When AI writes 80 per cent of code, who reviews it?
Anthropic reports that AI wrote more than 80 per cent of code adopted into its production codebase. What that means for reviews and approvals.
More than 80 per cent of the lines of code adopted into production at Anthropic in May 2026 were attributable to Claude, according to the company. The figure describes Anthropic’s own development environment and is not a benchmark for other teams. In June, the company released Claude Fable 5, a model based on Mythos 5. Fable 5.1 and Mythos 5.1 have been available since September.
Key takeaways
- Anthropic’s code figure is an internal measurement. It covers lines of code attributed to Claude that were adopted into production in May 2026.
- Fable and Mythos have different safeguards. Fable 5 is generally available, while access to Mythos 5 was restricted.
- Teams need clear rules for adopting AI-generated code. Responsibilities, tests and approvals belong in the development process.
What the 80 per cent actually measures
The figure comes from an analysis by Anthropic. The company counts lines of code attributed to Claude and adopted into its own production codebase. According to Anthropic, the attribution has gaps. Lines not attributed to Claude also include automatically generated code and other artefacts.
The percentage does not mean that Claude decides on approval itself, nor does it demonstrate an equivalent productivity gain. Anthropic says developers direct and review the work. For other companies, the figure’s relevance and comparability depend on factors including tasks, tools and measurement methods.
Fable 5 and Mythos 5 in use
Anthropic introduced Fable 5 and Mythos 5 on 9 June 2026. Fable 5 uses Mythos 5 as its underlying model and adds safeguards for general use. At launch, Mythos 5 was limited to a restricted group, including partners in the Glasswing security project.
Anthropic’s transparency overview has also listed Fable 5.1 and Mythos 5.1 since September 2026. Fable 5.1 is generally available. According to the provider, Mythos 5.1 is accessible through trusted access programmes and Claude Security.
Review safeguards and routing
For Fable 5, classifiers check requests relating to cybersecurity, biology and chemistry, among other subjects, as well as attempts to copy model capabilities by collecting responses. According to Anthropic, recognised requests in its web, desktop and mobile applications are automatically routed to the latest Claude Opus model. Users are notified which model handled the request.
For an integration, teams should therefore check the information for the specific interface and model version in use. When testing security-related tasks in particular, they should examine how safeguards affect permitted requests and how the system handles a refusal or rerouting.
Build review and costs into the process
A high proportion of generated code changes the requirements for acceptance. Teams should define who is responsible for the substance of changes, which tests must pass before a merge and when an additional security review is needed. Automated checks can help. The approval process must establish whether a change fits the system and its security requirements.
For operations, it is also advisable to attribute model usage and costs to teams or use cases. This allows the effort involved in generation, testing and rework to be considered together. A pilot with clearly defined tasks provides a more reliable basis for decisions than applying another organisation’s code percentage.
Frequently Asked Questions
What does Anthropic’s figure of more than 80 per cent mean?
It is the proportion of lines adopted into Anthropic’s production codebase that were attributed to Claude in May 2026. It is not a measurement for other companies.
Who reviews AI-generated code?
According to Anthropic, developers direct and review the work. Other teams should define who is responsible and establish approval rules for their own changes.
Is every Fable request routed to Opus?
No. Anthropic describes routing for recognised topics in its web, desktop and mobile applications. Users are notified which model handled the request.
Editor's Picks
cloudmagazinWhen the AI bill blows the cloud budgetcloudmagazinKubernetes as the default OS for AI: clusters as a compliance issuecloudmagazinCursor Cloud Agents: coding even when the laptop is closedMore from the MBF Media Network
MyBusinessFuture54.5 per cent use AI, yet the Mittelstand still lags behindDigital ChiefsAI has broken hiring: what matters nowSecurityTodayAPI security: the blind spot behind every integrationImage source: AI-generated (July 2026)
Translated from the German original using artificial intelligence. The German version is authoritative.

