AIREITERMuse Spark Open generally costs $1.25 per million input tokens and $4.25 per million output tokens. The right endpoint depends on its checkpoint, context cap, retention terms, and regional controls—not a small difference in a live speed reading.
Provider decision matrix
Muse Spark Open providers mostly publish the same token rates. Context limits, checkpoint evidence, privacy controls, availability, and routing behavior separate the options.
| Route | Input / 1M | Cached input / 1M | Output / 1M | Published context | Checkpoint evidence | Retention and region evidence |
|---|---|---|---|---|---|---|
| Modal shared endpoint | $1.25 | $0.15 | $4.25 | 1M | Names Muse Spark Open, NVFP4, Muse Model License | Model page does not publish a region or retention term |
| Groq limited preview | Not published on model page | Not published | Not published | 131,072 | Names meta-models/muse-spark-open; TruePoint serving | Direct terms and account pricing require verification |
| OpenRouter | $1.25 | $0.15 on supporting endpoints | $4.25 | Up to 1,048,576 | Route is meta/muse-spark-open; underlying provider can be selected | Provider filters and ZDR controls; separate US/EU in-region endpoints |
| Vercel AI Gateway | $1.25 | $0.15 where listed | $4.25 | 1M on five displayed providers | Route is meta/muse-spark-open; provider allowlists supported | ZDR routing option; Fireworks was the only displayed row explicitly marked US |
| Self-hosted official weights | Infrastructure cost | Your cache cost | Your compute cost | Up to 1,048,576 | Exact BF16 or NVFP4 repository revision can be pinned | You control storage and region, subject to the Muse Model License |
The $1.25/$0.15/$4.25 figures are a model-token baseline. Account credits, gateway terms, omitted cache support, and provider-specific billing rules can change the effective bill.
Best default: use a gateway with an explicit provider allowlist rather than accepting an opaque route. Choose a verified full-window endpoint for workloads above 131K tokens; consider Groq only when its preview access and 131K cap fit the job.
Muse Spark Open is not a Muse Spark 1.3 alias
Muse Spark Open is an officially released model with separate checkpoint names. Meta does not identify it as the same checkpoint as the closed muse-spark-1.3 endpoint, so production documentation should not treat the names as interchangeable.
Meta released Muse Spark Open on October 15, 2026. The official deployment guide names meta-models/Muse-Spark-Open-816B-A42B for BF16 and meta-models/Muse-Spark-Open-816B-A42B-NVFP4 for NVFP4. Meta describes 816 billion total parameters, 42 billion active per token, and a 1,048,576-token context window.
The Meta Model API price card lists muse-spark-1.2, muse-spark-1.3, and Contributor variants as separate hosted products. They share a 1M context and the same Standard price card, but Meta's Open launch and deployment documentation do not map Muse Spark Open to 1.3.
A provider slug is not a cryptographic identity. Self-hosters should pin a repository revision; hosted users should record the model slug, provider, route, date, and available revision metadata.
The sticker price is uniform, but the bill is not
The prevailing Muse Spark Open API price is $1.25 per million fresh input tokens, $0.15 per million cached input tokens, and $4.25 per million output tokens. Modal, Vercel's displayed providers, and OpenRouter's listed providers converge on that rate.
| Workload | Input | Cached input | Output | Estimated cost |
|---|---|---|---|---|
| Short agent turn | 10,000 | 0 | 2,000 | $0.0210 |
| Repository question | 30,000 | 0 | 2,000 | $0.0460 |
| Repeated large-context turn | 12,000 fresh | 90,000 cached | 1,000 | $0.03275 |
| Same turn without cache | 102,000 fresh | 0 | 1,000 | $0.13175 |
Use this formula:
cost = fresh_input × $1.25/M + cached_input × $0.15/M + output × $4.25/M
A 90% cache hit on a reusable 100K-token prefix cuts the example from $0.13175 to $0.03275, a 75% reduction. That saving requires an endpoint with cache-read billing and a prompt prefix that remains cacheable; Nebius did not display a separate cache-read price in the OpenRouter provider table captured on October 16, 2026.
Output discipline matters because output costs 3.4 times more per token than fresh input. Vercel's model documentation states that reasoning tokens typically count toward output usage, although reporting and billing can vary by provider.
The $0.10 input and $0.20 output Contributor rate belongs to Meta's hosted Muse Spark 1.2 and 1.3 Contributor products, not Muse Spark Open. Meta's pricing-mode documentation says Contributor prompts, code, files, context, and responses may be used to train future Meta models; Standard content is not used to train Meta models.
Context limits vary by deployment
Muse Spark Open supports 1,048,576 tokens at the model level, but an API provider can expose less. Groq's official limited-preview page caps both context and output at 131,072 tokens.
Modal publishes a 1M context for its shared endpoint. Vercel's model table showed 1M for Fireworks, Baseten, Together AI, Modal, and Nebius on October 16, 2026. OpenRouter publishes an exact 1,048,576-token model context, although the selected endpoint still determines operational behavior.
| Surface | Context shown | Max output shown | Important caveat |
|---|---|---|---|
| Official self-hosted vLLM configuration | 1,048,576 | Request-defined within context budget | Meta advises lowering --max-model-len when the workload needs less context |
| Modal shared endpoint | 1M | Not stated on model page | Verify request and modality limits before migration |
| Vercel displayed providers | 1M | 1M | Prompt and response share the context budget |
| OpenRouter model route | 1,048,576 | Endpoint-dependent | Pin a provider if the limit must remain stable |
| Groq limited preview | 131,072 | 131,072 | Five-image maximum; organization approval required |
Meta's deployment guide notes that full-attention KV memory grows with context and recommends lowering --max-model-len when the workload needs less. Count prompts, media, tool schemas, tool results, reasoning, and output against the endpoint's shared budget.
