Oleksandr Kuzmenko, EditorPublished 7 min readAI-assisted · editor-reviewed
TL;DR
- Prompt-cache read pricing dropped 60% to $0.20 per million tokens, heavily reducing recurring agent run costs.
- Opus 5.5 leads Terminal-Bench 4.0 at 66.4% and CursorBench 4.0 at 57.8%.
- A Fast mode option in Claude Code allows 2.5x generation speed at $8/$40 per million tokens.
Key facts
- Input Token Price
- $4.00 per 1M tokens
- Output Token Price
- $20.00 per 1M tokens
- Prompt Cache Read Price
- $0.20 per 1M tokens
- Fast Mode Input / Output Price
- $8.00 / $40.00 per 1M tokens
- Terminal-Bench 4.0 Score
- 66.4% (vendor-reported, xhigh effort)
- CursorBench 4.0 Score
- 57.8% (vendor-reported)
Source: Key facts — Input Token Price, Output Token Price, Prompt Cache Read Price
Source: Key facts. Input Token Price: $4.00 per 1M tokens. Output Token Price: $20.00 per 1M tokens. Prompt Cache Read Price: $0.20 per 1M tokens. The chart compares only values that share a unit. Each figure is copied from the story.
Metric chart
- $4.00 per 1M tokensInput Token Price
- $20.00 per 1M tokensOutput Token Price
- $0.20 per 1M tokensPrompt Cache Read Price
- Input Token Price$4.00 per 1M tokens
- Output Token Price$20.00 per 1M tokens
- Prompt Cache Read Price$0.20 per 1M tokens
Pricing and Performance Adjustments
Anthropic has made Claude Opus 5.5 available across its API and interactive platforms. Token pricing sits at $4 per million input tokens, $20 per million output tokens, and $0.20 per million cache read tokens. Compared to Opus 5, base pricing represents a 20% drop for standard tokens and a 60% reduction for prompt-caching operations, resulting in an estimated 40% net cost drop across autonomous agent workflows. A dedicated Fast mode accessible through Claude Code provides up to 2.5x token throughput at $8 per million input and $40 per million output tokens.
Agentic Benchmarks and Real-World Audits
On the Terminal-Bench 4.0 autonomous coding benchmark, Opus 5.5 achieved a 66.4% success rate at extra-high thinking effort, outpacing GPT-6 Astra (57.9%) and Opus 5 (52.3%). It scored 57.8% on CursorBench 4.0 and 54.4% on FrontierCode v1.1. In empirical refactoring tests, an early partner audited and repaired a 200,000-line repository in under 3 hours, consuming 60% fewer tokens than Opus 5 required for equivalent tasks. Anthropic also translated the C codebase of HAProxy to Rust in 9.5 hours, passing regression test suites.
Developer Safeguards and Tier Upgrades
Anthropic reports that Opus 5.5 is more resistant to prompt injection and less likely to perform destructive file operations without explicit confirmation. Pro, Max, Team, and seat-based Enterprise plans receive expanded five-hour window quotas alongside a manual rate limit reset that users can trigger on demand. High-risk cybersecurity and biological evaluation tooling continues to be gated behind verification programs.
Try it in 2 minutes
claude --model claude-opus-5-5 --thinking-effort highWhen to use
- Repository-scale migrations, large refactors, and complex architectural audits requiring sustained multi-step reasoning.
- Agentic loops inside Claude Code and Cursor where extensive prompt caching keeps recurring input overhead minimal.
When NOT to use
- Lightweight single-file script completions where smaller models like Haiku or Sonnet deliver sufficient accuracy at lower absolute cost.
- Budget-constrained bulk generation jobs that do not benefit from prompt caching or extensive thinking time.
What to do today
- Switch agent configurations in Claude Code and Cursor to test Claude Opus 5.5 against your existing test suites.
- Review prompt-caching architectures to take advantage of the $0.20/M token read price on repetitive context blocks.
- Test Claude Code Fast mode on time-critical interactive refactoring loops.
Step names are taken from What to do today.
Step names are taken from What to do today. Switch agent configurations in Claude Code and Cursor to test Claude Opus 5.5 against your existing test suites.. Review prompt-caching architectures to take advantage of the $0.20/M token read price on repetitive context blocks.. Test Claude Code Fast mode on time-critical interactive refactoring loops.
- Switch agent configurations in Claude Code and Cursor to test Claude Opus 5.5 against your existing test suites.
- Review prompt-caching architectures to take advantage of the $0.20/M token read price on repetitive context blocks.
- Test Claude Code Fast mode on time-critical interactive refactoring loops.
What the community says
Cool maybe this makes the 20$ sub less of a joke. I deemed 5.0 unworthy of spending time fighting with and personally considered it by far the worst release of 2026 by either of the two major labs.
Real-world use cases typically involve big, hairy, crufty, tech debt laden codebases and benchmarks do not. Success in a benchmark essentially boils down to do the tests pass.