Claude Opus 5: pricing, benchmarks, how it differs from Sonnet 5
Anthropic released Opus 5 on July 24, 2026, and sells it in one sentence: nearly the same intelligence as Fable 5, at half the price. I've gathered everything you need to know before switching to it, along with the numbers and places where the more expensive model still wins.
as of: 11 November 2026
Same rate as Opus 4.8 and half the price of Fable 5 ($10/50). Fast mode: about 2.5× faster at double the rate, only in Anthropic's API. Standard response limit is 128K tokens, in batch mode up to 300K with beta header enabled.
Benchmarki
| Test | Opus 5 | Fable 5 | Opus 4.8 | co mierzy |
|---|---|---|---|---|
| SWE-bench Pro | 79.2% | 80.3% | 69.2% | coding on real repositories |
| Frontier-Bench v0.1 | 43.3% | 33.7% | 21.1% | tasks at the edge of model capabilities |
| ARC-AGI-3 | 30.2% | - | 1.5% | abstract reasoning |
| GDPval-AA v2 | 1861 | 1747 | - | economically valuable work |
| OSWorld 2.0 | 70.6 | - | - | computer operation |
| BrowseComp | 90.8% | - | - | web search |
All numbers come from Anthropic's launch materials, meaning from the manufacturer, not from independent measurement. The Frontier-Bench result was obtained by the company's own run on mini-SWE-agent, five attempts per task. A dash means the manufacturer did not provide a result for that model.
What this means in practice
Fable 5 scores slightly higher on SWE-bench Pro and remains Anthropic's recommendation for multi-day agent work. Opus 5 wins where cost matters: at half the rate, it lands within 0.5 percent of Fable's peak score on CursorBench 3.2. On top of that comes the effort slider, set separately for each query, so real cost depends on how much you let the model think, not just on token price. How much that works out to in USD on specific tasks, I calculate in cost calculator, updated daily. A full compendium of the more expensive model, along with the timeline of its withdrawal and return, I maintain in the Fable 5 hub.
Frequently Asked Questions
How much does Claude Opus 5 cost?
$5 per million input tokens and $25 per million output tokens. That's the same price as Opus 4.8 and half the price of Fable 5 (10/50). Fast mode works about 2.5 times faster at double the rate and is only available in Anthropic's API.
Is Opus 5 better than Fable 5?
Not in every task. Fable 5 is marginally higher in SWE-bench Pro (80.3 to 79.2 percent) and remains Anthropic's recommendation for multi-day agent work. Opus 5 leads in GDPval-AA v2 and OSWorld 2.0, and costs half as much as Fable, so as an everyday model it performs better.
What does the effort slider do?
It's a per-query setting: low, medium or high reasoning effort. That's why the cost depends more on the task than on the token rate. Anthropic reports that on CursorBench 3.2, Opus 5 comes within 0.5 percent of Fable 5's peak score while spending half as much per task.
What context window does Opus 5 have?
One million tokens, with a standard limit of 128 thousand output tokens. In batch mode (Message Batches) you can reach 300 thousand output tokens by enabling the beta header. A large window doesn't mean perfect memory though, worth testing with your own documents.
Where does Opus 5 work?
In the API as claude-opus-5, in the Claude app (default model on the Max plan) and in Claude Code. It launched on July 24, 2026, less than two months after Mythos 5, Fable 5 and Sonnet 5.
// follow
New tests and price changes, straight to your feed reader
Free RSS, no account and no email needed.
Follow Quantslant →Sources: Anthropic - launch materials · model documentation · API pricing