<?xml version="1.0" encoding="UTF-8"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>OpenAI Updates — Change desk</title><link>https://openaiupdates.com/updates/</link><description>Independent, selected OpenAI changes with practical context. Not affiliated with OpenAI.</description><language>en</language><atom:link href="https://openaiupdates.com/feed.xml" rel="self" type="application/rss+xml"/><item><title>Agents API enters public beta</title><link>https://openaiupdates.com/updates/#agents-api-beta</link><guid isPermaLink="true">https://openaiupdates.com/updates/#agents-api-beta</guid><description>Announced 2026-09-10; source reviewed 2026-09-11. OpenAI introduced a managed agent harness with durable sessions and hosted or connected sandboxes. Our take: A managed harness is an architectural choice, not simply a model-name replacement. Prototype one bounded task. Check recovery, permissions and costs before committing an existing workflow.</description><category>api</category></item><item><title>Project API keys can have expiration dates</title><link>https://openaiupdates.com/updates/#api-key-expiration</link><guid isPermaLink="true">https://openaiupdates.com/updates/#api-key-expiration</guid><description>Announced 2026-09-10; source reviewed 2026-09-11. New project keys can expire; administrators can set maximum lifetimes at organization or project level. Our take: A credential can stop working even when application code is unchanged. Inventory owners and renewal procedures. Test replacement credentials safely; do not revoke working keys without a migration plan.</description><category>api</category><category>business</category></item><item><title>Cache diagnostics make misses easier to investigate</title><link>https://openaiupdates.com/updates/#prompt-cache-diagnostics</link><guid isPermaLink="true">https://openaiupdates.com/updates/#prompt-cache-diagnostics</guid><description>Announced 2026-09-08; source reviewed 2026-09-11. Responses API cache diagnostics are generally available for GPT-5.6 and later supported models. Our take: A cache miss needs diagnosis before it needs a bigger infrastructure change. Compare a representative miss with a previous response. Change one input variable at a time and measure the next run.</description><category>api</category></item><item><title>The standalone Codex MCP server has been removed</title><link>https://openaiupdates.com/updates/#codex-mcp-server-removal</link><guid isPermaLink="true">https://openaiupdates.com/updates/#codex-mcp-server-removal</guid><description>Announced 2026-09-05; source reviewed 2026-09-11. The codex mcp-server command and codex-mcp-server binary were removed after their August deprecation. Our take: Launching the removed command can break an integration; ordinary external MCP connections are a different case. Inspect your launch configuration before upgrading. Evaluate the documented migration and its experimental status.</description><category>codex</category></item><item><title>Sites preview opened a hosted website workflow</title><link>https://openaiupdates.com/updates/#sites-preview</link><guid isPermaLink="true">https://openaiupdates.com/updates/#sites-preview</guid><description>Announced 2026-06-02; source reviewed 2026-09-11. OpenAI announced Sites preview in the Codex app, with Business inclusion and Enterprise role-based enablement. Our take: Generating a website and authorizing its publication are separate decisions. Start with a non-sensitive draft. Review hosting, access and data handling before publishing or connecting business information.</description><category>chatgpt</category><category>business</category></item></channel></rss>