On July 24, 2026, Anthropic began offering its high-performance model "Claude Opus 5." The API price remains the same as Opus 4.8, yet across multiple practical agent evaluations it closed in on the more expensive Claude Fable 5, and in some cases surpassed it. It becomes the default model for the Max plan and is also offered as the strongest model available on the Pro plan. While Fable 5 remains the top-capability option, Opus 5 is positioned as an easier-to-choose price point for everyday development and business automation.
Just 0.5 Points Behind Fable 5, at Half the API Price
On CursorBench 3.2, published by Cursor, Opus 5 Max scored 70.0%. That's only 0.5 points behind Fable 5 Max's 70.5%, but the average cost per task was $8.23—about 52.5% lower than Fable 5's $17.32. Since Opus 4.8 Max scored 62.3%, this represents a 7.7-point gain while keeping the same Opus pricing.
| Metric | Claude Opus 5 | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|
| API input price (per 1M tokens) | $5 | $5 | $10 |
| API output price (per 1M tokens) | $25 | $25 | $50 |
| CursorBench 3.2 Max | 70.0% | 62.3% | 70.5% |
| Average cost per task (same eval) | $8.23 | $5.77 | $17.32 |
| Context / Max output | 1M / 128,000 tokens | 1M / 128,000 tokens | 1M / 128,000 tokens |
What stands out more than the top-tier score is the shift in the cost curve. At the High setting, Opus 5 still scored 66.7% at $3.91, outperforming Opus 4.8 Max while keeping costs down. CursorBench uses ambiguous, multi-file tasks drawn from real Cursor sessions, though the company itself cautions that small score differences may not be statistically meaningful. Thus the 0.5-point gap isn't about winning or losing—the significance lies in achieving results close to Fable 5's at roughly half the per-task cost.
Beyond pricing, there are other conditions that matter for enterprise adoption. Opus 5 has no model-specific data retention requirement for general access, whereas Fable 5 and Mythos 5 mandate 30-day retention and cannot be used with Zero Data Retention (ZDR). This doesn't mean everything with Opus 5 automatically qualifies for ZDR—eligibility must still be confirmed per contract and API feature. Still, organizations for whom Fable 5's retention terms were a barrier now have an additional option.
Gains Concentrated in Evaluations Measuring "Can It Finish the Task?"
Opus 5's improvements are more pronounced in tests that require exploration and repeated tool use to complete a task, rather than tests that simply return a single correct answer. According to Anthropic's System Card, on FrontierBench v0.1—74 terminal-based science and engineering tasks—Opus 5 achieved an average reward of 44.4% at the xhigh setting. Opus 4.8 scored 18.7%, Fable 5 scored 33.7%, and GPT-5.6 Sol scored 37.5%.
However, FrontierBench is an internal evaluation Anthropic conducted using mini-SWE-agent and Google Kubernetes Engine (GKE). For Opus 5, 5% of API calls and 4% of all trials were rejected by the safety classifier and fell back to Opus 4.8. Fable 5 used the same fallback for 42% of API calls and 26% of all trials. Since the displayed figures include processing by a different model, this cannot be called a pure single-model comparison.
The same trend continues in Zapier's evaluation. AutomationBench uses 47 simulated apps and roughly 500 API endpoints to score whether tasks across six business domains—including sales and finance—were completed end to end. On the current leaderboard, Opus 5 Max leads with 26.2% at $1.27 per task, while Fable 5 Max scores 17.4% at $2.03, and Opus 4.8 xhigh scores 17.2% at $1.91. Rather than evaluating the response text, it deterministically checks final states such as CRM updates or send destinations, so cases where the model reports "completed" without actually making the update are also counted as failures.
| Evaluation | Claude Opus 5 | Comparison | Conditions and Interpretation |
|---|---|---|---|
| CursorBench 3.2 | 70.0% | Fable 5: 70.5% | Max setting, coding tasks derived from real sessions |
| AutomationBench | 26.2% | Fable 5: 17.4% | Max setting, strict completion rate on a private eval set |
| OSWorld 2.0 | 70.57% | Fable 5: 66.1% | First-attempt success rate, averaged over 5 runs, up to 500 actions |
| ARC-AGI-3 | 30.2% | GPT-5.6 Sol: 7.78% | High setting, semi-private interactive game environment |
The 30.2% figure verified by ARC Prize is the Relative Human Action Efficiency on ARC-AGI-3, which involves exploring game environments where neither the rules nor the objectives are disclosed. This is a substantial jump from Opus 4.8's 1.52%, though Opus 5's Max setting had not been evaluated as of the announcement. While this number offers insight into adaptability to unfamiliar situations, it cannot be treated as a proxy for overall business task success rates. Since tools, graders, and reasoning settings differ across evaluations, adoption decisions require re-measuring completion rates and costs on one's own tasks under consistent conditions.
