Most people do not need the most powerful AI model for every task. They need one that is good enough to finish ordinary work quickly without turning every email, spreadsheet or slide deck into a premium-model decision.
That is the useful part of Anthropic’s new Claude Sonnet 5.5, released September 28. Anthropic says it runs more than 30% faster than Sonnet 5 and costs up to 30% less per task in its testing, while keeping the same API pricing. It is positioned for well-scoped everyday work, coding, and polished documents, slides and spreadsheets.
The useful rule
Use Sonnet for work you can describe clearly. Move up to Opus when the hard part is judgement, ambiguity or a long chain of decisions.
Start with these five jobs
1. Turn rough notes into a finished document
Give Sonnet the notes, audience and desired outcome. Ask for a finished first draft rather than a generic outline. This is a good fit for briefs, internal updates, proposals and meeting summaries where the task is clear and you can review the result.
2. Build or clean up a spreadsheet
Use it when you already know what the spreadsheet needs to do: organise a tracker, structure data, create formulas, clean categories or turn a recurring report into a reusable template.
3. Create a first-pass slide deck
Specify the audience, number of slides, decision the deck needs to support and source material. The model’s job is to get you from blank page to something reviewable, not to make the final strategic call.
4. Fix a well-scoped coding problem
Anthropic specifically positions Sonnet 5.5 for fixing bugs and everyday coding. The key phrase is well scoped: provide the failing behaviour, relevant files, expected result and constraints instead of asking it to wander through an entire codebase.
5. Iterate quickly
Speed matters when the workflow is conversational: rewrite this paragraph, adjust this table, change the tone, fix this function, try another layout. A faster model can be more useful than a stronger model when you expect several short review cycles.
The 30-second model test
Before choosing a model, ask yourself:
- Can I describe the desired output precisely? Start with Sonnet.
- Can I easily check whether the result is right? Start with Sonnet.
- Does the task involve conflicting evidence, major ambiguity or consequential judgement? Consider Opus.
- Am I repeating this hundreds or thousands of times? Cost and speed matter more; test Sonnet and, when available, the lighter Haiku tier.
A prompt that makes the choice for you
Copy this
I need to complete this task: [describe task].
Before doing it, classify the work as one of these:
A. well-scoped execution
B. iterative creation
C. complex judgement
D. high-volume repetitive work
Tell me which model tier is appropriate and why. Prefer the fastest/lower-cost tier that can reliably complete the task. Only recommend a more capable model when you can identify the specific part of the task that requires it.
Then give me the shortest workflow to complete the task and the checks I should make before using the output.
Where the bigger model still earns its place
Do not turn “Sonnet is faster” into “Sonnet should do everything.” Anthropic positions Opus 5.5 for complex work requiring careful judgement. That distinction matters for ambiguous strategy, difficult research synthesis, consequential decisions and long tasks where several uncertain steps depend on one another.
For those jobs, the cost of a weak decision can matter far more than the cost of extra tokens.
Test the upgrade on work you already know
A new model is easiest to evaluate on a task where you know what a strong answer looks like. Reuse a difficult research brief, spreadsheet analysis, coding problem or long document you have already completed. Give the new model the same inputs and constraints, then compare missed evidence, correction time and how much of the output survives your review.
This avoids the novelty trap. A polished answer to a new prompt can feel impressive because you have no baseline. A familiar task exposes whether the model is actually more useful for your work.
Where the human should stay
Model choice does not remove the need for review. Keep human ownership over factual claims, financial or legal decisions, external publishing, sensitive data, important client communication and anything where a plausible-looking mistake would be difficult to detect.
The bigger shift is not another benchmark
The interesting part of Sonnet 5.5 is not that another model moved up a leaderboard. It is that capable AI work is becoming cheaper and faster enough that the sensible default may increasingly be use the smallest model that reliably finishes the job.
That is a more useful habit than automatically clicking the model with the biggest name.
Sources
Anthropic – Introducing Claude Sonnet 5.5, September 28, 2026
