AI News Briefing — Meta releases Muse Wonder, its first frontier model

Model releases

[2026-10-12] Meta releases Muse Wonder, its first frontier model, on the Meta Model API as muse-wonder-1.0: $4 per million input tokens, $0.20 cached and $20 output, with a context of about 1M tokens. The stronger Wonder Full is not sold through the API; it comes only with Muse Code’s $50 and $200 monthly plans. Artificial Analysis puts Wonder Full at 56 and Wonder at 52 on its index, against 58 for Opus 5.5. Neither model has downloadable weights. (official, source, source)

The API model is the weaker of the two, so benchmark tables quoting Wonder Full describe something an API customer cannot call.

For Engineering Managers / Tech Leads: A team weighing Muse Wonder is pricing muse-wonder-1.0 at $4/$20 per million tokens, not Wonder Full, which comes only with the $50 and $200 Muse Code plans. Approve the trial against the API model’s score of 52, and keep any Muse Code plans as a separate per-developer line.
[2026-10-12] Anthropic says its distillation defenses cut what extractors harvest by roughly 90–99%. Two parts reach ordinary API users: API accounts created on or after August 31 cannot edit Claude’s preserved thinking, and since September 24 requests blocked by the extraction classifier are billed. Anthropic puts that classifier’s false-positive rate below 0.1%. (official)

A blocked request now shows up on the invoice, so an unexplained run of refusals in a high-volume batch job is worth checking against the bill.

[2026-10-12] OpenCode 2.0 is now the main release of the open-source coding agent. A background service starts automatically and serves the terminal, desktop and web apps, so a session can move between them, and scripts can attach to it. Code Mode becomes the default for every MCP server: the agent writes JavaScript that calls several tools in one run instead of loading each tool definition into context. (source)

The embeddable @opencode/sdk is still under development, so building on it means tracking a moving target.

For Software Developers: Run an existing MCP setup through OpenCode 2.0 and debug it differently: a multi-step tool sequence now happens inside one JavaScript run, so read the script the agent wrote when a step fails, not just the individual tool calls.

[2026-10-12] Spotify’s merged code changes per month rose from about 8,100 to 17,000 year over year, driven by its Fleet Management automation and the Honk background agent. An incident review found no material direct contribution from AI-written code, but change volume had outpaced some verification controls, so Spotify is widening rollback capacity and scheduling automated changes for the owning team’s working hours. (source)

Spotify’s fixes target recovery, not authorship. Faster rollback and changes landing while their owners are online help whether a person or an agent wrote the diff.
[2026-10-12] Octopus surveyed 379 platform professionals: teams that build most platform features from scratch are 3.5 times more likely to say their platform lacks a clear strategy. Only 17.5% work mainly that way; 51.3% mostly buy tools. Agents make a scratch build cheaper to start, not cheaper to maintain. (source)

[2026-10-12] Enterprise admins can now attach a user to GitHub Copilot’s OpenTelemetry data through a managed telemetry setting. Identity capture adds the GitHub login, local OS account and host name, while prompts, responses, tool arguments, tool output and policy detail are five separate switches, all off by default. It covers Copilot CLI, the Copilot app and VS Code Agent Host sessions. (official)

Per-user cost reports become possible without collecting prompts, which can carry source code and file paths.

For Platform / DevOps Engineers: Enable identity capture in the managed telemetry setting and leave the five content switches off. Spans from Copilot CLI, the Copilot app and VS Code Agent Host then carry the GitHub login and host name, which is enough for the collector to group usage per person.

[2026-10-12] Anthropic’s Frontier Red Team tested whether small open models can copy themselves onto cheap devices. On an eight-node emulated Android fleet with an open ADB port standing in for the exploit, GLM-5.2 reached all eight nodes in four of five trials; models of 30B parameters or fewer handled single steps but broke when chaining them. Separately, Gemma4-26B-A4B ran across 15 Android TV boxes at about 0.39 tokens a second. (official)

The exploit step was deliberately trivial, so treat the results as an upper bound. The defensive advice is old: no default passwords and no exposed debug ports.

GPT-5.5 in ChatGPT and Codex: retires October 14; the API keeps it.
Step 5 Preview’s weights: not on Hugging Face yet; due October 15.
Mistral Large 4 weights: still preview only; promised by the end of October.