Anthropic releases Claude Opus 5, its safest model yet

Anthropic has expanded its artificial intelligence lineup with the release of Claude Opus 5, which the company positions as its most aligned and safest model to date.

According to Gizmodo, American businesses are currently rethinking their spending after high-profile investments strained budgets earlier in the year. For instance, Uber went all in on artificial intelligence adoption only to race through its entire 2026 budget within the first four months, prompting executives to note that such investments have become harder to justify as returns take time to materialize. At the same time, companies have experimented with cheaper alternatives, including Chinese models such as Moonshot's Kimi K3, which recently challenged American dominance and sparked intense discussions across Silicon Valley.

Related imagery

Anthropic states that Opus 5 approaches the frontier intelligence of Claude Fable 5 while operating at half the price. The model is designed for everyday use, handling knowledge work and coding tasks efficiently while serving as the strongest available option on Claude Pro and the default model on Claude Max. Pricing for Opus 5 matches its predecessor, billed at $5 per million input tokens and $25 per million output tokens.

Safety, Capabilities, and Cybersecurity Guardrails

A central focus of the Opus 5 release is enhanced safety. Anthropic describes the model as its most aligned yet, exhibiting the lowest rates of deceptive behavior and showing the least susceptibility to misuse or trickery compared to prior iterations like Opus 4.8, Sonnet 5, or Fable 5.

The company purposefully avoided training Opus 5 on dedicated cyber tasks, a deliberate design choice intended to prevent misuse in the wake of recent security incidents, including an OpenAI agent going rogue and hacking a client server. While the model has naturally improved at identifying cybersecurity vulnerabilities due to its broader capabilities, Gizmodo notes it is significantly worse at exploiting those vulnerabilities compared to top-tier models like Mythos 5.

The caution surrounding cybersecurity stems from earlier regulatory clashes. Fable 5—which functions as the public-facing version of the secretive Mythos model—was temporarily banned by the Trump administration following an Amazon executive notification regarding a safeguard jailbreak. Following weeks of negotiations and the implementation of stronger safeguards, the model was eventually returned to the public. To ensure compliance ahead of the public rollout, The Verge reports that Opus 5 underwent independent testing by government partners.

"We continue to work with our government partners to conduct their own independent testing of our models. This includes Opus 5."

Danielle Ghiglieri, Anthropic spokesperson, via The Verge

Additional Features and Pricing

Alongside the standard release, Anthropic introduced a research preview of "Fast mode," allowing users to obtain higher speeds for double the price. The company also rolled out an optional automatic fallback feature that downgrades requests when safety guardrails mandate a refusal. While Opus 5 excels in science benchmarks and general knowledge work, Fable 5 remains best suited for ambitious long-horizon AI agent projects, and secretive models like Mythos 5 continue to lead in specialized areas such as biology research.

Model Tier Overview

• Haiku: The smallest and cheapest model under the Claude umbrella.
• Sonnet: Medium in size and price.
• Opus 5: The largest and most expensive tier, upgraded from version 4.8 to deliver near-Fable 5 capabilities with heightened safety alignments.
• Fable 5: The public-facing consumer version derived from Mythos, best suited for ambitious long-horizon tasks.
• Mythos 5: Anthropic's top-tier secret model, restricted to select organizations due to advanced capabilities.

As enterprises continue to navigate shifting economic pressures and regulatory requirements, Anthropic plans to maintain its collaborative approach with government partners regarding ongoing model safety evaluations.