Anthropic dropped Claude Opus 5.5 on September 24, and the first thing you'll notice is not the benchmark scores — it's the price. Cache reads are $0.20 per million tokens, 60% cheaper than Opus 5's $0.50. Inputs are $4 and outputs $20 per million, down 20% from $5 and $25. Since cache reads make up the majority of agentic and coding costs, that's where the real money moves.
But the headline number is the net effect: Opus 5.5 costs 40% less to run than Opus 5 on typical workloads, and it uses fewer tokens per task. On FrontierCode v1.1, at default effort, it beats GPT-6 Astra's top score for about a fifth of the cost per task. On Terminal-Bench 4.0 it matches Astra for about 40% of the cost. That's not a small adjustment to the calculus — it's a reset.
The Price Table
| Prices per 1M tokens | Claude Opus 5.5 | Claude Opus 5 |
|---|---|---|
| Cache reads | $0.20 | $0.50 |
| Input tokens | $4 | $5 |
| Output tokens | $20 | $25 |
| Cache writes | $5 | $6.25 |
Fast mode is also available in Claude Code and the Platform at 2.5x speed, at $8 input / $40 output. The 40% drop comes from two directions: cheaper per-token pricing plus fewer tokens per task. GitHub's CPO noted that in their testing, Opus 5.5 solved more terminal tasks than Opus 5 in less than half the steps. Optiver's team says it matched Opus 5's quality in about half the turns and output tokens, cutting workload cost by 40–50%.
What Changed: The Benchmarks
Here's where Opus 5.5 lands against the current frontier, with the results table from Anthropic's table:
| Benchmark | Opus 5.5 | Fable 5.1 | Opus 5 | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 55.8% | 52.3% | 57.9% | 37.3% |
| FrontierCode v1.1 | 54.4% | 50.3% | 48.0% | 53.3% | 47.5% |
| CursorBench 4.0 | 57.8% | 51.8% | 46.6% | — | 41.7% |
| GDPval-AA v2.1 | 1846 Elo | 1,735 | 1,708 | 1,542 | 1,588 |
| AutomationBench | 40.0% | 31.4% | 26.9% | 41.4% | 28.8% |
| OSWorld 2.0 | 81.8% | 80.7% | 74.0% | — | — |
Anthropic itself cautions that at this level, benchmark margins are a less reliable guide to real-world differences — in their own use, the gap to Fable 5.1 is narrower than the scores suggest. The lead that matters is efficiency. On the cost-per-task charts, Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost. On FrontierCode, default-effort Opus 5.5 scores 54.6%, beating Astra's top score (53.3) at a fifth of the cost per task.
Why This Matters for a Developer
I'd reached for smaller models or mid-tier Claude tiers for most agent work because the per-step cost of high-effort Opus was hard to justify for long-running tasks. Now, agents that run for hours — codebase migrations, overnight refactors, multi-file audits — have a lot of the cost pressure removed.
A concrete data point: one early tester audited and fixed a 200,000-line codebase in under three hours; Opus 5 took over 20 hours and used 2.5x as many tokens. Another completed a 680,000-line code migration in less than a day. When you're running Claude Code unattended for 18 hours, those numbers are the difference between "an experiment" and "everyone's default workflow."
The Catch: Safeguards and New Speed Limits
Anthropic is wrapping Opus 5.5 in Fable-class safeguards. It's comparable to Claude Mythos 5.1 in biology and cybersecurity, so most cybersecurity tasks are re-routed to Opus 4.8. If you need to bypass those for legitimate research, you can apply to the Life Sciences Verification Program or wait for the expanded Cyber Verification Program. The safeguards are coupled to output — on the benchmark runs, when safeguards intervened, cybersecurity tasks were completed by Opus 4.8, which suppresses the headline scores.
Also: preserved thinking is now active. This means API users can't edit Claude's prior context in an attempt to extract its reasoning. It applies to accounts created on or after August 31, 2026. And "thinking" mode can no longer be switched off — the ability to the difference with your own agent if you're used to faster iterations at lower effort. Also, a rate-limit reset is now available, which users can save and use whenever they choose.
What This Means for Your Toolchain
If you've been routing your coding agents through Opus or Sonnet based on cost, this release muddies the line. Opus 5.5 outperforms Opus 5 on every benchmark, while you're paying less. That's the model, not the wrapper — it's worth comparing it against your own workload to see if I've any local models should lose their job.
Next Steps and Availability
Opus 5.5 is live now on all platforms, including AWS, Google Cloud, and Azure. On the Claude Platform, you start by pointing at claude-opus-5-5. Anthropic says Sonnet 5.5 and Haiku 5.5 follow in "coming weeks," so this is an era is just beginning.
Sources:
- Anthropic: Claude Opus 5.5 announcement
- Anthropic: Opus 5.5 documentation
- Anthropic: preserved thinking docs
- Zipping: 4-std dev notes on Opus 5.5 benchmarks
_Note: This is wider than canonical versions, and the exact limits of the number of Anthropic releases can be verified scenarios are went over.
** - Anthropic. **Verify details.
End (relationship to earlier coverage: I'm new to this site, so we have a past round on GPT-6 Astra's frontier capability shift that starts the pattern — but this release resets it.
Sources: