Claude Opus 5 Explained: What Changed, Pricing and Who Should Switch

Anthropic launched Claude Opus 5 on July 24, 2026, describing it as a thoughtful, proactive model built for everyday coding and knowledge work.

The company says Opus 5 leads evaluations including Frontier-Bench and GDPval-AA, approaches Fable 5 on selected coding tests at half the cost per task, and remains behind the restricted Claude Mythos 5 on advanced cybersecurity work. It is now the default model for Claude Max and the strongest Claude model available to Pro subscribers.

Those claims make Opus 5 sound like another straightforward model upgrade.

It is more consequential than that.

The most important change is not a single benchmark score. Anthropic has narrowed the practical gap between Opus and Fable while giving customers more control over how much intelligence—and therefore how much token usage—they apply to each task.

That changes how Claude users should build their AI stack.

The 60-second answer

Claude Opus 5 is Anthropic's strongest Opus model so far.

It keeps the same $5 input and $25 output pricing per million tokens as Opus 4.8, compared with $10 input and $50 output for Fable 5. It is available to Claude Pro, Max, Team and Enterprise users, through the Claude API and across Anthropic's supported cloud platforms.

For developers, the model identifier is claude-opus-5.

Opus 5 has a one-million-token context window, supports output of up to 128,000 tokens and offers five effort settings: low, medium, high, xhigh and max.

For most serious Claude users, Opus 5 should now be the premium model tested first.

Fable 5 still makes sense when the task is exceptionally difficult, long-running or valuable enough that even a small improvement in capability outweighs the additional cost.

For ordinary writing, summaries, basic research and simple coding, both models are likely to be unnecessary.

What actually changed with Claude Opus 5?

1. Opus performance moved much closer to Fable

Anthropic says Opus 5 substantially outperforms Opus 4.8 across coding, automation, computer use, scientific work and professional analysis.

On Frontier-Bench v0.1, Anthropic reports that Opus 5 surpassed every other tested model and achieved more than twice the performance of Opus 4.8 at a lower cost per completed task.

On CursorBench 3.2, Opus 5 at maximum effort reportedly finished within 0.5% of Fable 5's peak result while costing half as much per task.

That does not prove Opus 5 is universally equivalent to Fable 5.

It does show the performance gap may now be small enough on some practical coding workloads that paying twice the standard token price becomes difficult to justify.

2. Effort controls are now central to the buying decision

Opus 5 supports five effort levels:

Effort is not a fixed token budget. It is a behavioral instruction that changes how deeply Claude reasons before and during a task.

At lower effort levels, the model can use fewer thinking tokens and respond faster. At higher levels, it spends more computation exploring, checking and refining its work. Anthropic recommends starting with xhigh for coding and agentic tasks, high for most other intelligence-sensitive workloads and max only when a task genuinely justifies unconstrained token spending.

This is the feature that matters most for buyers.

The headline API price tells you what each token costs. The effort setting helps determine how many tokens the model consumes to complete the job.

A team no longer needs to make one permanent decision between a cheaper model and a more capable one. It can run Opus 5 at lower effort for routine stages, increase effort for difficult reasoning and reserve Fable for the small number of tasks where Opus still falls short.

That is a more useful form of cost control than choosing one model for everything.

Related: The Unlimited AI Subscription Is Dying

How strong are the benchmark results?

Anthropic reports strong results across several categories:

These are meaningful results, but they need careful interpretation.

Some figures come from Anthropic's own testing setup, and benchmark performance does not guarantee the same advantage inside every codebase, business process or research workflow.

Anthropic notes, for example, that its Frontier-Bench figures came from an internal run using a particular agent harness and backend, averaged across five attempts per task.

The responsible conclusion is not that Opus 5 has conclusively beaten every other model.

It is that Anthropic has provided enough evidence for existing Claude users to test Opus 5 seriously against their own work.

