Meta

Executive Summary

What it is: Meta re-entered the coding-agent market in force: Muse Code (terminal multi-agent CLI, beta Aug 5, GA Aug 31 with subscription plans), the Muse Spark 1.2 flagship ("enhanced coding capabilities," computer use, multimodal) on the self-serve Meta Model API and OpenRouter, and Muse Glimmer, a 29.6B Apache 2.0 local-agent model. Source: https://developer.meta.com/ai/

What to watch out for: The self-reported 82.9% Terminal-Bench is unverified (independent common-harness rank: 14th), subscription prices were published only as an image, Spark 1.2's open-weight license is still pending, and enterprise controls are absent. Source: https://developer.meta.com/ai/resources/blog/muse-code-new-plans-and-features/

Bottom line: Meta's return to open weights plus a real terminal agent is the month's biggest strategic shift, and the Vals Index values it at $0.69/test (5th overall); treat it as a promising beta, not a production platform, until pricing is machine-readable and benchmarks verify. Source: https://local-ai-zone.github.io/blog/ai-updates-august-2026.html

Key Terms

  • Muse Code - Meta's terminal coding agent ("multi-agent coding in your terminal"), built for Muse Spark. Entered beta August 5; exited beta August 31 with inter-session messaging (agents pass messages over local Unix sockets), Workflow (orchestrated subagent teams), Rewind, an SDK (developer preview, Muse Session Protocol), and rolling subscription plans. Sources: Meta – Muse Code , Meta – Muse Code New Plans And Features
  • Muse Spark 1.2 - Meta's new flagship (August 5-10): "enhanced coding capabilities," end-to-end agentic workflows including computer use, native multimodal perception (audio, video, images, documents). Open weights announced by Mark Zuckerberg (release pending at month-end). Available via Muse Code, the Meta Model API, and OpenRouter. Sources: Meta – Muse Spark , Local-Ai-Zone – Ai Updates August 2026
  • Muse Glimmer - 29.6B dense multimodal open-weight model (27.8B LM + 1.8B frozen ViT), Apache 2.0, 128K native context, distilled from Muse Spark for "always-on local agents" on a single consumer GPU (August 10). Sources: Meta – Muse Glimmer , Local-Ai-Zone – Ai Updates August 2026
  • Meta Model API - self-serve OpenAI-SDK-compatible API for Muse models (dev.meta.ai), including Muse Image at $0.01/image. Source: Meta – Meta Model Api
  • Muse Session Protocol (MSP) - the open protocol behind the Muse Code SDK: local stdio communication between a program and the Muse engine, no network. Source: GitHub – Muse Code Sdk

Latest Changes

Changes since the 2026-07 report.

July watch items verified:

  • Muse Spark 1.1 evolution (US-only preview): superseded. Spark 1.2 launched with "expanded global access" in public preview across Muse Code and the Meta Model API. Source: Meta – Ai
  • Llama 4 open-weight line: confirmed abandoned in practice. The developer site relegates Llama 4 to a legacy listing; no successor appeared. Meta's open-weight energy moved to Muse Glimmer (shipped) and Spark 1.2 (announced). Source: Meta – Llama 4

New August changes:

  • [NEW ENTRANT to the coding-agent market] Muse Code beta (Aug 5) and GA (Aug 31): Meta entered the tracked coding-agent category in earnest: multi-agent terminal coding with persistent subagents and full auditability, 1M-token context sessions, inter-session messaging, Workflow orchestration, Rewind, and a TypeScript SDK. GA adds three subscription plans (prices published only in blog imagery; see Transparency Gaps). Sources: Meta – Muse Code New Plans And Features , Local-Ai-Zone – Ai Updates August 2026
  • New flagship: Muse Spark 1.2 with coding as the headline capability. Meta-reported Terminal-Bench 82.9% (not yet independently verified); independent testing ranks Muse Code 14th on common-harness Terminal-Bench but 5th overall on the Vals Index at $0.69 per test. Source: Local-Ai-Zone – Ai Updates August 2026
  • Open-weights return: Spark 1.2 open weights announced (modified Llama Community License); Muse Glimmer shipped Apache 2.0 on August 10. Source: Local-Ai-Zone – Ai Updates August 2026
  • Ecosystem: Muse Image on the Meta Model API at $0.01/image; cookbooks for multi-agent orchestration, subagent fan-out, computer use, a GitHub agent built with OpenCode, and search grounding. Source: Meta – Ai

