Anthropic has released Claude Opus 5, a new artificial intelligence model built for coding agents, complex business tasks and professional knowledge work. Its main selling point is not simply that it is more intelligent than the previous Opus model. Anthropic is positioning it as a practical balance between performance, speed and price.
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens. That is the same price as Claude Opus 4.8 but half the API price of Claude Fable 5. Anthropic says the new model approaches Fable-level performance on many defined coding and knowledge-work tasks, although Fable 5 remains the stronger option for some highly autonomous projects that continue for several days.
The release could matter most to businesses that are already using AI agents in production. For these teams, the real cost of a model includes more than token charges. Failed attempts, unnecessary tool calls, slow responses and human review can all make an apparently affordable model expensive.
This article explains Claude Opus 5 pricing, benchmark results, technical specifications, safety controls, availability, limitations and how it compares with other Claude models.
Key takeaway: Claude Opus 5 is designed to provide much of Anthropic’s frontier-level performance without forcing businesses to pay Fable 5 prices for everyday work.
What Is Claude Opus 5?
Claude Opus 5 is Anthropic’s newest Opus-tier model and the first Opus release in the company’s fifth-generation model family.
Anthropic describes it as a thoughtful and proactive model designed for daily use. It became the default model for Claude Max and the strongest Opus model available to Claude Pro users when it launched on July 24, 2026.
The model is aimed at demanding tasks such as:
- Writing and reviewing software
- Debugging large codebases
- Operating AI agents
- Analysing financial information
- Reviewing legal documents
- Building reports and presentations
- Conducting scientific analysis
- Automating multi-step business processes
Opus 5 sits between Claude Sonnet and Claude Fable in Anthropic’s product range. Sonnet is generally better suited to lower-cost, high-volume work. Opus 5 targets difficult everyday tasks, while Fable 5 remains Anthropic’s premium option for the longest and most autonomous projects.
Who Is Claude Opus 5 Built For?
Claude Opus 5 is most relevant to developers and organisations that need more judgement and consistency than a lower-cost model can provide.
A customer-support chatbot answering predictable questions may not require an Opus-level model. A coding agent that must investigate a bug, inspect several files, run tests and correct its own mistake may benefit much more.
The same applies to financial, scientific and legal tasks. The model’s value depends on whether its stronger reasoning reduces errors, retries and manual checking.
Claude Opus 5 Technical Specifications
Official cloud documentation provides several important details that were missing from much of the launch-day news coverage.
| Specification | Claude Opus 5 |
|---|---|
| Release date | July 24, 2026 |
| Context window | 1,000,000 tokens |
| Maximum output | 128,000 tokens |
| Supported inputs | Text, images and PDFs |
| Output format | Text |
| Knowledge cutoff | May 2026 |
| API model name | claude-opus-5 |
| Reasoning | Supported |
| Computer use | Supported |
| Function calling | Supported |
| Prompt caching | Supported |
| Google Cloud status | Generally available |
Google Cloud lists a maximum input of one million tokens and a maximum output of 128,000 tokens. AWS documentation also lists the May 2026 knowledge cutoff and support for adaptive reasoning.
A large context window allows the model to process extensive codebases, lengthy reports or collections of documents in one workflow. However, a larger advertised context window does not guarantee equal accuracy across every part of a very long prompt. Businesses should still test retrieval quality and instruction-following on their own documents.
Claude Opus 5 Pricing
Claude Opus 5 uses the following standard API rates:
- Input: $5 per million tokens
- Output: $25 per million tokens
These prices match Opus 4.8 and are half the rates charged for Claude Fable 5.
Realistic Cost Example
Consider an agentic coding session that uses:
- 200,000 input tokens
- 20,000 output tokens
The estimated Claude Opus 5 cost would be:
- Input cost: $1
- Output cost: $0.50
- Total: $1.50
The same token usage at Fable 5’s listed price would cost approximately $3.
This is only a basic comparison. A model that uses fewer turns or solves a problem without restarting may produce a lower total cost even when its token rate is higher. Conversely, a cheaper model may cost more if it repeatedly fails.
Claude Opus 5 Fast Mode
Anthropic also offers Fast mode, which runs Opus 5 at around 2.5 times its standard speed. Fast mode costs twice the model’s normal base price.
For example, a large workflow using one million input tokens and 100,000 output tokens would cost approximately:
- Standard mode: $7.50
- Fast mode: $15
Fast mode may make sense for interactive coding, time-sensitive operations or customer-facing applications. It may be unnecessary for offline research and batch processing where speed is less important.
Cost warning: Token price is only one part of the bill. Teams should also measure retries, latency, tool calls, human-review time and the cost of serious errors.