Claude Opus 5 pricing

Opus 5 keeps the same base pricing as Opus 4.8 and costs half as much as Fable 5 at standard API rates.

Anthropic also offers Fast mode, which runs at around 2.5 times the default speed and costs twice the standard Opus 5 rate.

That creates an important tradeoff.

Fast mode removes the headline token-price advantage over standard Fable 5. It may still be worthwhile where response time matters more than raw model capability, but it should not be treated as a free performance upgrade.

The right comparison is not simply price per token. It is:

• Total tokens consumed
• Time taken
• Number of corrections required
• Whether the task was completed properly
• How much human supervision was needed

A model with a cheaper published price can still cost more if it uses substantially more tokens or repeatedly needs correction.

Claude Opus 5 versus Claude Fable 5

The data-retention difference deserves more attention than it will probably receive in launch coverage.

Anthropic says Opus 5 does not require additional data retention for general access. Fable 5, by contrast, requires 30-day retention for safety monitoring.

For organizations handling sensitive or confidential work, that may make Opus 5 preferable even before pricing and benchmarks are considered. Businesses should still assess the full terms applying to their account, platform and cloud provider rather than treating this as a complete privacy review.

The practical distinction is now straightforward:

Opus 5 is the model to run regularly. Fable 5 is the model to escalate to.

Who should switch to Claude Opus 5?

Existing Opus 4.8 users

Opus 4.8 users have the clearest decision.

Opus 5 costs the same at standard API rates and Anthropic reports substantial improvements in difficult coding, tool use, analysis and long-running work. Unless a production workflow depends on a specific Opus 4.8 behavior, Opus 5 should be tested immediately.

That does not mean replacing the model blindly. Run both versions against a representative task set before moving critical production workloads.

Fable 5 users

Fable users should not necessarily abandon it.

Instead, they should identify which tasks genuinely need it.

A sensible model-routing system might use:

• Opus 5 at low or medium effort for routine subtasks.
• Opus 5 at high or xhigh effort for serious everyday work.
• Fable 5 for problems where Opus fails, requires excessive correction or produces materially worse outcomes.

For many teams, that could preserve most of Fable's practical value while materially reducing overall model spend.

Developers and AI-agent builders

Opus 5 appears particularly well suited to:

• Debugging unfamiliar systems
• Root-cause analysis
• Large codebase changes
• Multi-file feature development
• Code review
• Long-running tool loops
• Business-process automation
• Agents that must verify their own work

Anthropic says the model is more persistent, more careful about checking results and more likely to address underlying causes instead of fixing only the visible symptom. Those are vendor-reported findings, but they target exactly the failure patterns that make production agents expensive to supervise.

Researchers, consultants and professional teams

The model may also suit work involving:

• Financial analysis
• Legal documents
• Due diligence
• Scientific research
• Complex spreadsheets
• Presentations
• Large document collections
• Multi-stage reports

Anthropic's early-access customers reported improvements in numerical reasoning, financial modeling, legal work, data analysis, slide production and long-horizon research. These are customer reports selected for a launch announcement rather than independent tests, so they should be treated as evidence of intended use rather than guaranteed performance.

Everyday Claude users

Most everyday users do not need to change their stack because Opus 5 exists.

If your regular tasks are short emails, simple summaries, basic brainstorming, casual chat or lightweight content, a cheaper and faster model will usually remain the better choice.

Paying for more reasoning only makes sense when the task benefits from more reasoning.

Four practical examples

A founder building a complex product

This user regularly asks Claude to inspect a repository, diagnose bugs, change several files, run tests and verify the result.

Recommendation: Test Opus 5 at xhigh effort. Keep Fable available for tasks Opus cannot complete reliably.

This is the type of workload where the new model's coding, persistence and self-verification improvements are most likely to matter.

An agency producing reports and presentations

This user combines research, spreadsheets, written analysis and finished client deliverables.

