Anthropic Cut Opus's Price From $25 to $20 a Million Tokens, Then Made It Your Default Claude Model
The per-token price drop is real on its own terms. Input fell from $5 to $4 per million tokens and output from $25 to $20, per TechCrunch's coverage of the launch. Cached tokens got the steeper cut: reading from the prompt cache now costs $0.20 per million tokens and writing to it costs $5, down from $6.25, per Anthropic.
Anthropic's bigger claim is that Opus 5.5 'performs at the level of Claude Fable 5.1 on most work,' its larger and pricier Mythos-tier model, per Anthropic's own announcement. On Terminal-Bench 4.0, a benchmark that scores models on real terminal-based coding tasks, Opus 5.5 hit 66.4% against Fable 5.1's 55.8%. On GDPval-AA v2.1, which grades models against real professional deliverables, it scored 1846 Elo against Fable 5.1's 1735.
| Claude Opus 5 | Claude Opus 5.5 | |
|---|---|---|
| Input | $5 / MTok | $4 / MTok |
| Output | $25 / MTok | $20 / MTok |
| Cache write | $6.25 / MTok | $5 / MTok |
| Cache read | higher legacy rate | $0.20 / MTok |
| Fast mode | not offered | $8 / $40 per MTok (in/out) |
| Default reasoning effort | High | Medium |
Source: Anthropic's Opus 5.5 pricing page and TechCrunch's report on the old Opus 5 rate.
The catch in the '40% cheaper' number
Anthropic's savings figure compares Opus 5 running at its old default reasoning effort, high, against Opus 5.5 running at its new default, medium. That's a fair comparison for most everyday use, since medium is what the default experience now delivers. But teams that built applications pinning Opus 5 explicitly to high effort don't get the full 40% just by switching models, and some existing integrations reportedly threw errors on the switch before developers updated their code to match the new defaults, per an independent review of the release. If your own API calls specify an effort level, check it before assuming the discount applies automatically.
If you just use Claude, you already have it
There's nothing to opt into. Opus 5.5 replaced Opus 5 as the selected model across Claude.ai, the mobile apps, Claude Code, and Cowork the day it shipped, per Anthropic's own system-prompt release notes. If you haven't manually picked an older model, you're already talking to Opus 5.5, and paying whatever your existing plan already charges.
Tested outside the building before it shipped
Anthropic says Opus 5.5 'was tested before release by external evaluators, including Frontier Design and METR,' and scored its best result yet on the company's own automated behavioral audit, per Anthropic. That's a concrete, dated instance of the outside-evaluator access CEO Dario Amodei pledged in a September 12 essay calling on the AI industry to slow its pace and let independent researchers inside frontier labs. It's worth noting as a real step, not a settled one: how independent any given evaluator actually is from the lab funding the work is its own live argument, and one worth watching rather than taking on faith.
Do I need to do anything to use Claude Opus 5.5?
No. It replaced Opus 5 as the default selected model in Claude.ai, the Claude iOS and Android apps, Claude Code, and Cowork on September 22, 2026. If you haven't manually chosen an older model, you're already using it.
How much cheaper is it, really?
The per-token price dropped a genuine 20% on output ($25 to $20 per million tokens) and 20% on input ($5 to $4). Anthropic's larger '40% cheaper to run' figure also factors in a lower default reasoning effort, medium instead of Opus 5's old default of high, which uses fewer tokens per answer.
Why might my app behave differently after the switch?
If your integration explicitly set Opus 5 to a specific effort level, that setting doesn't carry over automatically, and some existing calls reportedly errored until developers updated them to match Opus 5.5's new defaults.
How does it compare to Claude Fable 5.1?
Anthropic says Opus 5.5 performs at the level of Fable 5.1, its larger and more expensive Mythos-tier model, on most work. On Terminal-Bench 4.0 it scored 66.4% against Fable 5.1's 55.8%.
What is METR and why does it matter here?
METR is an independent AI-evaluation organization. Anthropic says Opus 5.5 was tested by METR and Frontier Design before release, a concrete example of the outside-evaluator access CEO Dario Amodei publicly pledged in his September 12 essay calling for the AI industry to slow down.
Try the tools ✦
Free browser tools that never upload your files.