Claude Opus 5 Arrives at 4.8 Pricing—But Watch for Breaking Changes
Claude Opus 5 is here—and at the same price as Opus 4.8, it cuts the cost gap with Fable 5 in half while matching most of its performance gains. That’s the headline, but the real story lies in three subtle changes that could quietly break code that worked last week.
A Pricing Upside with Hidden Traps
Opus 5’s input/output pricing matches Opus 4.8 at $5 and $25 per million tokens, exactly half of Fable 5’s $10/$50 rates. On benchmarks, it trails Fable 5 by just 1.1 points on SWE-bench Pro while outperforming it on OSWorld 2.0 at a third of the cost. The catch? Three behavior shifts now force developers to audit their integrations before migrating.
The first surprise is thinking being enabled by default. In Opus 4.8, omitting the parameter meant no internal reasoning—now it triggers adaptive thinking, which risks truncating responses if max_tokens is set too tightly. Disabling thinking outright is also restricted: setting thinking=disabled with effort=xhigh or max now returns a 400 error, a change that breaks routes that previously allowed this combination.
Prompt Caching Gets Cheaper—and Shorter
Opus 5 lowers the minimum cacheable prompt size from 1,024 to 512 tokens, making caching viable for shorter reusable prompts without code changes. For teams running many small, repeated calls, this could quietly reduce costs—but only if the new cache threshold is accounted for in existing workflows.
The model’s rate limits also operate independently from Opus 4.8’s shared bucket, meaning migrating traffic won’t free up headroom on legacy limits. Developers need to check their tier’s specific constraints before assuming capacity will scale smoothly.
Why it matters
For teams balancing cost and performance, Opus 5 flips the script: Fable 5’s premium is harder to justify for most hard tasks, but the migration isn’t plug-and-play. The silent defaults and tightened controls around thinking and effort levels demand immediate audits to avoid runtime failures. Those leveraging prompt caching will see immediate benefits, while others risk hidden costs from truncated responses or rate-limit misconfigurations. The real win isn’t just pricing—it’s whether your code can handle the change.
Source: DEV Community. AI-assisted editorial synthesis — TechnoExpress.