Anthropic launches a new AI model with Claude Opus 5 that approaches the performance of the top model Fable 5 – at half the price. The company maintains the costs of the predecessor Opus 4.8: five dollars per million input tokens, 25 dollars per million output tokens. Opus 5 replaces Opus 4.8 as the standard model in the Claude Max subscription.
Opus 5 costs half of Fable 5
Anthropic has not changed the prices for Opus 5 compared to the predecessor model Opus 4.8. According to the official price list of the company, one million input tokens cost five dollars, and one million output tokens cost 25 dollars. Fable 5, which remains the most powerful model in the lineup, charges exactly double:
- Claude Opus 5: 5 dollars per million input tokens, 25 dollars per million output tokens
- Claude Fable 5: 10 dollars per million input tokens, 50 dollars per million output tokens
- Batch processing Opus 5: 2.50 dollars per million input tokens, 12.50 dollars per million output tokens
- Batch processing Fable 5: 5 dollars per million input tokens, 25 dollars per million output tokens
For cached inputs (prompt caching), both models carry the usual surcharge of 25 percent for writing and a 90 percent discount for reading from the cache. Opus 5 is available via the Claude API under the model identifier claude-opus-5, as well as in Claude.ai, Claude Code, and Claude Cowork. Across the subscription plans, it replaces Opus 4.8 as the standard model in Claude Max and remains the strongest option in the cheaper Claude Pro tier.
Effort levels control speed and cost
A central new feature of Opus 5 is configurable effort levels. According to a report by Fortune, users can choose between low, medium, and high effort, allowing them to weigh response speed, token consumption, and result quality against each other. At low effort, the model saves tokens and responds faster; at high effort, it checks intermediate results more thoroughly and works longer on a task. Anthropic also promotes Opus 5 as a model that operates more independently than its predecessors: it is meant to recognize and correct errors in its own intermediate steps without users having to steer every step individually.
Two further additions target developers. The Claude API now lets developers change the available tools in the middle of an ongoing conversation without invalidating the existing prompt cache – previously, any tool change forced an expensive rebuild of the cache. If the model declines a request for safety reasons, the API also automatically routes the request to another model instead of simply returning an error. Both features are aimed mainly at long-running agentic workflows in which tools and context change frequently.
Fable 5 remains ahead on benchmarks
In its own benchmark tests, Anthropic reports significant progress over Opus 4.8: on the Frontier-Bench v0.1 test, Opus 5’s score more than doubles while costing less per task. On the coding benchmark CursorBench, Opus 5 trails Fable 5 by only about 0.5 percentage points – at half the price. On complex problem-solving tasks in the ARC-AGI-3 test, Opus 5 reportedly reaches three times the score of Opus 4.8, while gains on chemistry and protein-analysis tasks range between 7.7 and 10.2 percentage points. These figures come from Anthropic’s own announcement and are independently unverified.
Opus 5 does not lead in every discipline: on tasks involving finding and exploiting security vulnerabilities, it ranks behind Mythos 5, the company’s model built specifically for such tasks. For autonomous tasks spanning multiple days, Anthropic still recommends Fable 5. That model has had a turbulent few weeks: after a temporary shutdown under US export controls in June, Anthropic permanently tied access to Fable 5 to the pricier Max and Team Premium tiers starting July 20. The lower price of Opus 5 is likely also a response to mounting competitive pressure: just at the start of July, SpaceXAI introduced Grok 4.5, a coding model that significantly undercut Opus 4.8 on price.
What matters now is whether companies and developers actually act on Opus 5’s price advantage or stick with Fable 5 for demanding work despite the higher cost. With the new effort levels, Anthropic puts the trade-off between speed, cost, and quality directly into users’ hands for the first time, rather than leaving it entirely to model choice. How much that dents demand for the pricier Fable 5 will likely only become clear over the coming weeks, once usage data comes in.
