Inside the Claude 5 Model Family and Fallback Routing Pipelines

Anthropic has rolled out its new Claude 5 family of AI models, designed to balance speed, cost, and capability across different use cases. While Claude Sonnet 5, Claude Opus 5, and Claude Fable 5 are open to the public, there is a fourth, restricted model in the mix: Claude Mythos 5. This one is kept strictly under wraps for high-security defense applications.

Let's take a look at how these models work and what sets each of them apart.

Understanding the Four Claude 5 Models

First up is Claude Sonnet 5. This model excels at autonomous execution, like running browser loops or entering terminal commands. Developers can currently access Sonnet 5 via API at a heavily discounted promotional rate. Through August 31, 2026, pricing sits at $2.00 per million input tokens and $10.00 per million output tokens, before stepping up to the standard rate of $3.00 input and $15.00 output. Sonnet 5 powers both the free and Pro tiers of Claude on the web, offering performance close to the older, pricier Opus 4.8 at a fraction of the cost.

Next is Claude Opus 5, the go-to choice for general development and office workflows. It serves as the default model for Claude Max and is the flagship option for Pro subscribers. Standard API pricing is $5.00 per million input tokens and $25.00 per million output tokens. For urgent workloads, a high-speed mode runs tasks 2.5 times faster for twice the price. Opus 5 shines in coding benchmarks, driving down task completion costs significantly. It also represents a major leap forward for scientific research, particularly in biology and chemistry workflows like molecular modeling and protein prediction.

Then we have Claude Fable 5, designed for complex, deep reasoning. Its release path has been eventful. After launching on June 9, 2026, a temporary regulatory pause on June 12 briefly halted its distribution, but it returned globally on July 1, 2026. As the most expensive public model, Fable 5 costs $10.00 per million input tokens and $50.00 per million output tokens. This pricing yields a massive 1-million-token context window and a 128,000-token output limit. It also supports dynamic compute allocation, letting developers adjust the model's reasoning effort on the fly. To put its power in perspective, Stripe used Fable 5 to migrate a 50-million-line legacy codebase in just 24 hours, a project that would typically take a human engineering team two months.

Finally, Claude Mythos 5 is a defense-grade variant of Fable 5 optimized for rapid scientific research. While it shares Fable 5's architecture, Anthropic completely stripped away safety filters related to cyber warfare and biological agents. Because of this, access is highly restricted, limited to defense agencies and military partners under Project Glasswing. In early trials, Mythos 5 designed proteins ten times faster than standard models, autonomously correcting its mistakes and identifying targets without human intervention.

Keeping Users Safe with Smart Routing

Anthropic relies on automated pre-screening to prevent misuse. Every prompt undergoes safety evaluations before reaching the primary model. Queries triggering high-severity safety flags are instantly redirected.

For example, cybersecurity-related prompts (such as network scanning) or chemical synthesis inquiries (like distillation) are routed to Claude Opus 4.8, which retains strict legacy safety guardrails.

Biology queries follow a different path. When a biology-related query is flagged on Fable 5, the system routes it to Opus 5 rather than the older Opus 4.8. This redirection ensures safe, legitimate biological research can continue under the model's specialized safeguards while still blocking harmful material.

This routing process introduces a slight delay of a few seconds, though it only impacts a tiny fraction of users. Less than 5% of conversations trigger the safety filter, meaning more than 95% of queries bypass the system with zero latency. Developers facing false positives can resolve them by providing clear, benign context at the start of their prompts.

Data retention policies also vary by tier. Unlike systems that store interactions for 30 days for safety audits, Claude Opus 5 has no data retention requirements for general access. Your inputs and outputs remain private and unstored during standard usage.