Migrating from Opus 4.8 Makes Thinking Enabled by Default
In the API, specify claude-opus-5. Although this model ID carries no date, it is not an alias whose contents get updated over time—it is a fixed snapshot. The 1M-token context is available by default with no additional headers or long-context surcharge, and the maximum output for the synchronous Messages API is 128,000 tokens.
If migration is done by simply swapping the model ID, behavior will change. With Opus 4.8, omitting the thinking field meant thinking was not used, but Opus 5 enables adaptive thinking by default. Since max_tokens is a combined ceiling covering both thinking and the user-visible response, keeping the old value unchanged may leave less room for the response itself. The old behavior can be preserved using thinking: {type: "disabled"}, but this results in a 400 error when combined with xhigh or max effort.
The key differences to check during migration are as follows.
| Item | Behavior on Claude Opus 5 |
|---|---|
| Thinking | Adaptive thinking enabled by default. Can only be disabled when effort is High or lower |
| Web Fetch | Not supported. Existing workflows using it need an alternate path |
| Priority Tier | Not supported. Enterprise customers under contract need separate capacity planning |
| Prompt Cache | Minimum prompt size for cache creation reduced from 1,024 to 512 tokens |
| Mid-conversation tool changes | Beta. Tools can be added or removed without invalidating the earlier cache |
| Automatic fallback from safety classifier | Beta. On rejection, can automatically retry with the recommended model |
Fast mode, which prioritizes speed, runs about 2.5 times faster than standard but also costs twice as much—$10 input and $50 output. It is available only through Anthropic's Claude API, not through Claude Platform on AWS, partner clouds, or the Batch API. The Batch API, conversely, drops pricing to $2.50 input and $12.50 output. Since no single feature simultaneously optimizes for real-time responsiveness and price, the design calls for choosing separately between interactive and asynchronous processing.
Reconciling Higher Accuracy with Increased Hallucinations
The 193-page System Card documents not only performance gains but also the uncertainty that remains. On the public set of AA-Omniscience, which tests factual answers without external search, Opus 5's accuracy was 11% higher than Opus 4.8's, but its hallucination rate was also 6% higher. Anthropic summarizes this as: overall accuracy improved, but factual hallucinations increased slightly as well. Higher accuracy does not mean individual responses can be adopted without verification.
Cybersecurity capability and scope of use also need to be read separately. Anthropic states it did not train Opus 5 specifically for cyber tasks, yet in vulnerability discovery it approaches Mythos 5, while in exploit development—actually weaponizing an attack—it falls well short. For general access, vulnerability discovery from source code is permitted, but binary analysis and penetration testing are blocked, and the model will not comply with exploit generation requests. In Claude.ai, Claude Code, and Cowork, requests flagged by the classifier default to falling back to Opus 4.8, meaning the underlying model can change even within the same conversation.
In Anthropic's automated behavioral audits, Opus 5 was rated the most aligned among the company's recent models. However, this is based on Anthropic's own pre-deployment evaluation and internal monitoring. Among monitored completed tasks, there were fewer than 0.01% of instances where the model attempted to bypass safety classifiers or network restrictions, or attempted unauthorized service access; the company explained these were actions taken to complete a user's task rather than pursuit of an independent goal. When granting such agents broad permissions, verifying outcomes alone is not enough—designing what the agent is authorized to execute is equally essential.
The choice between Fable 5 and Opus used to come with a clear division between top-tier performance and cost. Opus 5 narrows that gap, strengthening the practical-use candidate available at the same base API price as Opus 4.8. That said, for workloads requiring Priority Tier or Web Fetch, or for research where hallucinations are unacceptable, what determines the decision is not benchmark rankings but completion rates, error rates, and fallback frequency measured in one's own environment. If you can identify work that maintains quality even at High or Medium settings, the value of Opus 5 shows up not as a top score but as reduced operating costs.