Sonnet 5 by Anthropic, the latest version of Sonnet 5 that Anthropic has rolled out, may very well be a thorny issue for those people who are currently using Claude Opus. Sonnet 5 has become the default model for Free and Pro and Anthropic has advertised this version as being the most agentic Sonnet they have developed yet, bridging a significant part of the gap with respect to Opus 4.8 on benchmarking of coding, reasoning, and tool usage capabilities. This was done without costing as much as half the amount it would cost with Opus 4.8.
Also read: Fable 5, Mythos 5 coming back: Anthropic still hasn’t answered key question
Numbers tell the story clearly. Sonnet 5 will come in at an introductory price point of $2 for million input tokens and $10 for million output tokens, and rise to $3 and $15 respectively post August 31st.
However, the change isn’t just about the benchmark scores. According to Anthropic, Sonnet 5 handles multi-turn reasoning tasks more independently by creating tests of its own, fixing problems, and otherwise doing what an agent does when given a task. This is the key point when considering the benefits of the new version for both developers and users of Claude Code, as it makes a big difference whether you have a system that will respond well to your prompts, or one that can be delegated some work.
The developers claim that Sonnet 5 also suffers from fewer hallucinations and sycophancies, as well as better protection against prompt injection.
Also read: Not thinking about storage in AI is a mistake: Dell’s Venkat Sitaram
And what about Opus 4.8’s premium? To be frank, it depends on your use cases. Opus is always better when you need to solve the toughest problems requiring the highest capabilities, and when the marginal improvement is worth its higher cost no matter how expensive it is – the tasks where you don’t want anything less than the ceiling. However, in the majority of routine programming, editing and agent use cases, where you don’t really care whether your tool is one of the best on the market, Sonnet 5 being this close in terms of quality at this price is going to be difficult to ignore for many organizations.
There is one important caveat for everyone planning API expenses. Sonnet 5 uses a new tokenizer, so it will charge 1.35x more tokens depending on input data, reducing its potential savings. It looks like Anthropic intentionally compensated for this during their introductory period, but it will end in late August.
For now, Opus 4.8 hasn’t been dethroned. But Sonnet 5 has changed the calculation enough that defaulting to the flagship model without a specific reason to need it looks a lot less obvious than it did a week ago.
Also read: Most global datacentres face climate risk: What happens to India’s AI boom?