Claude Opus 5: What Actually Changes for Your Vibe Coding Project

A model swap that looks like one line of code is rarely just one line of code. Moving from Claude Opus 4.8 to Claude Opus 5 genuinely is a changed model ID – and you still get back a system that behaves differently: longer answers, higher token spend and, in one place, an entirely new failure mode. This piece places the official numbers as of 31 July 2026 and turns them into an upgrade you can verify.

Anthropic released Claude Opus 5 on 24 July 2026 and frames it as a step change over Opus 4.8 rather than an incremental update. The largest gains it names are in deep reasoning, long agentic task chains and scaling results with additional compute at response time.

What is new in Claude Opus 5?

The model ID is claude-opus-5. Claude Platform lists a context window of one million tokens – as both the default and the maximum, with no smaller variant available. Output runs up to 128,000 tokens. The model is available through the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, as well as inside the Claude applications themselves.

Pricing stays at 5 US dollars per million input tokens and 25 US dollars per million output tokens, unchanged from Opus 4.8. One easily missed improvement: the minimum length for a cacheable prompt drops from 1,024 to 512 tokens. Shorter recurring system prompts now benefit from caching without any code change at all.

Why thinking now runs by default

This is the change with the most side effects. On Opus 4.8 a request ran without thinking unless you explicitly enabled it. On Opus 5 it is the other way around: the model decides when and how much to think as soon as you say nothing.

In practice that means three things. Your responses take longer. They cost more output tokens. And because max_tokens caps thinking and answer text together, a tight value that was fine on Opus 4.8 can now cut off mid-sentence. Anyone upgrading should re-measure that value against a real task rather than carrying it over. The systematic way to do that is in keeping AI costs under control.

The effort ladder and its new trap

Effort controls how many tokens Claude spends on thinking, tool calls and the response. The ladder runs low, medium, high, xhigh, max, and the default is high. Anthropic describes Opus 5 as the first Opus model that converts extra effort into better results more reliably – so the level you pick carries more weight than it did on Opus 4.8.

The opposite direction is just as notable: the documentation says low and medium still produce strong quality at a fraction of the tokens and latency. So the interesting question is no longer only how high you can go, but how far down you can come before quality breaks.

One genuine trap: thinking can only be disabled up to level high. Combine thinking: disabled with xhigh or max and the API answers with a 400 error. This is validated per request, so a later call in the same session can fail even though earlier ones went through.

When is Opus 5 worth it for vibe coding?

The centre of gravity is still where several decisions depend on each other. Anthropic names multi-file features, larger refactors and work that runs end to end without leaving stubs behind. Code review and bug finding are described as strengths too, with a high hit rate and few false positives.

For a known single-file change, a short summary or a standardised classification, Opus remains oversized. The best default is not the strongest model but the smallest configuration that reliably meets your quality contract. On Opus 5, that smallest configuration can explicitly be a low effort level of the same model.

A practical case: upgrading from 4.8 to 5

Step one is the model ID. Then come the two behaviour changes. Check every call that disables thinking: if it runs at xhigh or max, you either lower the effort level or drop the thinking field. Then check every call that never set thinking at all – there the thinking share is new and max_tokens may now be too tight.

Step two is the prompt. If it contains instructions like "add a final verification step" or "have a subagent check the result", take them out. Opus 5 verifies its own work anyway; carried-over verification instructions produce over-verification rather than extra safety. The same applies to passages encouraging delegation to subagents: Opus 5 already delegates more readily on its own.

Step three is measurement. Run the same real task on 4.8 and on 5 and log token spend, latency, number of correction rounds and your own review time. Only that comparison shows whether the switch pays off in your case.

What else changes in behaviour

The documentation names four points, and all four are noticeable day to day: user-facing responses get longer. Written files and reports get longer. In agentic sessions the model narrates its progress more often. And in multi-agent setups it delegates more readily.

These are defaults, not bugs – and defaults can be moved with prompting. If your product needs short answers, say so explicitly. If a coding agent narrates too much, ask for silence between tool calls. The pattern mirrors what AI agents are: permissions, scope and communication style belong in the brief.

What limits remain despite better scores?

Anthropic reports clear leads for Opus 5 in its own evaluations, including more than doubling Opus 4.8's score on one internal benchmark. Those are vendor results under defined conditions, not a promise about your application. They are a good reason to run your own comparison, never a substitute for one.

One more detail deserves attention: with thinking disabled, the documentation notes that Opus 5 can occasionally write a tool call into its visible text instead of actually triggering it. The run then looks successful although nothing happened. That is another argument for leaving thinking on and steering through effort instead.

Checklist for your upgrade to Opus 5

• Every call with thinking disabled is verified to run at effort high or lower.
• max_tokens has been re-measured for calls that previously ran without thinking.
• Verification and delegation instructions have been removed from older prompts.
• Effort levels are compared from an identical starting state instead of raised blindly.
• Token spend, latency, correction rounds and review time are logged for both 4.8 and 5.
• Tests, diff and user path are checked independently of the model's summary.
• Critical changes are approved by an accountable human before release.

Claude Opus 5 moves the point at which a strong model pays for itself – upward in capability and downward in cost per usable result at the same time. The gain does not come from the model name, though. It comes from treating the upgrade as a measurable change: adjusted limits, tidied prompts, and an acceptance step that stays yours.