Claude Opus 5.5 in 11 slides

Swipe or use the arrow keys. Hover the cover to tilt it, click it to replay. Try the calculator, effort dial, prompt kit and breaking changes.

Anthropic release, 22 Sep 2026
Claude Opus
5.5
Fable-level work.
40% cheaper than Opus 5.
Swipe
Pricing per 1M tokens

Every token
got cheaper

Input−20%
$5.00
$4.00
Output−20%
$25.00
$20.00
Cache reads−60%
$0.50
$0.20
Opus 5Opus 5.5

Cache writes $5 (was $6.25). Fast mode: $8 in / $40 out for up to 2.5× speed.

Drag the sliders

What would
you pay?

Opus 5
Opus 5.5

List prices, millions of tokens a month. Excludes fast mode and batch. Opus 5.5 also uses fewer tokens per task, so real savings run higher.

Token consumption
−40%
cost per task,
fewer tokens too
Box tokens vs Opus 5
33%
Kiro tokens vs Opus 5
50%
Rogo output tokens vs Opus 5
40%
200k-line audit tokens vs Opus 5
40%

Full track = what Opus 5 used on the same job. Early-tester reports. Output also streams 30%+ faster.

Tap a level on the dial

Effort is
your new dial

low medium high xhigh max default
Thinking is always onCan't be disabled. Effort sets depth and cost.
medium is the defaultMatches or beats Opus 5 at high on coding and knowledge work.
Use cases

Where it shines

680k lines
Migrated in under a day
72%
Known bugs caught in review at low effort
1846 Elo
Knowledge work on GDPval‑AA
81.8%
Computer use on OSWorld 2.0
How to prompt Opus 5.5

Prompt less.
Tune effort.

Drop “think step by step”

It decides how much to think. Replies start sooner.

Give agents a checklist

A text-only end of turn is a report, not “done”.

Name what to avoid

“No cream background, no pill buttons” beats “don't look generic”.

Tag pasted text

Wrap it in tags so hidden instructions aren't followed.

effort: "medium", // set it, then sweep max_tokens: 128000, // thinking counts here
How to prompt Opus 5.5, part 2

Two quiet traps

Changing effort resets the cache

effort changed cached cache lost

Fix: change effort per message (beta). The cache stays.

Asking for its reasoning can be refused

Write out your reasoning refusal

Fix: set display: "summarized" and read the thinking blocks.

Pick the problem you see

Prompt kit


    
    

  

  
Written from the patterns in Anthropic's “Prompting Claude Opus 5.5” guide. Official wording at platform.claude.com.







  
Tap a change to see the fix

  
Four breaking
changes

  

    



  







  
Available today

  
Same envelope.
Better engine.

  

    
1Mcontext tokens

    
128kmax output

    
30%+faster output

  

  
claude-opus-5-5API, AWS, GCP, Azure

  
What would you
hand it first?

  
Sources: anthropic.com/claude-opus-5-5 and the Opus 5.5 docs on platform.claude.com. Tester numbers are self-reported.