Claude Opus 5 vs Fable 5
The most common question about the launch is whether Opus 5 is better than Claude Fable 5.
The answer depends on the task.
Anthropic reports that Opus 5 matches or exceeds Fable 5 on several bounded evaluations. These are tasks with a defined goal and a clear point at which the model can check whether it succeeded.
Fable 5 remains more suitable for highly ambitious, long-running projects that may operate autonomously for hours or days. Anthropic product leader Dianne Penn advised customers to choose Opus 5 for value and Fable 5 for “days-long, very autonomous projects.”
Where Opus 5 May Be the Better Choice
Choose Claude Opus 5 when:
- The task has a measurable outcome.
- Price matters.
- You need strong coding and debugging.
- The workflow requires repeated self-checking.
- A project lasts minutes or hours rather than days.
- You need fewer cybersecurity restrictions than Fable 5.
- You are replacing Opus 4.8 at the same API price.
Where Fable 5 May Remain Stronger
Choose Fable 5 when:
- A project runs autonomously for several days.
- The task involves a long chain of connected decisions.
- Maximum capability matters more than API price.
- The system must maintain goals over an unusually long horizon.
- Your organisation has tested Fable 5 and measured a better cost per successful result.
Calling Opus 5 universally “better than Fable 5” would therefore be misleading. It may be more economical and stronger on selected tests without being the best choice for every workload.
How Claude Opus 5 Performs on Coding and Agent Benchmarks
Anthropic reports strong performance across software engineering, computer use, automation and professional knowledge work.
On Frontier-Bench v0.1, the company says Opus 5 more than doubles Opus 4.8’s performance while achieving a lower cost per completed task. On CursorBench 3.2, its maximum-effort result comes within 0.5% of Fable 5’s peak score at half the cost per task.
Anthropic also reports that:
- Opus 5 scores three times higher than the next-best model on ARC-AGI 3.
- Its AutomationBench pass rate is around 1.5 times the next-best model at the same cost.
- It surpasses Fable 5’s strongest OSWorld 2.0 result at just over one-third of the cost.
- It improves by 10.2 percentage points over Opus 4.8 on an internal organic chemistry test.
- It improves by 7.7 percentage points on a protein-related evaluation.
These numbers are noteworthy, but they should be read carefully. Most were published by Anthropic or launch partners. They are not the same as independent, peer-reviewed production tests.
Benchmark results can change according to:
- Prompt design
- Reasoning-effort setting
- Available tools
- Token budget
- Number of attempts
- Scoring method
- Task duration
- Whether cost is matched across models
A model that wins a 30-minute benchmark may not be better at maintaining a complex project over three days. Likewise, a model with a lower maximum score may still be more economical in production.
How Claude Opus 5 Checks Its Own Work
One of Anthropic’s main claims is that Opus 5 is better at verifying and correcting its output.
In one company-reported test, the model received an image of a machine component but no direct visual tool for inspecting it. It created a computer-vision pipeline, extracted the geometry and reconstructed the component as a three-dimensional model. Anthropic says competing models failed after five attempts under the same setup.
In another example, Opus 5 investigated a software bug, identified its root cause and corrected an edge case that an existing community patch had missed.
This type of behaviour matters because many AI errors are not obvious. A coding agent can produce software that appears correct but breaks during unusual inputs. A financial agent can calculate the right total while applying the wrong business rule.
Better self-verification could reduce:
- Superficial bug fixes
- Invalid assumptions
- Incomplete tests
- Incorrect tool use
- Human-review time
- Repeated agent runs
However, self-checking does not make the model error-free. A model may confidently verify work using the same mistaken assumptions that produced the original answer. External tests and human oversight remain necessary for high-impact decisions.
Enterprise Use Cases
Claude Opus 5 is being marketed as more than a coding model.
Software Development
Development teams can use it for debugging, code review, feature implementation, test generation and large-scale codebase changes.
The strongest use case may be complex issues where finding the root cause matters more than writing a small block of code.
Financial Analysis
Launch partners reported improvements in numerical reasoning, table analysis and specialist financial workflows. One customer said its tests showed higher accuracy with fewer turns, fewer tool calls and less completion time, although these remain vendor-selected testimonials.
Legal Work
Legal teams may use the model for contract review, first-pass redlining, due-diligence support and document comparison.
Confidentiality, jurisdiction-specific law and professional review remain essential. Claude Opus 5 should support legal professionals rather than replace final legal judgement.
Business Automation
On AutomationBench, Anthropic says the model completed multi-step business tasks such as reviewing account-health information, identifying customers at risk and triggering the correct follow-up process.
