Now Anthropic offers three “frontier-grade” versions of Claude, and this nomenclature is not kind to you. The Sonnet 5.5 became available on September 28. The Opus 5.5 is above it. The Fable 5.1 is above it, falling into the Mythos line. You have the impulse to go for the biggest brand. But your bill tells you otherwise. Let’s see which one you should pay for.
Also read: Apple’s Vision Pro, Samsung’s Android XR, Meta VR Glasses: The battle for the future of XR
Sonnet 5.5 has been designed to perform well-scoped tasks such as bug fixes, formatting documents, slide shows, and spreadsheets. Sonnet 5.5 has a cost of $2 for each million tokens of input and $10 for each million tokens of output. Sonnet 5.5 produces output 30% faster compared to Sonnet 5. Anthropic claims that the cost is reduced up to 30% due to less token use. Box, an early tester, found it to be 2.4 times faster than its predecessor. It also continues the Claude tradition of beating a Pokemon game – this time Pokemon Red while working only from screenshots.
Now the awkward bit for Opus. According to Terminal-Bench 4.0, an agent coding benchmark, Sonnet 5.5 outperforms Opus 5.5 with 70.6% versus its best 66.4%. In the GDPval-AA test that measures practical performance across 44 occupations, it scores 1844 versus Opus’s 1846. This is barely a rounding error. On CursorBench, it lags behind by a margin of two points. At Medium effort, which is the default effort level on the Claude applications, it wins in Terminal-Bench performance compared to Sonnet 5 by less than a tenth of the price per task. Sonnet is the default choice. There is one drawback, though – it has new cybersecurity measures.
Opus is twice as expensive: $4 per million input tokens, $20 per million output tokens. The cost of cache reads is the same for both at $0.20 per million, making cache-intensive models closer. So, what do you get for the additional investment? According to Anthropic, the choice is clear – Opus 5.5 is better when it comes to solving complex tasks that require judgment, outscoring Sonnet across the majority of public benchmarks. On FrontierCode, it scores 54.4% compared to Sonnet’s best 52.1%.
Also read: AI Notetaker to AI Teammate: Fireflies CEO on what people are handing over to AI
Humanity’s Last Exam with tools gives 67.7% versus Sonnet’s 64.5%. Think about it in terms of labour division – Sonnet implements the strategy, while Opus comes up with it. One early tester at Creator said to let Opus set the architecture and trust Sonnet to build it. That pairing costs less than running Opus on everything.
Fable 5.1 falls under the Anthropic’s Mythos category. It is based on the same model as Mythos 5.1 but has additional safeguards for biology, cybersecurity, and LLM research and development. Its pricing via the API is $10 per million input tokens and $50 per million output tokens, which is 2.5 times Opus 5.5 and five times Sonnet 5.5. The price for cache reads is $0.25 per million, and the context window is 1M tokens. According to Cognition, Devin will move its Opus 5 traffic to Fable 5.1 on the day of the launch due to cache pricing.
And now to the nitty-gritty stuff. Since Fable is a Covered Model, companies who do not retain data require explicit permission from Anthropic, and it is excluded from the Priority Tier. Fable and Mythos availability have been inconsistent as well; access was suspended on June 12 due to export controls in the United States and resumed on July 1.
Begin with Sonnet 5.5. It handles all kinds of routine work and is cheapest. Switch to Opus 5.5 whenever a job goes wrong because the model got confused, but not because it was running too slowly. Look for Fable 5.1 when Opus gets stuck on your most challenging problems and your data policy permits it. Indian startups and freelancers that earn in rupees but pay in dollars have a fivefold price difference between a budget line item and a footnote.
Effort settings should be just as important as the models. Anthropic says that Sonnet 5.5 at Low or Medium effort will outperform the best of Sonnet 5 on multiple benchmarks for one-tenth the cost. Adjust effort first before upgrading the model.
Also read: Deploying voice AI agents at scale: Tata Comms, TTBS target India’s SMBs