Anthropic launches Message Threads beta to cut Claude latency by 35 percent
Claude now remembers your long chats for five minutes. You get faster replies, but you still pay for every token.
Tools · Source: Unite.AI
What happened
Anthropic released a public beta for message threads on the Claude Platform on November 2, 2026. Developers building long-running agents no longer need to send the entire conversation history on every single API call. The new system divides the request lifecycle into two parts. First, you send a create request with the full conversation, system prompt, and tools. After that, you send continue requests containing only the new messages and the previous response ID.
This changes the math for heavy agent workflows where latency matters. Anthropic says requests carrying 500,000 to 1 million tokens see their median time to first token drop by about 35 percent. The server holds the thread in encrypted memory instead of forcing a massive payload upload every turn. It skips the redundant API validation and tokenization steps for the older context.
The thread stays alive for exactly five minutes after the last request. Every new response restarts that timer. If you wait longer, the thread expires. You then have to send the full conversation payload again in a new create request. The data never hits durable storage. Anthropic deletes it completely no more than one hour after its last use.
Key facts
- 35 percent — Median drop in time to first token for 500k to 1M token requests
- 5 minutes — Idle time before a message thread expires on the server
- November 2, 2026 — Date Anthropic launched the message threads public beta
Why it matters
Building long-running coding agents or computer-use agents just got much more responsive. Before this update, uploading massive image-heavy or document-heavy contexts repeatedly created severe network drag. Now, the API skips the redundant upload step. Your agent feels significantly faster to the end user because the server already has the context loaded in memory. The official SDK even includes a thread helper to manage this state automatically.
But you must understand this is purely a network shortcut, not a pricing discount. Claude still reads the entire conversation history on every single turn to generate its response. Your token usage stays exactly the same. Standard token rates and existing rate limits apply to every call. You also lose out on zero data retention and HIPAA compliance if you use this beta feature. Developers must weigh the speed upgrade against these strict privacy limitations.
For builders
Faster agents for heavy workflows
End users win with 35 percent faster response times on massive contexts. Builders save on network overhead and upload times. However, engineering teams must build logic to handle the strict five-minute thread expiration.
You still pay for every token
Anthropic still charges you for the full context window on every turn because the model reads everything. Founders hoping for a massive cost reduction on long agent sessions will be disappointed. You pay the exact same bill.
Maintain your own conversation history
Threads expire quickly and cannot be retrieved, listed, or deleted via the API. Your application remains the single source of truth. You must keep a local copy of every file and message to replay the full context when a thread drops.
My take
This is a brilliant engineering trick to make Claude feel faster without Anthropic giving up a single cent of token revenue. I love the speed boost for heavy coding agents, but the five-minute timeout is brutal for real-world user sessions. You still have to build robust fallback logic to resend the whole payload anyway, which adds complexity to your backend.
Original reporting: Unite.AI. This is my rewrite and opinion.