Plans

Muse Code subscription plans launched August 31: three monthly tiers advertised as "more affordable Meta Model API usage than pay-as-you-go pricing." Exact prices were published as an image without machine-readable text (undisclosed here). Pay-as-you-go Meta Model API rates at dev.meta.ai remain the alternative. Source: Meta – Muse Code New Plans And Features

API Pricing

Meta Model API (OpenAI-SDK-compatible): per-token rates for Muse Spark 1.2 are listed at dev.meta.ai; Muse Image is $0.01/image. A full public rate table was not extractable this cycle; the Vals Index measured Muse Code economics at $0.69 per test. Sources: Dev , Local-Ai-Zone – Ai Updates August 2026

Model Performance / Benchmarks

Benchmark Score Notes
Terminal-Bench (Meta-reported) 82.9% Muse Spark 1.2 claim, not independently verified on official leaderboards. Source: Local-Ai-Zone – Ai Updates August 2026
Common-harness Terminal-Bench (independent) 14th place Independent testing. Source: Local-Ai-Zone – Ai Updates August 2026
Vals Index (independent) 5th overall at $0.69/test Best value positioning among frontier agents. Source: Local-Ai-Zone – Ai Updates August 2026

Latest News

Meta Returns to Coding Agents and Open Weights (August 2026)

After 15+ months without a Llama successor, Meta Superintelligence Labs shipped a full stack in one month: the Muse Spark 1.2 flagship with coding as its lead use case, the Muse Code terminal agent (beta Aug 5, GA Aug 31), Muse Glimmer as an Apache 2.0 local-agent model, and announced open weights for Spark 1.2. Zuckerberg framed open weights as competitive necessity. Sources: Meta – Ai , Local-Ai-Zone – Ai Updates August 2026

Muse Code GA Feature Set (August 31, 2026)

Inter-session messaging between parallel agent sessions (local Unix sockets, agent-driven), Workflow subagent orchestration with a live control room, double-Esc Rewind over the event log, and an SDK exposing sessions, tools, and permissions over the open Muse Session Protocol. Source: Meta – Muse Code New Plans And Features

Community Signals

Independent evaluation coverage highlighted the gap between Meta's self-reported 82.9% Terminal-Bench and its 14th-place common-harness ranking, while crediting the Vals Index value result ($0.69/test, 5th overall). Source: Local-Ai-Zone – Ai Updates August 2026 . Meta's agentic security posture (Muse Code + EU AI Act rules) was noted in third-party August comparisons. Source: Neuralcoretech – Ai Coding Agents August 2026 Muse Code

Enterprise Readiness

Feature Available? Details
SSO (SAML/OIDC) No Not documented for Muse Code / Meta Model API.
SCIM No Not documented.
Audit logs Partial Muse Code advertises "full auditability" of multi-agent runs; enterprise audit infrastructure not documented. Source: Local-Ai-Zone – Ai Updates August 2026
IP indemnity No Not documented.
Data residency No Not documented.
HIPAA No Not applicable to current offerings.
Air-gapped / on-prem Partial Muse Glimmer enables fully local agents (Apache 2.0, consumer GPU). Source: Meta – Muse Glimmer
SLA No Not published.
Admin controls (RBAC) No SDK exposes permission control per session; organizational admin controls not documented. Source: GitHub – Muse Code Sdk

Transparency Gaps

Gap Details Severity
Subscription pricing The three Muse Code plans are shown as an image; prices are not machine-readable or listed in text. High
Self-reported benchmarks The 82.9% Terminal-Bench claim is unverified and diverges from independent 14th-place ranking. High
Spark 1.2 open-weight license "Modified Llama Community License" terms are not published yet. Medium
API rate table Muse Spark 1.2 per-MTok pricing is not surfaced on the public site in extractable form. Medium
Muse availability regions "Expanded global access" is undefined; July's US-only restriction history is not updated with a region list. Low