Claude Opus 5 is available today. If you work with code, analysis, or products that need long-form reasoning, this matters to you: Opus 5 promises intelligence close to the most advanced models, but at half the price and with more efficiency for everyday use. Sounds interesting, right?
Performance and cost: more for less
Opus 5 presents itself as a faster, cheaper version compared to its predecessor, Opus 4.8. Anthropic says that on programming and knowledge work tasks—like Frontier-Bench and GDPval-AA—Opus 5 sets a new reference. It isn’t the top performer in every niche (Mythos 5 still leads in offensive cybersecurity), but it shines where most teams care: productivity and cost per task.
- In software engineering benchmarks, Opus 5 outperforms other models and doubles the performance of Opus 4.8 with a lower cost per task.
- In CursorBench, at maximum effort, it’s 0.5% behind Fable 5 while costing half as much per task.
- On novel-problem challenges like ARC-AGI 3, its score triples that of the next-best model.
What’s the practical result? More tasks solved using fewer tokens and with lower latency. For teams who bill by usage or want to iterate quickly, that means less friction and more predictable costs. Want fewer surprises on your bill?
Real use cases: examples that matter
Opus 5 doesn’t just impress on paper: in internal tests and with early customers it showed behaviors we often ask of AI but rarely see in practice.
- Faced with an image of a mechanical part without direct viewing access, Opus 5 wrote its own computer-vision pipeline, extracted the geometry, and rebuilt the 3D model in FreeCAD. It repeated the task successfully where other models failed.
- It found the root cause of a bug in a popular package manager and fixed it correctly; another model only patched the symptom.
- An engineer used Opus 5 to build a complete market data feed in a single session; lacking an external validation source, Opus 5 created its own test harness.
These examples highlight something important: Opus 5 is more methodical, verifies its work, and can iterate more autonomously across long workflows.
Science, visuals and specialized work
Opus 5 also improves on scientific and visual tasks. In internal life-science evaluations it surpasses Opus 4.8 across the board: from organic chemistry (notable improvements inferring structures from spectroscopy) to predicting effects of mutations in proteins.
On the visual side, it produces stronger outputs: better animations, games, and 3D work, and cleaner deliverables for presentations and dashboards. In short, the polished stuff you show stakeholders gets easier to produce.
Judgment and consistency: what changes day to day
One phrase testers repeat: Opus 5 has better judgment. It doesn’t rush into writing— it spots logical flaws in its plan, verifies results by independent methods, and explains why an answer is correct.
For product managers and teams that rely on steady quality, that means fewer manual reviews, fewer back-and-forths, and more confidence delegating complex tasks to the AI. Wouldn’t that free up time for higher-value work?
Security and alignment
Anthropic reports Opus 5 is their most aligned model so far: better scores in automated audits, less deceptive behavior, and reduced susceptibility to misuse instructions. Important points:
- Opus 5 does not push dual-use capabilities to the limit. In biology and offensive cyber it remains behind Mythos 5.
- While Opus 5 can identify vulnerabilities nearly as well as Mythos 5, it does not reach the same effectiveness at developing exploits.
About filters and guardrails:
- Its cybersecurity classifiers are less restrictive than Fable 5’s, allowing you to find bugs in code, but they block more sensitive tasks like binary scanning, penetration testing, and exploit generation.
- When a request is flagged on Claude.ai or in Claude products, it defaults to falling back to Opus 4.8. In the API you can configure flagged requests to automatically fallback to another model.
- Companies and research centers in the Cyber Verification Program (CVP) can access a version with fewer restrictions for legitimate security work.
In biology, the same mitigation routes apply: Opus 5 is the most capable general model available for research, but Anthropic remains cautious with long-running autonomous tasks that could pose risks.
How to get started and pricing
Opus 5 is available today on all platforms. Announced prices:
- 5 USD per million input tokens
- 25 USD per million output tokens
There’s also a Fast mode that runs about 2.5× speed at roughly double the cost. In the API the model is named claude-opus-5, and you can enable automatic fallbacks if you want.
Useful betas launched with Opus 5:
- Mid-conversation tool changes on the Claude platform without invalidating the prompt cache.
- Automatic fallbacks in the API so flagged requests redirect to the best available model.
What does this mean for you?
If you work in product development, financial analysis, scientific research, or business automation, Opus 5 promises to save time and tokens without sacrificing quality. For people in security or biology, the recommendation is to evaluate case by case: Opus 5 helps a lot to discover problems, but for exploitation or high-risk research you still need controls and specialized models.
Opus 5 isn’t cosmetic: it’s a bet on making models more useful in real flows—more reliable and more economical. Can you imagine delegating the first pass of a code review or creating a complex report? With Opus 5, that possibility is closer.