Message threads are now in public beta on the Messages API. The API caches your conversation on the server between turns, so instead of resending the full conversation, you send only new messages and the previous message ID, and Claude starts responding sooner.

We recently integrated message threads in Claude Code. On requests with 500K to 1M tokens, median time to first token is about 35% lower. In long conversations, the median request went from about 800 KB to 1.8 KB.

To use it, add the thread parameter and beta header to your requests. It's available on the Claude Platform, Claude Platform on AWS, Claude in Microsoft Foundry, and Claude on Google Cloud at standard token rates. Full blog:

i wake up and new update

TL;DR 👇

here's a breakdown of this big improvement, put together by Opus 5.5!

That is what I was looking for. But you produce new features so fast that I cannot make it on testing 😅😅

No shots at all, love Claude. This week I noticed usage burned the entire thing. Is this from duet not allowing thinking effort to be selected or the follow up?

everyone spent the last two years writing custom sqlite wrappers to juggle context arrays just for Anthropic to finally hold the state server side.

Need weekly reset plsss

Where is the rust sdk?

Wow this is huge for my on device personal agent powered by Claude

not a day without a claude release

is it cheaper as well?

That means less token consumption right?

Is this similar to what codex does with websockets?

Chill everyone, Fable 5.5 will come tomorrow.

does the cached thread bill at the 10% cache-read rate or full price

@maudegregori something for our agents to look at

On my way to improve performance of customer support agent 🏃‍♂️🏃‍♂️

so is the cached history billed like a prompt cache read or as full input every turn? that's what decides if this is a speed feature or a cost feature

Sending only new messages plus the previous message ID is a clean fix for long sessions. The 35% lower median time to first token on 500K to 1M token requests shows how much of the wait was resending context. Practical gain for anyone building agents.

Five minutes at the coffee machine and the thread expires

if the server holds the history, reproducing a bad turn means pulling what it actually had for that message ID, not what your app thinks it sent

the server remembers so my wallet can forget

Nice update

sending only the new message instead of the whole history is a big deal for anything with long conversations. first-token time is the whole feel of a chat product. can’t wait to test this.

the gotcha in the docs: a thread doesnt cache your prompt for you you still want a top-level cache_control on every request, otherwise each turn sends less but nothing gets cached

not resending the whole conversation every turn is a nice quality of life thing for anyone building chat tools!

Did fable 5.5 compose this music?

hmm interesting ok I did wonder why we were sending the whole thing back every time, seemed like it was easy to screw up compared to some kind of session id token

i can't trim old tool output from the history anymore

Server side conversation caching can reduce repeated context transmission but developers also need clear guarantees around state isolation retention deletion and consistency across sessions

previous message id as the cursor

Why not just use OpenResponses format?