Unverified50% confidenceEventExact time
Bug 3 was confirmed on Haiku 3.5 and rolled back on September 4; Anthropic believes it may have affected some Sonnet 4 and Opus 3 requests on the Claude API, and Opus 3 was rolled back on September 12; the final fix was to switch to exact top-k and normalize some operations to fp32, accepting a slight efficiency loss.
1
Sources
50%
Confidence
Long-term
Relevance
10/19/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedBug two was that a configuration error in Claude API's TPU servers caused incorrect token generation, occasionally assigning high probability to tokens that should rarely appear, resulting in English prompts getting Thai or Chinese replies and code syntax errors; this problem affected Opus 4.1 and Opus 4 from August 25 to 28 and Sonnet 4 until a rollback on September 2, and third-party platforms were not affected.73% similarUnverifiedBug three originated from code deployed in August 2025 that improved the token selection logic (removing a temporary workaround from December 2024), which inadvertently triggered a latent bug in the XLA:TPU compiler, causing the approximate top-k algorithm to return completely wrong results under specific batch sizes and model configurations, and the same prompt might work one time and fail the next.66% similarUnverifiedAnthropic's Opus and Sonnet 5.5 have removed the no-reasoning option, while Haiku 5.5 retains it but performs poorly in that mode.65% similarUnverifiedIn tests with real Agent logs, Haiku 4.5 and Opus 5.5 showed leakage (P=0.050), while Haiku 5.5 and Sonnet 5.5 held.63% similarUnverified@ai-sdk/zai 3.0.6 follows semantic versioning; the major version number 3 indicates it evolves in sync with the AI SDK 5.x mainline, and the patch digit indicates a backward-compatible bug-fix update.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1072493API
curl https://kongchang.com/api/v1/knowledge/claims/1072493MCP
get_claim(id=1072493)