Anthropic shipped Claude Opus 5 on July 24, 2026. If you blinked, you missed it, because this is the fourth Claude 5 model the company has pushed out in under two months.
Here is the short version. Opus 5 gets close to the intelligence of Claude Fable 5, Anthropic’s most capable publicly available model, and it does it at half the token price. It is now the default model on Claude Max and the strongest model you can reach on Claude Pro.
That is the headline everyone ran with. The more interesting story is sitting in the settings menu and the pricing page, and that is what we are going to dig into.
Claude Opus 5 at a glance
- Released: July 24, 2026
- API model string: claude-opus-5
- Price: $5 per million input tokens, $25 per million output tokens
- Same price as: Opus 4.8, the model it replaces
- Half the price of: Claude Fable 5
- Default on: Claude Max, and the strongest option on Claude Pro
- Fast mode: around 2.5x the default speed, at twice the base price
- Data retention: none required for general access
What exactly is Claude Opus 5?
Opus 5 is Anthropic’s new everyday workhorse. It is not the smartest thing the company has built, and Anthropic is refreshingly upfront about that.
The Claude 5 lineup is a bit confusing right now, so here is how it stacks up:
- Mythos 5 sits at the top. It is not publicly available. A small number of trusted organisations get access through Anthropic’s Project Glasswing, mostly because of what it can do in cybersecurity.
- Fable 5 is the same underlying model as Mythos with extra safety measures bolted on for biology, cyber and AI research. It launched in early June and is the most capable model you can actually buy.
- Opus 5 is the new arrival. Near-frontier, half the price, built to be run all day.
- Sonnet 5 landed at the end of June for high-volume work.
- Haiku is still on 4.5 and is the only tier yet to get a 5-series upgrade.
So Opus 5 is not the ceiling. It is the floor being raised, which for most teams is the more useful thing.
How much does Claude Opus 5 cost?
Five dollars per million input tokens. Twenty five per million output. Exactly what Opus 4.8 cost, so if you are already running Opus in production, your bill does not move and your model gets better.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Claude Opus 5 | $5 | $25 |
| Claude Opus 4.8 | $5 | $25 |
| Claude Fable 5 | $10 | $50 |
That is the whole pitch in one table. You give up a slice of capability against Fable 5 and you halve your inference cost.
For a lot of workloads that trade is obvious. For some it is not, and we will get to which is which.
What is actually new in Claude Opus 5?
The effort dial
This is the feature that deserves more attention than it got.
You can now tell the model how hard to think. Low, medium, high, and up through extra high and max on the API. Crank it up when the problem is genuinely hard. Dial it down when you just need a competent answer and you would rather not pay for a philosophy degree.
Anthropic says Opus 5 holds a surprising amount of its performance at lower effort levels while burning noticeably fewer tokens. On Zapier’s AutomationBench, even the lowest effort setting reportedly passes more tasks than any other model at any setting.
If you run agents at scale, this is the line item that matters. Model choice used to be your only cost lever. Now you have two.
Fast mode
Same model, roughly 2.5 times the speed, twice the base price. It is available on the Claude Platform and through usage credits in Claude Code.
Worth it for anything a human is sitting and waiting on. Probably not worth it for a batch job running overnight.
Mid-conversation tool changes
In beta, and genuinely useful if you build agents. You can now swap which tools Claude has access to partway through a conversation without blowing up your prompt cache.
Anyone who has built a multi-stage agent knows the pain here. You either loaded every tool up front and watched the model get distracted, or you started a fresh conversation and paid to re-read your entire context. Now you can hand it a different toolkit at each stage and keep the cache warm.
Automatic fallbacks
Also in beta. When a safety classifier flags a request, instead of the call just failing, it can route to a different model automatically.
In Claude.ai, Claude Code and Claude Cowork this happens by default, with flagged requests falling back to Opus 4.8. On the API you opt in.
This is a quiet but real reliability improvement. A hard block in the middle of an automated pipeline is an outage. A fallback is a slightly different answer.
What do the Claude Opus 5 benchmarks say?
Anthropic’s own numbers, so read them with the usual pinch of salt. That said, some of these gaps are wide enough to be interesting.
- Frontier-Bench v0.1: Opus 5 comes out ahead of every other model, and more than doubles what Opus 4.8 managed, at a lower cost per task.
