If you use AI every day to code, write, or automate, you know the trade-off: the smartest model is almost always the one that drains your budget fastest. That's exactly the knot Claude Opus 5 unties.
Announced by Anthropic on July 24, 2026 and available immediately, Opus 5 comes close to the frontier intelligence of Fable 5 at half the price, while becoming the new state of the art on coding and knowledge work. Designed for daily use, it combines speed, rigor, and an excellent performance-to-cost ratio.
In this article, we'll break down everything you need to know about Claude Opus 5: its real-world performance, pricing, benchmarks (the standardized tests used to compare models against each other), concrete use cases, and limitations. Whether you're a developer, an analyst, a content creator, or simply curious about AI, you'll find a complete and actionable analysis here.
What is Claude Opus 5?
Claude Opus 5 is the most accomplished language model in Anthropic's Opus line. It succeeds Opus 4.8. Its defining trait: delivering intelligence close to the best models on the market while staying token-efficient, and therefore faster and cheaper.
A quick refresher if you're new to this: a token is the unit of text the model processes and that you pay for, roughly three quarters of a word in English. The fewer tokens a model burns to reach the right result, the smaller the bill.
Where does Opus sit in the lineup? Anthropic stacks five tiers: Haiku (fastest and cheapest), Sonnet (the balanced option), Opus (the generalist high end), Fable (the most capable publicly available model), and Mythos (reserved for a handful of trusted organizations). So Opus 5 is not Anthropic's most powerful model in absolute terms, but it does offer the best intelligence-to-price ratio for everyday use.
In practice, Opus 5 is now the default model on Claude Max and the most powerful model available on Claude Pro. It's built to be used every day, across tasks ranging from software development to scientific research to business process automation.
A "thoughtful and proactive" approach
Anthropic describes Opus 5 as a "thoughtful and proactive" model. That means it doesn't just answer: it iterates, verifies, and corrects until it succeeds. During early-access testing, several companies reported that the model spontaneously built its own validation tools when it couldn't find any.
One striking example: on a 3D reconstruction task for a machine part, with no direct access to the technical drawing, Opus 5 wrote its own computer vision pipeline (a program that "reads" an image to extract shapes) to pull the geometry from the raw pixels, then rebuilt the complete part. No competing model solved this task after five attempts.
What sets Opus 5 apart is its judgment: it thinks harder before writing a single line of code, catches its own logical faults during the planning phase, and checks its work the way an experienced professional would.
Performance and cost-effectiveness
Opus 5's real strength lies in its balance between intelligence and cost. At the same price as its predecessor Opus 4.8, it delivers markedly better performance. Anthropic offers an effort setting that lets you optimize either for maximum intelligence or for conserving tokens to get faster, cheaper results. Think of that dial as choosing between "take your time and be exhaustive" and "get straight to the point."
Excellence in software engineering
On development tasks, Opus 5 is impressive:
- Frontier-Bench v0.1: Opus 5 surpasses every other model and more than doubles Opus 4.8's performance, at a lower cost per task.
- CursorBench 3.2: at max effort, it lands within 0.5% of Fable 5's peak score, but at half the cost per task.
- FrontierCode 1.1: Opus 5 approaches Fable 5's level at half the cost, with particular strength in hard debugging and root-cause analysis (tracing a bug back to its actual origin rather than papering over the symptom).
Reasoning and problem solving
The results are just as remarkable on knowledge work tasks, measured notably by GDPval-AA:
- ARC-AGI 3: on this evaluation of novel problem solving, meaning puzzles the model has never seen during training, Opus 5's score is three times higher than the next-best model's.
- Zapier AutomationBench: its pass rate is roughly 1.5 times that of the next model, at the same cost. Even at its lowest effort setting, it passes more tasks than any other model.
- OSWorld 2.0: on this computer use benchmark (where the model drives a mouse, keyboard, and software itself), it outperforms every model at any given cost, surpassing Fable 5's best result at just over a third of the cost.
Zapier confirmed that Opus 5 topped its AutomationBench leaderboard without spending more tokens than previous Claude models. On a complete churn-prevention sequence, it hit 100% where previous models failed.
Scientific research
Opus 5 is a significant step forward for scientific research compared to Opus 4.8. It scores better on every single one of Anthropic's life sciences evaluations, covering structural biology, organic chemistry, and bioinformatics.
