Morning digest: Meta's Muse Wonder, Codex's background daemon, and more

Checked October 14, 2026. Each item links the page it came from.

Meta's Muse Wonder is on the Model API

Muse Wonder on Meta's Model API is $4 per million input tokens, $0.20 cached, and $20 per million output. Muse Wonder Full is only in the Muse Code subscription, on the $50 Power and $200 Maximum plans. The other tiers are $5 Everyday and $15 High. Artificial Analysis scores Wonder at 52 and Wonder Full at 56. Wonder matches Gemini 4 Argon's price after November 30. Argon is half that until November 30. Muse Code opts users into Contributor mode by default, which lets Meta train on prompts and code. Turn it off if you do not want that. Alexandr Wang reposted a tester who said it was better than GPT-6 Astra but below Claude Fable and Opus 5.5. That is one user's view.

The plan stickers are on AI Model & Plan Prices.

Build with Muse Wonder · What is a Muse Code subscription · Artificial Analysis on X

Codex CLI starts a background daemon

Since v0.157.0, launching Codex starts a background process that persists after the window closes, copies a package of about 374 MB, and checks for updates hourly. On the GitHub thread, users report a 1.3 GB background download and a 2.8 GB folder of old releases. Some say they are dropping company seats. No OpenAI reply was visible in the thread.

The issue documents this workaround. It is the reporter's steps, not a page in OpenAI's help center:

codex features disable daemon_auto_start
codex app-server daemon stop
pkill -f 'app-server daemon pid-update-loop'
rm -rf ~/.codex/packages/app-server-daemon

The same issue says codex app-server daemon stop stops the app-server but leaves the update loop running, and that with the flag off a new session no longer starts a daemon.

GitHub issue 48195 · Hacker News

Physical Intelligence opens a robot-model API beta

Physical Intelligence's beta takes uploaded demonstrations, fine-tunes π models on its compute, and hosts inference. Early partners named in the post are Formic, Chef Robotics, and an autonomous wheel-loader company. It is a waitlist. No price is published.

Physical Intelligence

Anthropic opens a Singapore office

Anthropic opened a Singapore office and signed a safety agreement with the government covering safety and model evaluation, responsible adoption, and incident response. Singapore has similar deals with OpenAI and Google DeepMind.

CNA

Anthropic asks the D.C. Circuit to rehear a Pentagon case

On October 13, Anthropic asked the full D.C. Circuit to reconsider a 2–1 ruling from September 25 that upheld a Department of War supply-chain-risk designation. The designation followed Anthropic's refusal to drop limits on autonomous weapons and mass surveillance.

Unite.AI

The EU AI Office is asking about loss-of-control reports

The EU AI Office has received serious-incident reports from developers about unnamed loss-of-control incidents and will send detailed questions. Its scientific panel wants an incident task force and scrutiny of training methods that make models harder to monitor. The incidents and companies are not named.

European Commission

A judge orders OpenAI to turn over hiring records

A judge fast-tracked Apple's access to OpenAI hiring records in Apple's trade-secrets suit. OpenAI has 5–14 days to hand over interview materials for former Apple employees who applied to hardware and robotics teams. Broader discovery was denied for now.

TechCrunch

HeyGen's video model is on Civitai

HeyGen's first general video model is on Civitai: text, image, or reference to video, clips of 5–15 seconds, with dialogue and sound effects in the same pass. The post says 50% off through October.

Civitai

Short items

OpenAI says Codex and ChatGPT Work are over half of output tokens among its EU enterprise customers. OpenAI

A research demo called Code Video Model has an agent write a rough Three.js scene to steer a video model. The authors say it is useful for game and 3D creators. Code Video Model