Anthropic launches message threads to speed up long AI agent sessions

Solving the history bottleneck

Anthropic recently launched a public beta feature called message threads for developers who build AI agents using Claude. This update addresses a common technical hurdle where agents must resend an entire conversation history every time they make a new request. For complex tasks, these histories can grow to include hundreds of thousands of tokens, which are the basic units of text that AI models process.

When a conversation becomes very long, sending the entire history repeatedly creates a significant delay before the model can start generating a response. Developers building coding agents, for example, often have sessions that involve hundreds of individual requests. Each of those requests normally carries every previous message and tool result, leading to slower performance as the session progresses.

By introducing threads, Anthropic allows an agent to send the full conversation history just once. After that initial step, the agent only needs to transmit the new messages along with the ID of the model's last reply. This change streamlines the communication between the developer's software and the underlying AI model without requiring the model to 'forget' any previous context.

Performance gains for heavy tasks

The primary benefit of this new system is a noticeable increase in speed for high-volume interactions. Anthropic reported that for requests ranging from 500,000 to 1 million tokens, the typical wait time before Claude begins answering fell by about 35%. This is particularly relevant for intensive workflows where the AI must maintain a deep understanding of a massive amount of information over a long period.

Anthropic has already implemented this feature for most users of Claude Code, which is its specialized coding agent. This means many developers are already seeing these performance improvements in their daily workflows. The goal of the update is specifically to reduce the latency—the time spent waiting for the AI to process the input—rather than to change the underlying logic of the model.

While the speed increase is significant, it is important to note that the model still reads the entire conversation on every turn. The thread system simply optimizes how that information is delivered to the system. Because the model still processes the full context, the pricing and rate limits for the API remain exactly the same as they were before the beta launch.

Technical details and data handling

From a technical standpoint, message threads are designed to be temporary and secure. The data is held in memory and remains encrypted throughout the process. Anthropic stated that these threads are never written to a physical disk, which provides a layer of security for the data being processed during a session.

However, because the threads are temporary, they are not permanent storage solutions for developers. The threads expire automatically five minutes after the last request is made. This means that developers are responsible for maintaining their own persistent copy of each conversation to ensure that the history is not lost once the thread expires.

There are also specific limitations regarding data privacy for this feature. Anthropic noted that message threads are not eligible for their zero data retention or HIPAA health-data setups. This means that while the threads are encrypted in memory, they do not fall under the specific privacy protections required for certain highly regulated industries or sensitive health information.

Availability and broader context

The message threads beta is available across a wide range of platforms for developers. It can be accessed through Anthropic's own platform as well as through major cloud providers, including AWS, Microsoft Foundry, and Google Cloud. This broad availability ensures that teams using different cloud infrastructures can adopt the faster response times quickly.

This update follows the recent release of the Opus 5.5 model, which was introduced in September as a more affordable option. While the new threads focus on speed, Anthropic has also been looking at ways to manage costs for long agent sessions. In September, the company cut cache prices by 75% to help manage the rising bills associated with long-running AI interactions.

By combining faster response times through threads with lower cache prices, Anthropic is attempting to make long-form AI work more viable for businesses. Developers can now build more complex agents that handle larger amounts of data with less friction. The beta serves as a testing ground to see how these optimizations perform in real-world production environments before a wider rollout.