- CursorBench 3.2: at max effort it lands within half a percent of Fable 5’s best score, for half the cost per task.
- ARC-AGI 3: this one tests novel problem solving, and Opus 5 scores three times the next-best model.
- Zapier AutomationBench: around 1.5x the pass rate of the next-best model at the same cost per task.
- OSWorld 2.0: a computer-use benchmark, where Opus 5 beats Fable 5’s best result at just over a third of the cost.
Anthropic also calls it the new state of the art on GDPval-AA, a benchmark built to look more like real client deliverables than isolated coding puzzles.
One caveat worth stating plainly: independent head-to-heads against GPT and Gemini have not landed yet. The model is days old. Every number above is either from Anthropic or from an early-access partner, and partner quotes in a launch post are marketing, however honest they might be.
The examples that are more convincing than the charts
Buried in the announcement are three stories that say more than any leaderboard.
On one Frontier-Bench task, the model was handed a drawing of a machine part and asked to rebuild it as a 3D FreeCAD model in code. The catch: it had no way to actually look at the drawing. So it wrote its own computer vision pipeline to pull the geometry out of the raw pixels, then reconstructed the part. Anthropic says no competing model solved it in five attempts.
Given a real bug in a widely used open-source package manager, it found the root cause and patched an edge case the community’s own fix had missed. A rival model fixed the visible symptom and declared victory.
And an engineer at a trading firm used it to build a market data feed for a new exchange in one session. There was no live feed to test against, so the model built its own test harness to verify it was parsing the exchange data correctly.
The common thread is not raw intelligence. It is the model noticing that it cannot verify its work, and then building the thing that lets it verify its work. That behaviour is what makes long-running agents survivable.
Claude Opus 5 vs Claude Fable 5: which should you use?
Rough rule of thumb.
Go with Opus 5 if you are doing everyday engineering, agentic workflows, knowledge work, data analysis, document review, or anything you run hundreds of times a day. The cost difference compounds fast and the capability gap is small.
Stay on Fable 5 if you need days-long autonomous runs, or you are working at the hard edge of a specialist domain where the last few percentage points genuinely change the outcome.
Note for security teams: Fable 5 still leads on offensive cyber work, and Mythos 5 leads outright. Opus 5 was deliberately not trained on cyber tasks.
One more consideration that gets overlooked. Opus 5 carries no data retention requirement for general access. Fable 5 does. If you are in a regulated environment with a hard zero-retention policy, that alone may decide it for you.
Is Opus 5 any good for scientific research?
This is the underreported part of the launch.
Because Opus 5 uses roughly the same biology safeguards as Opus 4.8 rather than the tighter Fable 5 set, it is now Anthropic’s most capable generally available model for scientific research. Biology requests that get blocked on Fable 5 will now route to Opus 5.
The gains are real. Anthropic says it beats Opus 4.8 on every one of its life sciences evaluations. The biggest jumps are in organic chemistry, where inferring molecular structures from spectroscopy data improved by 10.2 percentage points, and protein work, where predicting how sequence variations affect function improved by 7.7 points.
Anthropic is also careful to say the model still struggles with long-running autonomous research, which is exactly where it thinks the biology risk actually lives.
How aligned is Claude Opus 5?
Anthropic’s automated behavioural audit rates Opus 5 as its most aligned model so far. It scores 2.3 on overall misaligned behaviour, the lowest of the recent lineup, and it sticks to Claude’s Constitution more closely than Opus 4.8, Sonnet 5 or Fable 5.
It also shows the lowest rates of deceptive behaviour Anthropic has measured, and it is the hardest of the bunch to trick into misuse. The company additionally calls it the safest yet at avoiding reckless actions with hard-to-reverse consequences.
If you are handing an agent write access to anything that matters, that last point is not a footnote. It is the whole risk model.
What about the cybersecurity guardrails?
This is where the launch gets genuinely contentious, and it is worth understanding before you commit.
Opus 5’s cyber classifiers are looser than Fable 5’s. The model can look for vulnerabilities in source code, which skews defensive. It is blocked from binary-based vulnerability scanning, penetration testing and exploit generation, which skew the other way.
Anthropic expects those classifiers to fire roughly 85 percent less often than they do on Fable 5. When they do fire in Claude.ai, Claude Code or Cowork, the request drops to Opus 4.8 rather than failing outright.
