Anthropic released Claude Opus 5 last Friday, July 24. It costs exactly what Opus 4.8 cost, and on the hard tests it roughly doubles Opus 4.8's score. If you're on a Claude Max plan you're already using it, because it became your default the day it shipped. Here's what changed, what the effort setting does to your bill, and when Opus 5 is the right model to reach for (and when Sonnet or Fable still is).

Where Claude Opus 5 sits in the lineup

"Claude" isn't one thing. It's a family of models at different levels of power, speed, and cost, and you pick the one that fits the job. We covered this in more depth in the Fable 5 write-up, so this is the short version.

Claude's tiers, lightest to most powerful

The rule has always been that more powerful is not the same as more appropriate. Bigger models cost more and run slower, so the work is in matching the model to the task. Opus 5 creates a grey area around that rule, because it pulls a lot of Fable's capability down into Opus-level pricing.

What actually changed

Start with the bill, since it didn't move. Opus 5 costs $5 per million words in and $25 per million out, the same as Opus 4.8, which is half what Fable charges on input.

Now the performance. On a hard software engineering test, Opus 5 more than doubled the old Opus score while spending less per task. On a test built around problems the model has never seen before, it scored about three times the next best model on the market. And on Zapier's benchmark for finishing real business tasks start to finish, it took the top spot, with even its lowest effort setting completing more tasks than any other model.

Three results from the launch worth knowing

It built its own eyes. Anthropic gave Opus 5 a mechanical drawing and asked it to rebuild the part as a 3D model in code, then deliberately took away its ability to look at the drawing. Opus 5 wrote its own image-processing code to pull the geometry out of the raw pixels, then built the part. It did this repeatedly. No competing model solved it in five tries.

The drawing test. No competing model solved it in five tries.

It found the real bug. Given a real defect in a widely used developer tool, Opus 5 traced the actual cause and fixed an edge case the human maintainers had missed. A rival model patched the visible symptom and declared victory.

It ran a churn save unattended. Zapier's CEO said Opus 5 took a raw account-health spreadsheet and ran an entire churn-prevention sequence on its own: flagged the at-risk accounts, alerted the right owner, summarized it for the retention team. Earlier models failed the task outright. Opus 5 went 100%.

Reasoning is what nearly every tester report came back to. Opus 5 "thinks" before it writes, and it catches its own bad reasoning during planning instead of after it has built the wrong thing. One engineer described proposing a design, having Opus 5 push back, insisting anyway, and watching it hold its position, explain which part of the idea was worth keeping, and offer a compromise.

The effort setting is the part that controls your bill

Like 4.8, Opus 5 has an effort setting you can dial from low all the way up to max. Low is cheap and quick. Max spends real money (read: usage) "thinking." The dial itself isn't the news, since 4.8 had it too. What's new is that the low end finally got good enough to trust.

On 4.8, turning the effort down meant accepting worse work, so most people left it high and quietly ate the cost. Opus 5 breaks that tradeoff:

• At its lowest setting, it still finishes more of Zapier's business tasks than any other model at any setting.
• A legal team held output quality flat while using 26% fewer tokens than 4.8 at max reasoning.
• A trading firm got better answers on its own benchmark using roughly a seventh of the reasoning tokens and under half the latency of 4.8.

Now think about what your team actually runs through AI in a week, or should be running through it. Reformatting a list. Drafting a routine email. Pulling three numbers out of a PDF. None of that needs a model's deepest reasoning, and plenty of teams are paying for it anyway, because nobody ever went back and touched the dial.

How to split the dial: cheap by default, expensive on purpose

Anthropic's own charts show Opus 5 beating every other model at any given cost across effort levels, and on some tasks its floor still beats rivals at their ceiling. There's also a Fast mode that runs about two and a half times quicker for double the price, for the times when the clock costs more than the tokens.

Your default probably changed

Opus 5 is now the default on Claude Max and the strongest option on Pro, so if you're on either plan you may already be using it without knowing. It's live everywhere: the Claude app, Claude Code, Cowork, the API, plus Amazon Bedrock, Google Cloud, and Microsoft Foundry.

For most teams, that changes the answer to "which model do we use?":

• Sonnet 5 still makes sense for simple, high-volume work where cost is the whole point.
• Opus 5 becomes the reasonable default for real work, especially anything multi-step or analytical.
• Fable 5 stays parked for the rare monster job, the huge project or the enormous migration, where its extra ceiling is worth the premium.

Where Opus 5 earns its money

Across very different industries, testers reported the same pattern: the harder and longer the task, the further ahead Opus 5 pulls.

A trading firm's engineer built a market data feed for a new exchange in one sitting, something previous models couldn't finish even with a detailed plan handed to them, and when Opus 5 found no live feed to test against, it built its own test harness.

On longer work, it holds the thread. One tester handed it a chief-of-staff role over their development systems for a weekend, and it built its own monitoring, ran the machines, and surfaced only the decisions that needed a human. And for anyone making documents for a living, testers called it the clearest jump they'd seen from building to revising things like full decks, with better visual judgment and fewer formatting messes.

Where Opus 5 isn't the answer

• It's overkill for simple work at volume. If you route everything through Opus 5 because it's the default, you're paying a premium for reasoning you're not using. Sonnet exists for a reason.
• It's not the top of the ladder. On the biggest and longest jobs, Fable still has more headroom. Opus 5 gets close on a lot of tasks, not all of them.
• It has walls around cybersecurity and biology work. It can find vulnerabilities in source code but blocks the offensive side, and flagged requests get quietly handed to the older Opus 4.8 instead. Anthropic says these limits trip about 85% less often than Fable's do, but if your work lives near those areas you'll feel it occasionally.

The short version

Opus 5 is Anthropic's new daily driver: near top-tier capability, unchanged price, noticeably better judgment, and no data-retention requirement of the kind Fable carries. The one thing to actually do differently is use the effort setting. It's been there since 4.8 and most people still ignore it, but this is the first version where dialing it down doesn't cost you work quality.

Keep reading