Opus 5.5 on the same task and judge the corrected output, time, and effort used.
- Keep Sonnet as your default where the answer passes. Use Opus for the task types where it reliably changes the result enough to justify the extra consumption or API rate.
If both models need extensive correction, the brief, source material, or review process may be the real problem.
Your actual plan allowance, company model access, and best result still depend on your setup and tasks. Anthropic’s performance figures are vendor claims, while the independent evaluation covers a particular test and effort setting. For a separate cross-provider choice, see my Opus 5.5 versus GPT-6 Sol comparison.
Update log: September 29, 2026 — initial publication, using Anthropic’s September 28 release and current model/help documentation plus the independent September 28 evaluation. Material changes to availability, prices or the decision will be dated here.
Sources and method
I compared current official release, model and usage documentation with the independent evaluation, then built the task-choice and fixed-token examples above. I have not run a controlled Sonnet-versus-Opus benchmark.
- Anthropic, “Introducing Claude Sonnet 5.5” — published September 28, 2026; checked September 29.
- Anthropic, Sonnet 5.5 model overview and Opus 5.5 model overview — prices and limits checked September 29.
- Claude Help, model and effort settings and model-switch behavior — checked September 29.
- Artificial Analysis, Sonnet 5.5 evaluation — published September 28; checked September 29. Benchmark results are specific to the published test and settings.