The gains are most notable in organic chemistry. For instance, it gains 10.2 percentage points on inferring molecular structures from spectroscopy data, a technique that identifies a molecule by how it absorbs light. On protein-related tasks, it gains 7.7 percentage points on predicting how variations in a sequence affect a protein's function.
Watch the nuance here: a gain of "10.2 percentage points" is not a 10.2% gain. Going from 60% to 70.2% success is indeed 10.2 points, but a 17% relative improvement.
Claude Opus 5 pricing
Good news for your budget: Opus 5 ships at the same price as Opus 4.8. Here's the API pricing table:
| Token type | Price (per million tokens) |
|---|---|
| Input tokens | $5 |
| Output tokens | $25 |
| Fast mode (around 2.5× faster) | 2× the base price |
The model is available through the Claude API under the identifier claude-opus-5. A Fast mode is also offered: it runs around 2.5 times faster than the default mode, at twice the base price on the Claude Platform and through usage credits in Claude Code.
Getting started with the API
Here's a simple API call to use Opus 5:
import anthropic
client = anthropic.Anthropic(api_key="your_api_key")
message = client.messages.create(
model="claude-opus-5",
max_tokens=1024,
messages=[
{"role": "user", "content": "Explain how prompt caching works."}
]
)
print(message.content)Unlike some other offerings, Opus 5 has no data retention requirements for general access, which simplifies compliance for many organizations.
Real-world use cases for Opus 5
Feedback from early-access customers paints a clear picture of where Opus 5 excels. Here are the main reported use cases.
Development and code review
Opus 5 writes clean, tight diffs (a diff is the exact list of lines added and removed in a file) with no dead code, and it spots subtle, codebase-specific issues. On the hardest agentic coding tasks, the Lovable team measures a 22% improvement over Opus 4.7, and above all far less variance from one run to the next, which is crucial for production reliability.
On a real bug in a popular open source package manager, Opus 5 found the root cause and fixed an edge case the community's patch had missed. A competing model addressed only the surface symptom.
Front-end development
Opus 5 checks its own work the way a human front-end developer would. On one benchmark, it opened its pages in a browser at desktop and mobile widths, caught a product hidden below the mobile fold (the line beyond which you have to scroll to see anything) and an off-screen checkout button, then fixed both before handing the work back. Companies also report the best animations, games, and 3D work ever produced by an Opus model.
Financial analysis and modeling
On the hardest financial modeling tasks, Opus 5 marks a clear step up in accuracy and efficiency. Across effort levels, it averages 9 percentage points more accuracy, with a third fewer turns and tool calls and 60% less time. On a trading benchmark, it uses roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8, meaning the delay between your question and its answer.
Legal work
Opus 5 crosses a threshold on legal agent work, with the biggest gains in areas like corporate governance and arbitration. On first-turn contract revisions (redlines, the annotations that propose changes directly inside the text), it scored the highest of any model tested, nearly double Opus 4.8, while maintaining quality with 26% fewer tokens on average at maximum reasoning.
What this actually changes for you, even solo
All these benchmarks can feel abstract when you're building your project alone from your bedroom or your office. Yet three very concrete shifts deserve your attention.
You can stop slicing your tasks to death. The reflex you built up with previous generations was to break every project into micro-steps so the model wouldn't lose the thread. Several customers report that Opus 5 handles in one pass work they would previously have fragmented. You get back the coordination time you used to spend gluing the pieces together.
The effort dial becomes your budget lever. When you're the only one paying for your tools, the question isn't "which model is best" but "what does one successful task cost me." Save max effort for genuinely hard problems and let low effort handle volume. Opus 5 stays strong even at the bottom of the effort range, which was rarely true before.
Automatic verification replaces the reviewer you don't have. On a team, someone reviews your code or your numbers. Solo, no one does. A model that opens its own page in a browser, tests its assumptions, and corrects itself before handing back the work fills part of that gap. It isn't a human code review, but it's a safety net that didn't exist before.
If you want to learn to actually pilot this kind of model day to day rather than just talk to it, our Claude Code course walks you through structuring your prompts, chaining tools, and keeping control over the result.
Alignment and safety
Anthropic presents Opus 5 as its most aligned model to date. Alignment is a model's ability to do what's expected of it without drifting, deceiving, or working around the rules. During pre-deployment testing, the automated behavioral audit showed that Opus 5 adheres to Claude's Constitution better than Opus 4.8, Sonnet 5, or Fable 5.
