Benchmark Meta Muse Spark 1.1 before adding its API to coding agents

Meta Muse Spark 1.1 is confirmed: Meta has released the agentic coding model and opened the Meta Model API in public preview for US developers. Confidence level: confirmed, with access limits. The practical question is whether its low-priced, long-context agent workflow is strong enough for your own coding, computer-use, and multimodal tasks.

What changed

Meta introduced Muse Spark 1.1 on July 9, 2026 as the updated model behind Meta AI's Thinking mode and as the first Muse Spark model available through the Meta Model API public preview. Meta positions it for coding, agentic workflows, computer use, tool calling, function calling, multimodal inputs, and long sessions.

The release matters because it moves Muse Spark from a mostly Meta-controlled product surface into a developer API. The Verge reports that the API preview is available to US developers and that new Meta Model API accounts receive $20 in credits. TestingCatalog also reports a 1-million-token context window, MCP/custom-skill support, parallel subagents, and context compaction claims from Meta's launch material.

Why this is early

TestingCatalog surfaced the developer-facing details quickly after Meta's announcement, including the Meta Model API preview and agent-specific claims. The official Meta launch page and Meta's evaluation report confirm the release, while The Verge independently confirms US developer preview access and the new API surface.

The early part is operational, not factual. Meta's claims about benchmark strength, tool-use gains, and workflow reliability still need hands-on validation in real repositories and agent stacks. Treat this as a launch worth testing, not a reason to swap production coding agents blind.

Key takeaways

• Meta Muse Spark 1.1 is now available in Meta AI Thinking mode and through the Meta Model API public preview.
• The model targets coding, computer use, multimodal perception, tool/function calling, MCP-style workflows, and parallel agent work.
• Independent reporting says the preview is currently for US developers, with $20 in credits for new API accounts.
• Meta's evaluation report says the unmitigated model hit high-risk thresholds in chemical/biological and cybersecurity domains, with mitigations applied before deployment.
• The useful next step is benchmark testing against your own agent traces, not relying on launch charts.

Availability and access

Developers can evaluate Muse Spark 1.1 through the Meta Model API public preview where eligible. The Verge reports US-developer availability, and TestingCatalog reports the model is also powering Meta AI's Thinking mode. Meta's broader consumer surfaces may expand over time, but API access, region limits, rate limits, enterprise controls, and final pricing should be checked in the Meta console before planning production use.

For agent builders, the strongest reason to test Muse Spark 1.1 is the combination of long-context claims, tool calling, multimodal input handling, and coding-oriented workflows. If your stack already routes through multiple models, add it as an eval lane first, then compare completion quality, tool-call error rate, latency, and cost per completed task.

Why it matters

Meta is entering the paid model API lane with a model aimed directly at coding agents and multi-app work. That changes procurement and routing decisions for teams that already compare OpenAI, Anthropic, Google, xAI, Z.ai, and open-weight models by task type.

The LinkLoot angle is simple: use Muse Spark 1.1 as a measured candidate for agent workflows, especially where the work involves screenshots, documents, app navigation, repository context, and tool orchestration. Keep a fallback model in place until you know how it behaves on your own failures. For practical agent stacks, see LinkLoot's guide to AI agent tools.

What to verify before you act

• Confirm your account can access the Meta Model API preview and whether your region is supported.
• Check final pricing, credit expiration, rate limits, context limits, and allowed data retention settings in the Meta console.
• Run repository-level coding evals against your existing GPT, Claude, Gemini, Grok, or GLM baseline.
• Review Meta's evaluation report for cyber, chemical/biological, prompt-injection, and deployment-safeguard notes.
• Test tool-call reliability, structured output, multimodal handling, and context compaction on real agent traces.

Source check

Confirmed by: Meta's official Muse Spark 1.1 announcement URL and Meta's Muse Spark 1.1 evaluation report confirm the model release, developer API direction, and safety-evaluation framing.

Confirmed by independent context: The Verge reports the public API preview for US developers, Meta AI Thinking mode availability, and $20 in credits for new API accounts. TestingCatalog adds fast model-release context, API preview details, and reported agentic workflow capabilities.

Early signal / context: Social posts embedded in TestingCatalog are useful for discovery, but LinkLoot is not treating them as primary evidence. API access changes, pricing, and regional rollout should be treated as update triggers.