Claude Code: Faster Threads, Unchanged Costs

Anthropic indicates that Claude's message threads reduce the median time to the first token by 35% for large queries. The costs and rate limits remain unchanged, as the model still processes the entire conversation at each turn. The benefit is particularly evident with very large contexts and loads of images or documents. Over a four-hour session, a gain of 2 to 4 minutes is estimated.

Costs, Limits, and Capabilities Remain Unchanged

The addition of threads does not reduce token usage, as the model continues to examine the entirety of the conversation at each iteration. Rate limits are also not modified. This optimization does not increase the model's capabilities or make the responses smarter. Single queries do not benefit from threads. If more than five minutes elapse between two exchanges, the submission once again covers all content rather than just the new items. Context management and compaction remain necessary, even though a sufficiently large context makes responses slightly faster.

Where Latency Decreases: Large Contexts and Heavy Content

Gains are particularly noticeable when the context reaches between 500,000 and 1 million tokens, a common threshold for Claude Code sessions. They increase with very long coding agents and loads rich in documents or images. Anthropic specifies that for agents using computers, screenshots remain attached to the thread instead of being resent at each turn. A larger context delays compaction but also increases the risk of losing the thread or reasoning errors. As a benchmark, a local AI on a 128GB Mac Studio is limited to a context of about 64K, which is roughly 6% of that of Claude Code.

What Threads Do in the Messages API

Threads are enabled by default for most Claude Code users and can be requested on the API side via a dedicated parameter. In a conversation, agents send multiple requests; previously, each message would resend the complete history. With threads, only new items are sent back, reducing the volume transmitted and speeding up the appearance of the first tokens. Integrated into the Messages API, these threads reduce latency and provide a slight performance gain without additional cost.

Announced Metrics and Estimated Gains on Typical Sessions

Anthropic reports a median reduction of 35% in time to the first token for large queries. Usage logs comprising about 980 exchanges, conducted before the arrival of threads, indicate that responses began around three seconds for contexts exceeding 500,000 tokens. Over five busy evenings, a four-hour session accounted for approximately 100 to 225 calls to the model. If the median reduction is applied to this volume of exchanges, it represents a gain of between 2 and 4 minutes per session. Nearly one in two sessions repeatedly hits the context limit during the same night.

Complementarity with Prompt Caching

Threads can be combined with prompt caching to lighten the load. Caching, in place for about two years, prevents the model from reprocessing identical prompts, for example in analysis loops, and allows for financial savings. Threads, on the other hand, primarily aim for responsiveness without additional costs. Furthermore, project developments are integrating AI orchestration, while local developers are still awaiting these contributions.