The underlying reasoning shows up in an Anthropic evaluation called OSS-Fuzz. Opus 5 is nearly as good as Mythos 5 at finding vulnerabilities, but well behind it at exploiting them. Finding is defence. Exploiting is offence. The safeguards are drawn along that line.
Security researchers who need fewer restrictions can go through Anthropic’s Cyber Verification Program. Existing members already have access to a less restricted build.
Fair criticism exists here. Some enterprises are not thrilled about safety classifiers deciding which model answers a given request. Automatic fallbacks soften that, but they do not remove it.
Where can you use Claude Opus 5?
- Claude apps: default on Max, strongest model on Pro
- Claude API: as claude-opus-5
- Claude Code and Claude Cowork
- Amazon Bedrock, Google Cloud and Microsoft Foundry
Anthropic has also published a dedicated prompting guide for the model, which is worth ten minutes if you are migrating existing prompts.
What this actually means if you build with AI
Three things stand out.
The competition moved from capability to cost. For three years the labs raced on what their best model could do on its best day. Opus 5 is a bet that the money is in the middle band of difficulty, where good-enough intelligence delivered efficiently beats brilliant intelligence delivered expensively. Given that inference spend is now a board-level line item at most serious AI shops, that bet looks well placed.
Tuning is becoming a real skill. Between the effort dial, Fast mode and automatic fallbacks, you now have several ways to trade cost against quality inside a single model. Teams that learn to route work properly will run circles on price against teams that just pick the biggest model and hope.
Release cycles have collapsed. Four models in under two months. Whatever you standardised on in May is already two generations old. If your architecture makes swapping models painful, that is now a competitive problem rather than a technical debt item.
For anyone building AI into products, the practical takeaway is boring but true. Stop optimising for the leaderboard. Start measuring your cost per completed task, because that is the number Opus 5 was designed to attack.
Frequently asked questions
When was Claude Opus 5 released?
Anthropic released Claude Opus 5 on July 24, 2026, and made it available across all its platforms the same day.
How much does Claude Opus 5 cost?
It costs $5 per million input tokens and $25 per million output tokens. That is the same price as Opus 4.8 and half the price of Claude Fable 5. Fast mode costs twice the base rate.
Is Claude Opus 5 better than Claude Fable 5?
On several coding and knowledge work benchmarks, yes, and at half the cost. Fable 5 still leads on the hardest long-horizon work and on offensive cybersecurity tasks. Anthropic positions Opus 5 as the everyday model rather than the most capable one.
What is the effort setting in Claude Opus 5?
It is a control that lets you decide how much computation the model spends on a task. Higher effort means better results on hard problems. Lower effort means fewer tokens and a smaller bill, with less capability loss than you might expect.
Which API model string do I use?
Use claude-opus-5 on the Claude API. The model is also available through Amazon Bedrock, Google Cloud and Microsoft Foundry.
Does Claude Opus 5 retain my data?
Anthropic states that Opus 5 has no data retention requirements for general access, consistent with previous Opus models. Fable 5 does have a retention requirement, which is a meaningful difference for regulated industries.
Can Claude Opus 5 do cybersecurity work?
Partly. It can search for vulnerabilities in source code. It is blocked from binary-based vulnerability scanning, penetration testing and exploit generation. Researchers needing broader access can apply to Anthropic’s Cyber Verification Program.
The bottom line
Claude Opus 5 is not the flashiest launch of the year. It is probably one of the more consequential ones.
Anthropic has taken most of what made Fable 5 impressive, cut the price in half, added controls that let you cut it further, and shipped it as the default for paying customers. Claude reportedly held around 40 percent of the enterprise large language model market by usage as of late 2025, according to a February 2026 analysis by Contrary Research, so this is not a fringe release. It resets the price-performance baseline that everyone else now has to answer.
The honest unknown is whether the efficiency claims hold up under real production load, and how often those safety classifiers get in the way for teams doing legitimate security work. Give it a month and we will know.
Sources: Anthropic’s Claude Opus 5 announcement, the Opus 5 system card, and reporting from TechCrunch, VentureBeat and Axios.
Building with AI and trying to work out who to hire? Our regional guides break down the market: software development companies in New York, the UAE, and Germany.