Claude Opus 5.5 is Anthropic’s newest Opus model, released September 22, 2026. It costs less than Opus 5, runs faster, and Anthropic’s own tests put it ahead of Fable 5.1, the more expensive model above it. This post lays out what changed, in plain words, with every number linked to where it came from.
Everything here comes from Anthropic’s announcements and pricing page, read on launch day. No independent testing is included. Treat the benchmark numbers as Anthropic’s claims, not independent results.
Three words you need first
- Token: the unit AI models read and write in. Roughly three quarters of a word.
- Per million tokens: how developers are billed. “$4 input” means $4 for every million tokens you send the model.
- Benchmark: a standard test, like an exam, used to compare models. Higher is better, but it is still a test, not your actual work.
Who is in this comparison?
Anthropic sells several Claude models at different price points. The four that matter here:
- Fable 5.1 (September 1, 2026): the most capable model the public can use. Built for long, hard jobs that run for hours. Also the most expensive.
- Opus 5.5 (September 22, 2026): the new one. Replaces Opus 5.
- Opus 5 (July 24, 2026): the model Opus 5.5 replaces.
- Sonnet 5 (June 30, 2026): the cheaper everyday model, and the default for Free and Pro users.
What does each one cost?
| Model | Input (per million tokens) | Output (per million tokens) |
|---|---|---|
| Fable 5.1 | $10 | $50 |
| Opus 5 | $5 | $25 |
| Opus 5.5 | $4 | $20 |
| Sonnet 5 | $2 | $10 |
Opus 5.5 costs 20% less per token than Opus 5. Anthropic says it comes out about 40% cheaper on typical workloads, because it also uses fewer tokens to get the same job done and cached (reused) input dropped from $0.50 to $0.20 per million. It is 40% of Fable 5.1’s per-token price.
If you pay a monthly subscription (Pro, Max, Team) instead of per token, these prices do not show up on your bill. What changes for you is how much work fits in your usage limits. Anthropic says it raised the five-hour usage limits across plans with this release. I wrote more about where subscription usage actually goes in Where Claude tokens go.
What changed from Opus 5 to Opus 5.5?
- Cheaper: see the table above.
- Faster: Anthropic says it writes its output more than 30% faster than Opus 5.
- Better on Anthropic’s tests: higher scores on every benchmark in the launch table (below).
- Plainer writing: Anthropic says it now puts the most important information first, uses less jargon, and follows writing instructions more closely. This targets a common complaint that Claude’s writing sounds formulaic.
- Safety: Anthropic reports its best score yet on its own automated behavior audit, with outside testing by METR among others.
How does Opus 5.5 compare to Fable 5.1?
On Anthropic’s launch table, Opus 5.5 scores higher than Fable 5.1 on all six listed tests, at 40% of the price.
| Test (what it measures) | Opus 5.5 | Fable 5.1 | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 (working in a command line) | 66.4% | 55.8% | 52.3% |
| FrontierCode v1.1 (hard coding tasks) | 54.4% | 50.3% | 48.0% |
| CursorBench 4.0 (coding inside an editor) | 57.8% | 51.8% | 46.6% |
| OSWorld 2.0 (using a computer like a person) | 81.8% | 80.7% | 74.0% |
| Chartography (reading charts) | 89.0% | 88.4% | 83.4% |
| GDPval-AA v2.1 (office work, Elo rating) | 1846 | 1735 | 1708 |
Two cautions. The gap is large on coding (Terminal-Bench: nearly 11 points) and small on computer use and charts (about 1 point), which is close enough that real-world results could go either way. And Fable is still pitched for a different job: very long, hands-off tasks that run for hours. A benchmark table does not measure that well.
Where does Sonnet 5 fit?
Sonnet 5 is half the price of Opus 5.5 and is the default model for Free and Pro users. Anthropic described it at launch as close to Opus 4.8, an earlier Opus model.
There is no fair score-for-score comparison with Opus 5.5 yet. Sonnet 5 was tested on older versions of the benchmarks (Terminal-Bench 2.1, not 4.0), so the numbers are not on the same scale. Putting them in one table would be misleading.
Anthropic also says Sonnet 5.5 and Haiku 5.5 are coming “in the coming weeks.” If you mostly use Sonnet, the more useful comparison is probably a few weeks away.
Which one should you use?
- Everyday questions, writing, light coding: Sonnet 5. Cheapest, and the default anyway.
- Harder coding, longer projects, work where mistakes cost time: Opus 5.5. It is cheaper and scores higher than Opus 5, so for most people it simply replaces it.
- Very long unattended jobs: Fable 5.1 is still the model built for that, but on Anthropic’s numbers, try Opus 5.5 first.
What is not known yet?
- Independent results. Every score above is Anthropic’s. Outside benchmarks usually appear within days to weeks.
- How the “40% cheaper” holds up. That figure depends on the kind of work. Your savings may be bigger or smaller.
- Sonnet 5.5 and Haiku 5.5. Announced, not released.
- Some requests go to an older model. Like Fable 5.1, Opus 5.5 routes certain cybersecurity and biology requests to a less capable model as a safety measure. Most people will never hit this.
Sources
- Introducing Claude Opus 5.5 (Anthropic, Sept 22, 2026)
- Claude pricing (Anthropic docs)
- Claude Fable (Anthropic)
- Introducing Claude Opus 5 (Anthropic, July 24, 2026)
- Introducing Claude Sonnet 5 (Anthropic, June 30, 2026)
- The Decoder’s launch coverage