Anthropic released Claude Opus 5 on July 24, 2026, and most of the coverage fixated on the easy numbers — a 1-million-token context window, pricing on par with Opus 4.8, half the cost of Fable 5. Those are real and worth noting. But the change that actually rewires how the model behaves day to day isn't any of them. It's that extended thinking now runs by default, on every request, whether you asked for it or not.
What Is "Extended Thinking" in Claude Opus 5?
Extended thinking — Anthropic calls it adaptive thinking in Opus 5's case — is the model's internal chain of reasoning before it produces a final response. Think of it as the model working through a problem privately, checking its own logic, before committing to an answer. In earlier releases, like Opus 4.8, you had to explicitly enable this. Leave the thinking field blank and the model replied without it. In Opus 5, that logic is flipped: reasoning is on unless you turn it down.
This matters for a reason that isn't immediately obvious. The model doesn't just answer questions now — it examines them first. That's a different kind of tool. Where a previous Claude might have fired back a quick response that was technically correct but lacked self-correction, Opus 5 pauses, processes, and verifies. Anthropic describes it as the model being "much stronger at verifying its work and iterating carefully until it succeeds" — and that behavior is baked in, not a setting you toggle on for hard problems.
How It Differs from Opus 4.8
The architectural difference is this: on Opus 4.8, a request that omitted the thinking parameter ran without thinking. On Opus 5, that same request triggers adaptive thinking automatically, and every thinking token bills at the output rate — $25 per million. The sticker price is identical. The real-world bill isn't. Developers migrating from 4.8 without adjusting their setup have reported Opus 5 spending roughly 2× the output tokens on comparable tasks. The price-per-token didn't move; the token count did.
The Effort Dial — and Why More Isn't Always Better
To give users control over this, Anthropic introduced an explicit effort parameter with five levels: low, medium, high, xhigh, and max. These let developers (and power users on the API) tune how much computational depth the model applies before responding.
The Counterintuitive Finding
- Medium effort is the sweet spot for most tasks. Anthropic's own data shows Opus 5's accuracy peaking at medium reasoning effort, not max.
- Pushing to max can actively hurt performance. On straightforward tasks, excessive reasoning causes the model to second-guess correct initial answers — a real phenomenon sometimes called "overthinking."
- Max effort is for genuinely hard problems. Mathematical proofs, open-ended research questions, complex multi-step code generation — that's where xhigh or max earns its cost.
- Low effort is a cost lever, not a quality sacrifice for simple tasks. Routine lookups, formatting work, basic summarization — low effort handles these well and cuts your token usage sharply.
- The effort toggle, not the per-token price, is where your bill is actually decided. That's the real control surface in this model.
On Opus 5, the same API request that previously ran without thinking now triggers adaptive reasoning by default — and every thinking token bills at $25 per million output tokens.
What Actually Changes About "Using Claude"
For everyday users on Claude.com, the shift is felt rather than configured. Responses take a visible beat longer on complex prompts — that's the thinking happening. But the outputs land with noticeably more internal consistency. The model catches its own errors before they reach you. You no longer need to include explicit verification instructions in your prompts; Opus 5 does that work unprompted.
It also means prompting for brevity has become the user's job now. Opus 5 generates longer responses by default — not because it's padding, but because it's showing more of its reasoning path. If you want concise output, you have to ask for it. That's a real change in behavior from previous Claude models, where brevity was closer to the default register.
Beyond that, the model's position in Anthropic's lineup has shifted. Opus 5 is now the default model on Claude Max and the strongest option available on Claude Pro. It's described by Anthropic as their "most capable generally available model for scientific research," with particular depth in biology tasks. On OSWorld 2.0 (a computer use benchmark), it outperforms every other model at a given cost and beats Fable 5's best results at roughly a third of the price. For most professional use cases, Fable 5 is no longer the obvious choice — Opus 5 is.
There's also a new Automatic Fallbacks feature in beta: when a prompt triggers a safety classifier, the API now routes the request to a less powerful model rather than returning an error. The practical effect is fewer dead ends in production pipelines. Opus 5 also engages safety classifiers 85% less often than Fable 5, which means less friction on legitimate edge-case prompts.
Fast mode is worth mentioning separately. It runs Opus 5 at roughly 2.5× default speed for double the base price — $10 per million input tokens and $50 per million output tokens. For latency-sensitive agentic pipelines, that tradeoff is often worth it. For most other work, default speed is fine.
Key Takeaways
- Extended thinking is on by default in Opus 5 — every request reasons before responding, and thinking tokens bill at the output rate of $25/million.
- The effort toggle (low → max) is the real pricing lever; accuracy peaks at medium effort, not max, for most task types.
- Opus 5 verifies its own work unprompted, generates longer responses by default, and is now the strongest model available on Claude Pro and the default on Claude Max.
The 1-million-token context window is a headline feature. The halved price versus Fable 5 is a legitimate selling point. But thinking by default is the design choice you'll actually feel — in response quality, in latency, and in how you need to prompt the model going forward. If you've been using Claude without touching the API, it's worth experimenting with explicit effort settings to understand where your use cases actually fall on that curve.
Frequently Asked Questions
Yes — adaptive thinking is on by default for all requests in Opus 5. On the API, developers can control the depth with the effort parameter (low through max) or disable it explicitly, but the out-of-the-box behavior applies reasoning to every prompt, regardless of complexity.
The per-token price is identical — $5/million input, $25/million output. In practice, many users find their API bills higher because Opus 5 generates roughly 2× the output tokens on comparable tasks, largely due to adaptive thinking running by default. Tuning the effort level down is the most effective way to manage this.