Anthropic announced Claude Opus 5 on July 24, 2026, billing it as a major upgrade for the Opus tier with the biggest gains in deep reasoning, agentic tasks (workflows where the AI model takes multi-step actions on its own), and test-time compute — the practice of letting a model spend more processing effort on a single query to produce better answers (Anthropic). The new model narrows the performance gap with Anthropic's top-tier Mythos-class Claude Fable 5 while cutting its price in half, and adds a research-preview "Fast mode" and an automatic fallback that routes declined requests to a lower-tier model.

Pricing stays the same as the previous generation: $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. (A "token" is roughly a word or piece of a word — the unit AI companies use to meter usage.) Anthropic says Opus 5 "comes close to the capabilities of Claude Fable 5 in many domains" while delivering substantially stronger complex coding performance (The Verge). On CursorBench, Opus 5 scores just under Fable 5 (Anthropic). Against external competition, Opus 5 is slightly cheaper than OpenAI's GPT-5.6 (The Verge).

A Deliberate Two-Tier Strategy

The model tiering is intentional. Anthropic positions Opus 5 for enterprise customers as best for knowledge work and biology, the ideal everyday model, while Fable 5 retains the mantle of "ambitious work and long-horizon AI agent projects" (The Verge). Fable 5, a Mythos-class model, was taken offline for several weeks alongside Mythos 5 following government concerns before being redeployed globally on July 1, 2026 with strengthened cyber safeguards (Anthropic; The Verge). Opus 5 enters at a lower price point without the Mythos-class baggage, offering enterprises a near-frontier option at $5/$25 rather than $10/$50.

Safety and Fallback Changes

Anthropic's system card (a technical document describing the model's behavior, limitations, and safeguards) details a notable shift in safety posture. Opus 5's protections match those of Fable 5, with one exception: Opus 5 now permits source-code vulnerability discovery at all access levels (Anthropic System Card). Anthropic describes Opus 5 as "the most aligned Opus model and is the least susceptible to being tricked into misuse" (The Verge). The new fallback feature automatically routes requests to a lower-tier model when Opus 5's safeguards mandate a decline, reducing dead-end interactions for developers building agentic pipelines that cannot tolerate hard refusals mid-task.

The Fast mode, launched in research preview, offers higher inference speeds at double the standard token pricing. For latency-sensitive agent workflows where cost is secondary to throughput, this gives teams a toggle without committing to a separate model deployment.

Anthropic spokesperson Danielle Ghiglieri confirmed the company continues to work with government partners on independent testing of its models, including Opus 5 (The Verge). Reuters characterized the release as an "efficiency upgrade" in its July 24 coverage (Reuters.

A Dense Summer of Announcements

Opus 5 lands within a packed summer of Anthropic activity. Claude Sonnet 5 shipped June 30 at $2/$10 per million input/output tokens, positioned as Anthropic's most agentic Sonnet yet with native browser and terminal tool use (Anthropic). The same day, Anthropic joined Amazon, Microsoft, Google, and other Glasswing partners in proposing an industry-wide framework for scoring jailbreak severity — that is, rating how hard it is to bypass a model's safety guardrails (Anthropic). On July 9, former Federal Reserve Chair Ben Bernanke was appointed to Anthropic's Long-Term Benefit Trust, and the company launched its "Inviting hard questions" public initiative. A $10 million commitment to Canadian AI research and the Claude for Teachers product both followed on July 14. A $20 million donation to Public First Action was announced July 21 (Anthropic).

The cadence tells its own story. Anthropic is layering frontier model releases, governance structures, and public-facing trust initiatives in parallel rather than sequentially. Opus 5 is the product release; the Glasswing jailbreak framework, the Bernanke trust appointment, and the government testing partnerships form the scaffolding around it. Whether that scaffolding meaningfully differentiates Anthropic in a market where every major lab publishes safety commitments is a question enterprise buyers will answer with procurement decisions, not press releases.

The practical question for engineering teams is simpler. Opus 5 at $5/$25 with Fable-5-adjacent alignment, permitted vulnerability discovery, and a refusal-fallback path may be the more deployable model for production agent workloads, even if Fable 5 retains the ceiling on raw capability. For organizations that hit Fable 5's price ceiling or its access restrictions, Opus 5 closes most of the gap at half the cost.