Message threads are now in public beta on the Messages API. The server caches conversation history between turns, so each new request only sends new messages and the previous message ID rather than the full conversation history.
Integrating message threads into Claude Code reduced median time to first token by about 35% on requests with 500K to 1M tokens. Payload size in long conversations dropped from roughly 800 KB per request to just 1.8 KB.
The new message threads feature, already integrated into Claude Code, is available on the Claude Platform, AWS, Microsoft Foundry, and Google Cloud at standard token rates. Add the thread parameter and beta header to requests to enable it.
A practical list of prompts Sumner uses when pushing Claude on performance optimization work: data-oriented design rewrites, replacing hash tables with flat arrays and pointer indices, packing data into upper bits, structure-of-arrays over array-of-structures, and explicitly deleting caches rather than adding them.
When doing performance work with Claude, Sumner sometimes resets expectations by explicitly telling it that 1% faster is not enough and the target should be 10x faster, which pushes the model toward more ambitious rewrites.
More specific prompts Sumner uses when pushing Claude on performance: replacing heap-allocated lists with stack-allocated ones with inline capacity, and swapping arrays of bools for bitsets.
Sumner describes staying on top of Claude's output during performance optimization by sampling code as it's being written, either directly or via GitHub Desktop, comparing it to repeatedly pausing and resuming a debugger to see where execution lands.
After telling Claude that 1% faster isn't enough and demanding 10x improvement, Sumner reports Claude came back with "only" 8x faster. The scare quotes do a lot of work here.
After extended iteration on performance work, agents tend to pursue 0.1% gains instead of hunting for 10x improvements. The problem isn't the profiling approach but the diminishing ambition over long sessions.
Accepted open-source maintainers in the Claude for OSS program now receive Ultra plan access instead of the previous tier.
The 5-hour usage limit that had been applied to the $200 Ultra plan was removed.
After one week on the new $200 plans, Tom Quinn's takeaway is that unlimited tokens are not necessary for high-quality output. He also called this the most expensive intelligence will ever be.
Tom Quinn says he has used the $200 Claude plans on weekends for a while and has never hit a meaningful usage limit, pushing back on claims that the plans are too restrictive.
In response to concerns about future price hikes for Opus, engineer Tom Quinn argued that intelligence will become abundant over time, with too many market forces pushing costs down.
Tom Quinn clarified that his comment about not hitting real limits on the $200 plans accounts for the 5-hour usage caps, not ignoring them.
During compaction in Claude Code, the entire conversation including skills and file reads is summarized, but immediately after, invoked skills are reattached and up to five of the most recently modified files are re-read. The full behavior is documented in the official docs.
Jarred Sumner noted that Claude's LGTM on a PR was its way of signing off, but users can trigger a full review by commenting "@claude review" on any pull request.
