Muse Spark 1.1 Brings 1M Context, But Thin API Docs

Meta Muse Spark 1.1 is the first Meta AI developer model I’d put into long-context coding evals immediately, and the last one I’d route production traffic through today without a vendor packet.

The reason is simple: Meta documents a 1 million token context window with active compaction in the official Muse Spark 1.1 launch post, while TechCrunch reports Reuters pricing of $1.25 per million input tokens and $4.25 per million output tokens in its July 9 launch report. That makes repo-scale coding cheap enough to test aggressively.

The missing part is operability: endpoints, SDK coverage, rate limits, retention terms, training opt-out language, and enterprise controls were not captured in the launch-day primary docs.

What changed in Meta Muse Spark 1.1?

Meta Muse Spark 1.1 changed Muse Spark from a consumer-first model into a developer API preview. The confirmed July 9 delta is public Meta Model API access, 1M-token context with active compaction, coding primitives, multimodal perception, and agentic workflow support.

The April 8 predecessor was introduced as the first Muse-family model from Meta Superintelligence Labs, available through meta.ai and the Meta AI app with private API preview access for selected users, according to Meta’s original Muse Spark announcement. July 9 is the developer-surface shift.

Meta’s launch language is specific enough to matter. The Muse Spark 1.1 post says the model has a 1 million token context window and active compaction that remembers actions, retrieves earlier information, and preserves critical steps for later work.

That maps directly onto a coding-agent failure mode. Long sessions lose the original requirement, stale tool output sticks around, and the model forgets why a file was edited three steps earlier.

The cautious read is that Meta has documented the mechanism at a product level. It has not published the compaction algorithm, failure modes, or account-level limits.

What is Muse Spark API pricing in 2026?

Muse Spark API pricing in 2026 is reported at $1.25 per million input tokens and $4.25 per million output tokens. Those figures come from Reuters via TechCrunch, so engineering teams can use them for eval planning, but procurement should wait for a first-party Meta pricing page.

The reported price changes the eval math for large contexts. A task with 800K input tokens and 80K output tokens costs about $1.34 at the reported Muse Spark 1.1 rate: $1.00 for input and $0.34 for output.

That’s the right unit for a first pass, but it’s still incomplete. Real cost is cost per accepted patch, including retries, test failures, reviewer repair time, and latency.

The Verge also reported $20 in free account credits for the public preview in its launch coverage. No captured first-party Meta page confirmed batch pricing, cached-input pricing, free-tier policy, or rate-limit tiers.

Is Muse Spark better than Claude for coding?

The sourced record does not establish a universal coding winner. Muse Spark 1.1 is the more interesting experiment when workload cost is dominated by large input context; the production decision still belongs to accepted-patch rate, latency, retry cost, and the controls documented by each provider.

The strongest Muse Spark coding case is economic. Long-context repo tasks are input-heavy, and input-heavy workloads punish expensive context windows.

Meta also frames the model around coding primitives for detecting and fixing complex bugs in the official launch post. Its developer guide is the right first-party source for build patterns rather than launch coverage.

Benchmarks need discipline here. Artificial Analysis says it independently evaluated the original Muse Spark at 52 on its Intelligence Index in its April evaluation. Scale Labs’ live SWE-Bench Pro private leaderboard, fetched in the July 9 audit, showed Muse Spark 1.1 at 51.5 using the mini-swe-agent harness.

Those two numbers answer different questions. AAII is a broad intelligence score for the original Muse Spark evaluation. SWE-Bench Pro is closer to coding-agent reality, but it is still a benchmark harness, not your repository.

Use public benchmarks to choose candidates. Use accepted-patch rate to choose a router.

What is still undocumented in the Muse Spark API?

The launch-day Muse Spark API gap is operational documentation. The research did not capture first-party endpoint references, SDK language lists, rate limits, data-retention terms, training opt-out language, knowledge cutoff, SLA commitments, or enterprise controls such as SSO, SCIM, IAM, audit logs, VPC, and data residency.

This is the part that decides production readiness. Endpoint compatibility cannot be verified without endpoint references. Load tests cannot be planned without rate limits. Security review stalls without retention and training-use terms.

