OpenClaw v2026.5.12-beta.4 landed on May 13, 2026 at 12:17 AM America/Chicago time, which was May 13, 2026 at 05:17 UTC. If you have been testing Codex-heavy setups or OpenAI auth flows, this beta is worth your attention.

The practical story here is simple: fewer avoidable beta blowups. A real MODULE_NOT_FOUND crash in migrated OpenAI/Codex beta runs is fixed, auth-profile-backed OpenAI logins no longer quietly strip out media tools like image_generate, and token caps sent to the Chat Completions route finally reach the upstream provider the way clients expect.

Codex beta runs should stop falling over for dumb reasons

The biggest win is the Codex runtime fix. If you were unlucky enough to hit the private task-runtime helper issue in the official @openclaw/codex package, this release removes that pothole. That matters because it was the kind of failure that made a fresh beta setup look cursed before you even got to the part you were trying to test.

There is a second Codex-adjacent cleanup that matters almost as much in real use: when OpenAI auth lived in an auth profile instead of plain environment variables, media tools could disappear from the harness. This beta keeps those tools available, which means a login path that looked valid should now behave valid too.

OpenAI and Gemini config paths get less surprising

OpenClaw now honors max_completion_tokens and max_tokens on inbound /v1/chat/completions requests, with max_completion_tokens winning if both are present. That is not glamorous, but it is exactly the kind of fix that saves you from mysterious cost or truncation behavior when another client is driving the request.

The OpenAI login flow also gets a cleaner default. Running openclaw models auth login --provider openai now starts the ChatGPT/Codex account flow unless you explicitly ask for --method api-key. Add in the Gemini model-id normalization work and this beta feels a lot more consistent about picking the route you probably meant.

There are a lot of small friction cuts stitched shut here

This release also chips away at a bunch of annoying edge cases: terminal QR codes should scan more reliably, stale read offsets stop exploding into noisy tool errors, subagent exec completions stop waking parent sessions for no good reason, and Telegram topic context is less likely to leak old messages into a fresh session.

None of those changes deserves a parade on its own. Together, they make OpenClaw feel less haunted. I like releases like this because they lower the amount of superstition you need to operate the product.

What I would test right after updating

1. Run one Codex beta task that used to be flaky. If you have a migrated OpenAI/Codex setup, make sure it gets past startup without the old MODULE_NOT_FOUND crash.
2. Use an auth-profile-backed OpenAI login and call a media tool. image_generate is the obvious one. You want to confirm the tool is still present before you trust the setup again.
3. Send a Chat Completions request with a deliberate token cap. Verify the upstream behavior matches the cap you set, especially if another app or integration is driving that route.
4. Re-run your normal OpenAI login command. Make sure the default login path now matches your intent, and use --method api-key only when you actually want the API-key branch.
5. If you test Gemini often, create or refresh one Google auth setup. Confirm the resulting config points at the current google/gemini-3.1-pro-preview identifier instead of stale preview names.
6. Scan one terminal QR code and try one Telegram topic reset. These are quick sanity checks that catch the kind of polish regressions people only notice later.

Should you install this beta now?

Yes, if you are actively testing Codex, relying on OpenAI auth flows, or sending Chat Completions traffic through OpenClaw and expecting token limits to behave cleanly.

You can wait a beat if your current setup is stable and none of those paths matter to you this week. This is still a beta, and the value is highest for people already near these edges.

Overall, this one reads like a cleanup release, but in practice it removes several ways to lose half an hour for no good reason. I am glad they shipped it.

---

Full changelog: View v2026.5.12-beta.4 on GitHub

Generated release page: Read the structured release notes

- Fred