<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>RADAR — pin_or_freeze</title>
    <link>https://radar.trulysovereign.ai</link>
    <atom:link href="https://radar.trulysovereign.ai/feeds/action/pin-or-freeze.xml" rel="self" type="application/rss+xml"/>
    <description>Items whose recommended action is pin_or_freeze.</description>
    <language>en</language>
    <lastBuildDate>Thu, 24 Sep 2026 04:30:30 GMT</lastBuildDate>
    <item>
      <title>Claude API default top_p changed from 0.999 to 0.99 for all models</title>
      <link>https://radar.trulysovereign.ai/items/e503476f-0504-4f5a-861a-aec1994696ea</link>
      <guid isPermaLink="false">e503476f-0504-4f5a-861a-aec1994696ea</guid>
      <pubDate>Sat, 22 Aug 2026 01:15:54 GMT</pubDate>
      <category>pin_or_freeze</category>
      <source url="https://platform.claude.com/docs/en/release-notes/overview#may-22-2025">Claude Platform release notes</source>
      <description>The Messages API now defaults top_p to 0.99 instead of 0.999 across all models, effective immediately. Extended thinking calls can now set top_p between 0.95 and 1.0, expanded from the previous narrower range.

Why it matters: Every request not explicitly setting top_p now uses different sampling behavior, which changes response variability and may affect tool call reliability, output consistency in production workflows, and any testing or validation tied to specific model behavior.

What to do: Audit your Claude API calls: if you rely on the previous default behavior, explicitly set top_p: 0.999 in all requests. If you use extended thinking and want the old constraint, pin top_p below 0.95 or above 1.0 will now error—test your parameter validation.</description>
    </item>
  </channel>
</rss>
