Claude Opus 5: same price, thinking-on by default, and mostly worth it
Claude Opus 5 keeps Opus 4.8 pricing and genuinely improves multi-file code refactors, but thinking-on by default quietly inflates your token bill 2-3x.
Claude Opus 5 is real and it's good at code, but the "near-Fable-5 at half the price" framing is doing a lot of work. Same $5/$25 per Mtok as Opus 4.8, 1M context, 128k output. I pushed it through a week of actual coding tasks and it held up — the benchmark gains translate on refactors, less so on the stuff people actually get stuck on.
Anthropic shipped this on July 24 after the "Honeycomb" leaks, and you can read the full spec sheet in Anthropic's Opus 5 launch post. This is their fourth Claude 5 model in under two months, which is either impressive velocity or a sign the version numbers stopped meaning anything.
Do the benchmarks translate to real coding?
Mostly yes, on structured work. I gave it a gnarly multi-file TypeScript refactor — moving a service layer behind an interface, updating ~40 call sites. Opus 5 did it in one pass with no broken imports, which Opus 4.8 fumbled about a third of the time. The 128k output ceiling matters here: it wrote the whole diff without me babysitting continuation prompts.
Where it's less convincing is debugging live systems. Given a flaky test with a race condition, it confidently proposed three fixes, two of which were plausible-looking nonsense. The 30.2% on ARC-AGI-3 tracks with my gut: it's better at generating than at reasoning about why something's already broken.
What's the gotcha?
Thinking-on by default is the thing that'll surprise you on the bill. Every request now burns reasoning tokens whether the task needs them or not, and those count as output at $25/Mtok. For a batch of simple codegen calls I saw token usage jump roughly 2-3x versus flipping thinking off. If you're running this in a loop or an agent, set the thinking budget explicitly or you'll eat the cost.
Also worth naming: it's behind Mythos 5 on cyber tasks. If security tooling or exploit reasoning is your workload, Opus 5 isn't the pick.
Is it benchmaxxing?
The HN crowd is arguing benchmaxxing vs real gains, and honestly it's both. Frontier-Bench and GDPval-AA numbers are up, and some of that is genuine — the refactor performance is not a benchmark artifact. But four models in two months means each release is a smaller delta dressed up in fresh SOTA claims. Opus 5 over Opus 4.8 is an upgrade, not a leap.
Should you use it?
Yes, if you're already on Opus and doing heavy code generation or large-context work — the price didn't move and the code quality did. Switch your default. But turn off thinking for cheap tasks, keep Mythos 5 in mind for security work, and don't believe the near-Fable-5 marketing until you've run your own eval. Verdict: a solid point release wearing a major-version badge.
FAQ
Is Claude Opus 5 more expensive than Opus 4.8?
No. Pricing is unchanged at $5/Mtok input and $25/Mtok output. The real cost difference comes from thinking-on by default burning extra output tokens, so effective spend can rise 2-3x on simple tasks unless you cap the thinking budget.
Is Opus 5 good enough to replace Opus 4.8 for coding?
For multi-file refactors and large-output codegen, yes — it completes tasks in fewer passes. For live debugging and race-condition reasoning it's only marginally better, so don't expect it to fix workflows that were already frustrating.
Should I use Opus 5 for security work?
Not as a first choice. Opus 5 trails Mythos 5 on cyber benchmarks, so for exploit reasoning or security tooling you're better off with a model that scores higher on those tasks.