Claude Opus 5.5

TL;DR Highlight

Opus 5.5 cuts cost 40% vs Opus 5, runs 30% faster, and leads coding agent benchmarks

Who Should Read

Backend developers weighing a model switch because coding agent and Claude API costs keep climbing

Core Mechanics

40% cheaper than Opus 5
- At default settings on typical workloads, it costs 40% less than Opus 5.
- Input and output tokens are $4 and $20 per million, 20% below Opus 5.
- Cache reads are $0.20 per million tokens, 60% below Opus 5.
Output is over 30% faster
- It generates output more than 30% faster than Opus 5.
- Fast mode in Claude Code and the Claude Platform is up to 2.5x faster at $8 input and $40 output per million tokens.
Tops agentic coding benchmarks
- On Terminal-Bench 4.0 it scores 66.4% vs Fable 5.1 at 55.8%, Opus 5 at 52.3%, and GPT-6 Astra at 57.9%.
- On CursorBench 4.0 it scores 57.8% vs GPT-5.6 Sol at 41.7%.
- Anthropic notes benchmark margins are a less reliable guide to real-world differences at this level.
Default effort beats Opus 5 at max
- On Terminal-Bench 4.0, default-effort Opus 5.5 beats max-effort Opus 5 for about a fifth of the cost.
- On FrontierCode at default effort (medium) it scores 54.6%, above GPT-6 Astra's best 53.3%, for about a fifth of the cost per task.

Community Reactions

Opinion from a user who stuck with 4.8 after Opus 5
- They found Opus 5 not worth the effort and used the May release 4.8 on the $20 tier.
- 4.8 became their grunt model for simple work, and it was their lowest Anthropic usage in two years.
Hope that the $20 tier becomes worthwhile
- One commenter suggests 5.5 might make the $20 subscription feel less of a joke.
- The raised usage limits and savable rate limit resets relate to that hope.

How to Apply

Rerun one existing agent task on 5.5
- Pick one coding agent task you run today and swap only the model to Opus 5.5.
- Record completion, token counts, and time side by side with the Opus 5 results.
Start at default effort and measure cost when raising it
- The comparisons use medium as the default effort, so check quality there first.
- Move up to high, xhigh, then max only if needed, and log cost at each level.
Recompute your cost sheet if you rely on caching
- Check the share of cache reads in your bill.
- Apply the drop from $0.50 to $0.20 per million and recalculate monthly cost.
Run an injection test before long autonomous runs
- Feed your web-reading agent a page containing malicious instructions.
- Compare how Opus 5 and 5.5 react and use it as a switching criterion.

Try It

copy & run

# Paste into Claude or Claude Code (quick eval prompt for comparing cost and quality before switching models)

I'm deciding whether to move my coding agent from Opus 5 to Opus 5.5.
I ran the same task on both models under the same conditions and am pasting the results.

Task description: {{task description}}
Opus 5 result: {{Opus 5 result summary}}
Opus 5.5 result: {{Opus 5.5 result summary}}
Opus 5 tokens/time: {{input tokens, output tokens, cache read tokens, elapsed time}}
Opus 5.5 tokens/time: {{input tokens, output tokens, cache read tokens, elapsed time}}

Prices (per 1M tokens)
Opus 5: cache read $0.50, input $5, output $25
Opus 5.5: cache read $0.20, input $4, output $20

Summarize in a table:
1. Cost per task comparison
2. Quality differences in the results
3. Whether to switch, with reasons

Next step: rerun the same task with different effort levels and compare how cost and quality change.

Terminology

effortA setting for how long the model thinks about a problem. Lower is faster and cheaper but may be less thorough on hard tasks.

cache readA cheap way to reuse a long prompt you already sent. Agents resend the same context repeatedly, so it makes up much of their cost.

prompt injectionAn attack that hides commands in web pages or documents to make the model misbehave. Think of a fake instruction slip tucked into the mail.

Terminal-BenchA benchmark that checks whether a model can finish complex multi-step professional tasks in a command line. It measures how well an agent gets work done in a shell.

EloA scoring method that ranks models by comparing them against each other. Higher numbers mean stronger.