Test above the expected 95th-percentile prompt size before launch. A successful 20K-token request does not validate a 500K-token workflow.
Retention and regional routing require endpoint-level proof
“No training,” “zero data retention,” and “in-region processing” are different controls. Confirm each one for the exact route that will receive production data.
| Surface | Training-use evidence | Retention/ZDR evidence | Region control |
|---|---|---|---|
| Meta Model API Standard Spark | Content is not used to train Meta models | No general Spark ZDR promise established by the cited price page | Product page says expanded global access but provides no country matrix |
| Meta Model API Contributor Spark | Content and responses may train future Meta models | Contributor is not a privacy tier | Availability varies by product and account |
| OpenRouter | Provider-specific policies | ZDR-only routing and provider metadata | us.openrouter.ai and eu.openrouter.ai |
| Vercel AI Gateway | Provider-specific policies | zeroDataRetention routing option | Provider metadata; captured table explicitly marked Fireworks as US |
| Direct hosted provider | Check direct agreement | Check direct agreement | Check deployment region and subprocessors |
OpenRouter's provider directory classified Modal, Groq, and Nebius Token Factory as zero-retention providers when checked on October 16, 2026. That directory describes routes in OpenRouter's ecosystem; it should not replace the security terms for a direct provider contract.
Use this verification sequence before sending private repositories or regulated data:
- Select the exact model and underlying provider instead of unrestricted auto-routing.
- Require ZDR or a contractual retention duration; “not used for training” is insufficient by itself.
- Select a US or EU regional endpoint when processing location is mandatory.
- Disable gateway, observability, and application-body logging or align their deletion periods.
- Store response metadata that identifies the route used.
- Recheck the policy after adding web search, file storage, remote tools, or other subprocessors.
Pick a route by deployment risk
The best Muse Spark Open API route depends on the constraint that would make the deployment fail. Speed snapshots change with load, prompt size, region, and routing, so treat them as observations rather than durable rankings.
| Requirement | Recommended route | Reason |
|---|---|---|
| Full-context repository or document analysis | Modal, a pinned 1M OpenRouter provider, or a verified Vercel provider | Published 1M context; avoids Groq's 131K cap |
| Strong displayed gateway throughput | Retest Baseten, then pin it if results hold | Vercel displayed 269 TPS and OpenRouter 249 TPS for Baseten on October 16, 2026; these are live telemetry snapshots |
| Low-latency workloads below 131K | Groq preview, if approved and priced acceptably | Groq publishes roughly 350+ TPS but caps context at 131,072 |
| ZDR through a gateway | OpenRouter or Vercel with ZDR enforcement | Both expose routing controls; verify the eligible endpoint and caching implications |
| Strict US or EU processing | OpenRouter in-region endpoint plus a provider allowlist | Region is enforced separately from provider identity |
| Maximum checkpoint reproducibility | Self-host an official repository revision | Hosted route names do not expose a checkpoint hash or runtime build |
| Cheapest non-sensitive experimentation | Muse Spark 1.3 Contributor, not Muse Spark Open | $0.10/M input and $0.20/M output, with training-use permission and lower Meta rate limits |
For a typical coding-agent team, pin one full-context provider for production and a compatible second provider for failover. Enforce retention and region at the gateway, cap reasoning output, and record provider metadata during evaluations.
Do not choose Groq solely for speed if the 131K cap or limited-preview approval blocks the workload. Do not allow gateway failover to cross a retention or regional boundary. Do not send private code to Contributor endpoints unless the organization accepts Meta's training-use terms.
Muse Spark Open API pricing FAQ
How much does Muse Spark Open cost per million tokens?
The common hosted baseline is $1.25 per million input tokens, $0.15 per million cached input tokens where supported, and $4.25 per million output tokens. Account credits, cache support, and provider billing rules can change the effective cost.
Is there a Muse Spark Open Contributor price?
No official Muse Spark Open Contributor endpoint appears in Meta's release or deployment documentation. The $0.10 input and $0.20 output Contributor prices apply to hosted Muse Spark 1.2 and 1.3 variants with training-use permission.
Are reasoning tokens billed as output tokens?
Vercel says reasoning tokens typically count toward output usage, with reporting and billing varying by provider. Budget them at the output rate unless the selected endpoint documents another rule.
Is the API OpenAI-compatible?
Modal calls its shared endpoint OpenAI-compatible, and Groq documents an OpenAI-style client example. Meta's deployment guide documents OpenAI-compatible HTTP for SGLang, vLLM, TensorRT-LLM, llama.cpp, MLX, and OpenAI/Anthropic-compatible HTTP for Unsloth. Parameter and modality support still varies.
Is there a free trial?
The Vercel model page stated on October 16, 2026 that users who had not made a payment received $5 of AI Gateway credits every 30 days. Treat this as a time-sensitive account offer, not a free tier attached to the model.
Sources: Meta release · Meta deployment guide · Meta pricing and rate limits · Modal model page · Groq model page · OpenRouter model page · Vercel AI Gateway model page
Related reading: Muse Spark Open's self-hosting, license, and benchmark tradeoffs