This could support:
- Customer-retention workflows
- Report generation
- Compliance checks
- Operational monitoring
- Data reconciliation
- Internal research
Scientific Research
Anthropic reports improvements in structural biology, organic chemistry and bioinformatics. The company nevertheless states that Mythos 5 remains stronger for the longest autonomous biological research tasks.
Research organisations should validate scientific outputs against experimental data and established literature.
Safety Controls and Automatic Model Fallback
Anthropic says its automated behavioural audit gave Opus 5 an overall misalignment score of 2.3, the lowest among its recent models. The company also says it observed lower rates of deceptive behaviour and less susceptibility to misuse. These are Anthropic’s internal findings, not an independent safety certification.
The model’s cybersecurity controls are less restrictive than those used for Fable 5.
Opus 5 may analyse vulnerabilities in source code, but its standard safeguards block or restrict:
- Binary-based vulnerability scanning
- Penetration testing
- Exploit generation
- Other higher-risk cyber activities
Anthropic expects Opus 5’s classifiers to intervene around 85% less often than those protecting Fable 5.
When a request is flagged in Claude.ai, Claude Code or Claude Cowork, the system can fall back to Opus 4.8. API users can also enable fallback behaviour.
This raises an important enterprise issue: a workflow may start with one model and finish with another. Regulated organisations should log fallback events and test whether model changes affect accuracy, tone, policy compliance or reproducibility.
Claude Opus 5 Availability
Claude Opus 5 launched across Anthropic’s platforms and cloud partners.
It is available through:
- Claude Pro
- Claude Max
- Claude Team
- Claude Enterprise
- Anthropic’s API
- Amazon Bedrock
- Claude Platform on AWS
- Google Cloud
AWS announced availability through Amazon Bedrock in regions including US East, Asia Pacific and Europe. Google Cloud lists the model as generally available with text, image and PDF inputs.
Developers using Anthropic’s direct API can access it with the model name:
claude-opus-5
Organisations should check regional access, quotas, data-processing terms and platform-specific prices before deployment.
Claude Opus 5 Limitations
Claude Opus 5 brings meaningful improvements, but it also has clear limitations.
Most Evidence Comes From Anthropic
The launch benchmarks and customer examples are useful, but many come from Anthropic and selected early-access partners.
Independent testing may produce different results.
It Can Still Make Mistakes
Self-verification reduces some errors but does not eliminate hallucinations, incorrect assumptions or incomplete analysis.
It Costs More Than Lower-Tier Models
High-volume, predictable work may be cheaper on Sonnet or another smaller model.
Fast Mode Doubles the Base Price
Higher speed may improve user experience, but it can quickly increase production costs.
Safety Fallbacks Can Change Behaviour
An automatic move to Opus 4.8 may affect output quality or consistency during a workflow.
Fable Still Leads on Some Long-Horizon Tasks
Opus 5 should not be presented as a complete replacement for Anthropic’s premium model.
Should You Upgrade From Opus 4.8?
Opus 5 is priced at the same standard API rate as Opus 4.8, making it an attractive upgrade candidate. Still, teams should not switch every production workflow without testing.
A practical migration process is:
- Select representative Opus 4.8 tasks.
- Save quality, cost, latency and failure-rate baselines.
- Run the same tests with Opus 5.
- Compare tool calls, retries and human-review time.
- Test maximum and lower effort settings.
- Inspect safety and fallback behaviour.
- Roll out gradually with monitoring.
Migration makes the most sense when Opus 5 improves task completion without creating unacceptable changes in speed, cost or behaviour.
Which Claude Model Should You Choose?
Choose the model according to the task rather than assuming the newest or most expensive option is always best.
- Use Sonnet for lower-cost, high-volume and predictable work.
- Use Claude Opus 5 for difficult everyday coding, analysis and agent workflows.
- Use Fable 5 for the longest, most autonomous and capability-intensive projects.
The final decision should come from internal evaluation. Public benchmarks can narrow the choices, but production data should determine the model you deploy.
Conclusion
Claude Opus 5 represents an important change in how leading AI companies compete. The goal is no longer simply to publish the model with the highest benchmark score. Anthropic is also competing on reliability, token efficiency, latency, safety controls and the total cost of completing real work.
For many developers and enterprise teams, Claude Opus 5 may offer a more practical balance than Fable 5. It provides major reported improvements over Opus 4.8 at the same base API price while approaching Fable-level performance on many clearly defined tasks.
That does not make it universally superior. Fable 5 still has a role in long-running autonomous projects, while Sonnet may remain the smarter choice for simpler, high-volume workloads.
Before moving a production system, test Claude Opus 5 against your own data, prompts, tools and success criteria. The best AI model is not the one that wins the most launch-day charts. It is the one that completes your work accurately, safely and at a sustainable cost.
