Anthropic has released Claude Opus 5, its newest flagship artificial intelligence model, making it available simultaneously across all major platforms on July 24. The model can be accessed through the company’s API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and the Claude applications. Anthropic positions the launch not as an incremental refinement but as a significant advance, stating that Opus 5 surpasses all competitors in coding and knowledge-intensive tasks.
Pricing and availability across subscription tiers
For subscribers on the Claude Max plan, Opus 5 is now set as the default model. Claude Pro users gain access to it as the most capable selectable option, while those on the free tier remain limited to Sonnet 5 and do not receive Opus 5 access at all. This stands in contrast to some competing offerings, such as Kimi K3, which can be tried without a paywall. Anthropic maintains that its most powerful model remains a paid product.
The company has drawn attention to what it describes as a cost advantage. Through the API, Opus 5 is priced at $5 per million input tokens and $25 per million output tokens—identical to the pricing of its predecessor, Opus 4.8. The reference to half price concerns Fable 5, Anthropic’s higher-tier model, making Opus 5 the more affordable option for near-premium performance. For users who prioritize speed, a Fast mode doubles throughput and is currently available exclusively through the Claude API. Private subscribers are billed a flat subscription fee rather than per token, insulating them from these usage costs directly.
Built-in reasoning and adjustable control levels
A practical shift in everyday use is that Opus 5 engages internal reasoning by default, unlike Opus 4.8, which required users to manually enable it. The model autonomously determines when to think and at what depth. Users can adjust this behavior across five settings: low, medium, high, xhigh, and max. High is the standard setting, though Anthropic recommends low and medium for many common tasks, noting they can produce solid results with far fewer tokens.
Two consequences follow from this design. Answers tend to run longer out of the box, and the model checks its own work without prompting. Both factors consume quota more quickly. Pro subscribers who regularly hit usage limits are advised to lower the reasoning level accordingly.
Technical specifications and safety architecture
Opus 5 operates with a context window of one million tokens, which serves as both the default and the maximum; no smaller variant is offered. Output can extend up to 128,000 tokens. In practical terms, this allows users to feed in extensive contracts, complete manuals, or large project folders in a single pass without splitting them into segments.
Anthropic describes Opus 5 as its most aligned model to date, registering the lowest rate of deceptive behavior the company has recorded. At the same time, safety filters for cybersecurity requests have been relaxed, triggering roughly 85 percent less often than on Fable 5. When a request is blocked, the Claude apps automatically fall back to Opus 4.8. On offensive security code generation, Anthropic’s own testing still places Opus 5 considerably behind Mythos 5.
It should be noted that the top benchmark results cited come from Anthropic’s internal evaluations. Independent head-to-head comparisons with models such as OpenAI’s GPT-5.6 and leading Chinese alternatives have not yet appeared. For the time being, Opus 4.8 remains available across all platforms.
Sources: www.anthropic.com, platform.claude.com