The tool-use claim also needs exact wording. Meta says Muse Spark 1.1 can zero-shot generalize to new native tools, MCP servers, and custom skills in the official launch post. That does not prove every preview account has every MCP or tool surface enabled on day one.

The same distinction applies to multimodal workflows. Meta’s July 7 Muse Image and Muse Video announcement establishes the media-model family and integration path, while the July 9 Muse Spark post documents native multimodal perception across images, video, and documents. Account-level availability still needs verification in your preview environment.

This may be a timing artifact. The API launch and this audit share the same date, July 9, 2026.

But launch timing doesn’t help your security questionnaire.

Is Muse Spark 1.1 open source?

Muse Spark 1.1 is a proprietary, closed-weight Meta AI developer model. The public record supports that through Meta’s hosted Meta Model API release framing, the absence of downloadable Muse Spark weights in the captured research, and Artificial Analysis describing Muse Spark as a closed model.

This is a strategic break from the Llama distribution playbook. The public evidence does not establish Muse Spark 1.1 model lineage; it does establish a hosted, closed-weight API release rather than a downloadable-weights release.

For teams that chose Llama for local inference, auditability, fine-tuning, or data sovereignty, Muse Spark 1.1 changes the operating model. You get Meta-hosted frontier-style access. You give up direct control of weights.

That trade can still make sense for long-context coding. Operating a 1M-token coding model with low latency and stable compaction is infrastructure-heavy work.

The procurement question is different. Llama can enter a private deployment architecture. Muse Spark 1.1 needs vendor documentation.

What should teams test Monday morning?

Teams should test Muse Spark 1.1 where 1M context can change the result: repo-wide analysis, multi-file bug fixes, long-session recall, tool-augmented workflows, and cost-latency endurance runs. Measure accepted patches, retries, latency, and total task cost before comparing benchmark headlines.

Start with tasks that currently burn context. Tiny prompts will hide the only launch-day advantage that’s clearly worth pursuing.

Run the same prompt envelope across Muse Spark 1.1 and your incumbent coding model. Record input tokens, output tokens, time to first token, total latency, retries, passing tests, and reviewer edits.

A practical first eval set:

1. One real bug fix from your backlog with issue text, relevant files, and tests.
2. One multi-file feature across three to six files.
3. One repo-wide migration analysis above 200K tokens.
4. One recall probe with key facts placed near the beginning and end of a large bundle.
5. One MCP-style tool trial, only if your account exposes the needed surface.
6. One multimodal task, only after your preview account confirms supported inputs.

Use a blunt pass bar. Tests pass. The diff is reviewable. Total cost drops. Manual repair doesn’t erase the savings.

What this means for you

US developers should put Meta Muse Spark 1.1 into long-context coding evals now. The combination of CONFIRMED_PRIMARY 1M context and REPORTED $1.25/M input pricing is too material to ignore.

Enterprise teams should ask Meta for a document pack before production traffic: endpoint references, SDK support, model IDs, rate limits, retention terms, training opt-out, audit logs, IAM, SSO or SCIM, private connectivity, data residency, and SLA terms.

Benchmark comparisons should use verified model names and visible sources. In this audit, the publishable coding benchmark is Scale Labs’ SWE-Bench Pro value of 51.5 for Muse Spark 1.1, and the publishable broad score is Artificial Analysis’ 52 for the original Muse Spark evaluation.

The operating position is narrow and useful: evaluate Muse Spark 1.1 where long context is expensive, measure cost per accepted patch, and hold production adoption until Meta turns the API preview into an operable platform.

Sources

• Meta: Introducing Muse Spark on the Meta Model API
• Meta developer guide: Build with Muse Spark
• Meta: Introducing the original Muse Spark
• Meta: Introducing Muse Image and Muse Video
• TechCrunch: Meta enters the crowded AI coding battle with Muse Spark 1.1
• The Verge: Meta Muse Spark Model API launch coverage
• Artificial Analysis: Muse Spark evaluation
• Scale Labs: SWE-Bench Pro private leaderboard