Recommendation: Start with high effort and compare quality against the current model. Increase to xhigh only for difficult analytical work.

Running every stage at maximum effort would probably waste tokens. The advantage is being able to raise intelligence selectively.

A marketer writing short social posts

This user needs quick captions, rewrites and email drafts.

Recommendation: Stay with a cheaper model.

Opus 5 may produce excellent work, but the task is unlikely to justify premium-model economics.

A cybersecurity researcher

Opus 5 may identify vulnerabilities in source code, but its safeguards block binary-based vulnerability scanning, penetration testing and exploit generation. Anthropic says it remains substantially behind Mythos 5 at turning identified vulnerabilities into functioning exploits.

Recommendation: Review the restrictions carefully before switching. The model may be useful for defensive source-code work, but it is not an unrestricted security model.

Safeguards and automatic fallbacks

Anthropic says Opus 5 is less restricted than Fable 5 in cybersecurity.

The model can identify vulnerabilities in source code, while classifiers block narrower categories including binary scanning, penetration testing and exploit generation. Anthropic expects those classifiers to intervene approximately 85% less often than Fable 5's.

When a request is flagged inside Claude.ai, Claude Code or Claude Cowork, it falls back to Opus 4.8 by default rather than ending with a simple refusal.

API developers can also enable automatic fallbacks so flagged Opus 5 or Fable 5 requests route to another available model.

This improves continuity, but it creates a visibility issue teams should account for: the model answering a request may not always be the model originally selected.

For sensitive or high-stakes workflows, log fallback events and make sure model identity is visible in evaluation and audit systems.

Anthropic also describes Opus 5 as its most aligned model to date. Its automated behavioral audit gave the model an overall misaligned-behavior score of 2.3, lower than the company's other recent models. That is an internal safety measure, not an independent capability score, and should be understood in that context.

Availability

Claude Opus 5 is available to Claude Pro, Max, Team and Enterprise users.

Anthropic says it is available natively through the Claude Platform and through Amazon Web Services, Google Cloud and Microsoft Foundry. Developers can access it using claude-opus-5.

It is the default model for Claude Max and the strongest Claude model available on Claude Pro.

Subscription availability does not mean unlimited use. Users should review the usage limits, credits and model-access rules applying to their specific plan.

The Choosely verdict

Claude Opus 5 is not important because it wins another set of benchmark charts.

It is important because Anthropic has made its premium model hierarchy more usable.

Opus 5 brings near-Fable performance on selected workloads into a model that costs half as much at standard API rates, has no special 30-day retention requirement for general access and gives users direct control over how much reasoning they want to purchase.

That makes it the most sensible premium Claude starting point for a large share of serious users.

Fable 5 still has a role. It remains the higher capability option for exceptionally difficult, long-running or valuable work. But it should no longer be the automatic answer whenever a task feels hard.

Our recommendation:

• Opus 4.8 users: Test Opus 5 now.
• Fable users: Move regular workloads to Opus 5 and keep Fable as an escalation path.
• Developers and agent builders: Test multiple effort levels, not just the default.
• Everyday users: Do not upgrade unless your current model is genuinely failing.
• Businesses: Compare retention requirements, safeguards and total completed-task cost—not only benchmark scores.

The most powerful model is not automatically the best model for your stack.

The best model is the least expensive one that completes your real work reliably.

Opus 5 makes that line easier to find.

What Choosely verified

Every material launch date, price, availability statement and benchmark figure in this article was checked against Anthropic's launch announcement or developer documentation on July 25, 2026.

• Anthropic: Introducing Claude Opus 5
• Anthropic: Claude Opus model and pricing information
• Claude Platform Docs: Effort settings
• Claude Platform Docs: What's new in Claude Opus 5
• Anthropic: Claude Fable 5 and Claude Mythos 5

Availability can vary by plan, geography, account eligibility, workspace controls and staged rollout. Recheck the relevant account and live pricing page before changing a production workflow.