A more reliable model
Opus 5 exhibits the lowest rates of deceptive behavior and is the least susceptible to being tricked into misuse. On the automated behavioral audit, it scores 2.3 on overall misaligned behavior, the lowest of Anthropic's recent models. It's also the safest model yet when it comes to avoiding reckless actions with hard-to-reverse consequences.
Safety and dual-use capabilities
Opus 5 does not advance the frontier in risky dual-use capabilities, meaning skills that serve defense and attack equally well. In rigorous evaluations conducted with private-sector and government partners, it remains behind Mythos 5 in both biology research and offensive cybersecurity. Full details are in the model's System Card.
Anthropic deliberately avoided training Opus 5 on cyber tasks. The model nonetheless improved as a result of becoming more generally capable: it finds vulnerabilities almost as well as Mythos 5, but stays substantially behind on exploiting them, that is, on turning a flaw into a material threat.
On OSS-Fuzz, a cybersecurity evaluation, Opus 5 is close to Mythos 5 at identifying software vulnerabilities, but considerably less successful at developing exploits (programs that leverage a flaw to take control).
Safeguards and automatic fallback
Opus 5's safeguards are designed to allow beneficial uses in both cybersecurity and biology. The cyber classifiers, the automated filters that analyze each request, let Opus 5 find vulnerabilities in source code but block "binary-based" scanning, penetration testing, and exploit generation. Anthropic expects these classifiers to intervene around 85% less often than they do for Fable 5.
In Claude.ai, Claude Code, and Claude Cowork, any flagged request falls back to Opus 4.8 by default. This fallback can also be enabled on the API. Worth noting in the other direction: biology-related requests that are blocked on Fable 5 now route to Opus 5 rather than Opus 4.8.
Beta releases shipping alongside Opus 5
Alongside Opus 5, Anthropic is releasing two updates in beta:
- Mid-conversation tool changes: on the Claude Platform, developers can now change which tools Claude can use without invalidating the prompt cache.
- Automatic fallbacks on the API: requests flagged by the safety classifiers on Opus 5 (or Fable 5) can be automatically routed to another model. That way, API requests always route to the best available model rather than being blocked.
Best practices for getting the most out of Opus 5
To make the most of Opus 5, here are a few concrete recommendations. Anthropic also publishes a dedicated prompting guide for the model.
Tune the effort setting
The effort parameter is your main lever. For complex tasks demanding maximum precision, push effort to high or max. For simple or high-volume tasks, dial it down to save tokens and gain speed: Opus 5 often stays strong even at low effort.
Take advantage of prompt caching
Prompt caching means having the API memorize the fixed part of your instructions so you don't pay for it on every call. Thanks to mid-conversation tool changes without cache invalidation, structure your prompts to maximize that reuse. It sharply reduces costs on long conversations and agentic workflows.
Let it verify its own work
Opus 5's strength is its ability to iterate and verify. Give it the room to test, validate, and correct, rather than over-fragmenting your tasks. Several customers report that it handles work they would previously have broken into small steps.
Turn on automatic fallbacks
If you work on sensitive use cases (cybersecurity, biology), enable automatic fallbacks on the API to avoid getting blocked. Your requests will route to the best available model instead of being rejected.
Frequently asked questions
How much does Claude Opus 5 cost?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. Fast mode, around 2.5 times faster, is billed at twice the base price.
Is Claude Opus 5 better than Opus 4.8?
Yes, clearly, and at the same price. It more than doubles Opus 4.8's score on Frontier-Bench v0.1, gains 9 accuracy points in financial modeling, and improves on every life sciences evaluation. On the hardest agentic coding tasks, Lovable even measures a 22% improvement over the Opus 4.7 generation. It's also a more reliable model, with less variance from run to run.
How do I access Claude Opus 5?
Opus 5 is available across all Anthropic platforms. It's the default model on Claude Max and the most powerful on Claude Pro. Developers can use it through the Claude API with the identifier claude-opus-5.
Is Claude Opus 5 suitable for cybersecurity?
Opus 5 can find vulnerabilities in source code, but its safeguards block binary-based scanning, penetration testing, and exploit generation. Enterprises and researchers in the Cyber Verification Program (CVP) have access to a version with fewer restrictions.
Does Opus 5 replace every other Claude model?
No. Opus 5 is the best generalist choice for coding, analysis, and scientific research, but Mythos 5 remains stronger for long-running autonomous biology tasks and advanced offensive cybersecurity. Fable 5, for its part, is still the most capable publicly available model, at a noticeably higher price.
