An earlier single-session attempt to have Claude work on the Riemann hypothesis went nowhere. The successful run started eight days before the announcement, with a different opening message that apparently set a more productive direction.
The small token overhead from auto mode's per-tool-call classifier no longer counts against usage limits on Pro, Max, and Team plans. Enterprise and API support is coming later.
While jogging, a user asked Claude to solve the Riemann Hypothesis. It didn't, but over 1.5 days it produced a proof that at least 67% of the nontrivial zeros lie on the critical line, up from the prior record of 41.6%. Analytic number theorists are reportedly excited.
A UI bug caused by a config caching issue made it appear Fable's plan entitlements had changed. The underlying billing was unaffected: any Fable usage on Max or Teams plans correctly drew from subscriptions. A config update was pushed; restarting Claude or waiting a few hours clears it.
An Anthropic team member highlighted that the Riemann zeros breakthrough came down to simply prompting Claude to keep going, quoting the original post about Claude's progress on the problem.
The person behind the Claude-assisted Riemann zeros result says they dropped out at 16 and contributed no math, just variations of "keep going" and "believe in yourself." The underlying mathematical heavy lifting came from recent human work that Claude synthesized.
Pasting the theorem from the Claude-assisted Riemann zeros proof into current models yields very different results: Sonnet 5 refuses to engage but can't explain why, Haiku 4.5 admits it doesn't know, and Opus 5 times out after thinking for 17 minutes.
After a user posted a screenshot showing an unexpected model appearing in Claude, bcherny said he was unaware of any intentional changes and began investigating.
Bcherny confirmed work is underway to make Fable's allowed compute the full compute budget for Max users, and that the team is about halfway there.
Users on Teams Standard seats don't get Fable access; it requires a Teams Premium seat. To reduce heavy token usage, disabling plugins with /usage and lowering effort with /effort are the recommended starting points.
After a user reported Claude Code becoming resource-heavy and sluggish, bcherny noted the team has shipped significant P99 RSS improvements over recent months and asked for specifics: CLI vs. Desktop, operating system, and which tasks or tools trigger the problem.
A user on a Ryzen 7 7800X3D with 32GB RAM reports Claude Desktop crashing nearly every session due to memory exhaustion, driven by spawning multiple PowerShell windows and background tasks. Bcherny is attempting to reproduce the issue with the shared specs.
A user reported burning through their 20x plan in a day or two and experiencing crashes. The Anthropic team member expressed surprise, asked about loop usage, and offered a diagnostic command (claude -p /usage | pbcopy) to investigate, while asking for crash reproduction steps.
When a message is flagged, Claude Code falls back to a different model by default. You can disable this in /config under 'Switch models when a message is flagged' so the conversation pauses instead of switching.
When a user was unsure whether a model switch was the right behavior, lydiahallie explained they could run /feedback and share the feedback ID via DM to flag potentially incorrect message flagging.
