
Muse Wonder vs Muse Spark: Upgrade or Route Per Task?
By Olivia Hart
MyClaw Editorial
MyClaw
Run Best-in-Class AI Agents Now
Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.
AI Takeaway
- Should you switch from Muse Spark to Muse Wonder? Not wholesale: keep Spark 1.3 as the default for routine agent traffic and route only long, failure-prone jobs to Wonder, because Wonder costs about four to five times more per token for a four-point gain on the main independent index.
- How much more does Muse Wonder cost than Muse Spark? Meta's price list, checked Oct 13, 2026, puts Spark 1.3 Standard at $1.25 per million tokens in and $4.25 out, Spark Contributor at $0.10 and $0.20, and Wonder at $4 and $20 with no Contributor option.
- How much better is Wonder on benchmarks? On the independent Artificial Analysis index, Wonder reaches 52 and Wonder Full 56 against 48 for Spark 1.3, with the biggest jumps on terminal and automation tasks.
- Is Wonder cheaper per finished task? No, in Artificial Analysis's single run a Wonder index task cost around $4.52, against roughly $1.60 on Spark 1.3, partly because it read about 28% more input.
- Can one agent use Spark and Wonder together? Yes, Meta says both take the same request format, so a router can send easy steps to Spark and escalate hard ones to Wonder by changing only the model ID.
Every Spark Version and Both Wonder Models in One Table
| Model | API ID | Released | Pricing tiers | Where you get it |
|---|---|---|---|---|
| Muse Spark 1.1 | muse-spark-1.1 | Jul 9, 2026 | Standard only | Meta Model API |
| Muse Spark 1.2 | muse-spark-1.2 | Aug 5, 2026 | Standard and Contributor | Meta's API and Muse Code |
| Muse Spark 1.3 | muse-spark-1.3 | Sep 2, 2026 | Standard and Contributor | Meta's API and Muse Code |
| Muse Wonder | muse-wonder-1.0 | Oct 12, 2026 (preview) | Standard only | Meta's API and Muse Code |
| Muse Wonder Full | Not on the API | Oct 12, 2026 | Subscription only | Muse Code paid plans |
Every model in the table shares a context window of about 1.05 million tokens and accepts text, images and video. Meta's own model list flags audio on Spark 1.3 as not fully supported, so test audio jobs before relying on it. Meta's docs now call 1.2 the previous version and 1.1 the original.
Muse Glimmer, Meta's open-weight model distilled from Spark, is a separate local option and not part of this decision.

Price Ladder: Contributor, Standard, Wonder and Wonder Full
All per-token prices below are Meta's, from its Model API pricing page on Oct 13, 2026.
| Tier | Input per 1M | Output per 1M | Cached input per 1M | Rate limits |
|---|---|---|---|---|
| Spark Contributor (1.3, 1.2) | $0.10 | $0.20 | $0.002 | 100 requests, 3M tokens per minute |
| Spark Standard (1.3, 1.2, 1.1) | $1.25 | $4.25 | $0.15 | 3,000 requests, 4M tokens per minute |
| Wonder Standard | $4.00 | $20.00 | $0.20 | Listed by third parties at Standard limits |
| Wonder Full | No token price | No token price | n/a | Muse Code plans only |
Wonder Full sits inside Muse Code's two top subscriptions, reported at $50 and $200 a month. Its subscription key is locked to the Muse Code command-line tool and cannot be reused as cheap API access.
The Contributor tier is the cheapest rung by a wide margin, but it has a real cost: Meta may use your prompts and completions to train future models. Keep it for prototypes, public data and load tests. Client documents, health or legal material and anything under a confidentiality agreement belong on Standard.
What the Multipliers Mean for a Real Workload
Take an agent that processes 10 million input tokens and 1 million output tokens a day, with no caching. At Meta's list prices, our arithmetic gives:
| Tier | Daily cost | 30-day cost |
|---|---|---|
| Spark Contributor | $1.20 | $36 |
| Spark Standard | $16.75 | $502.50 |
| Wonder | $60.00 | $1,800 |
That is roughly 3.6 times more on Wonder than on Spark Standard for identical token counts. Real workloads shift the gap: Wonder tends to read more input per task, while it writes slightly less output. Artificial Analysis logged about 7.69M input tokens on each Wonder index task versus 6.03M on Spark 1.3, and around 54,000 output tokens against 60,000. Its resulting cost per task was about $4.52 for Wonder and $1.60 for Spark, close to a 2.8x gap. That is one lab's harness, so measure your own. For Wonder's full pricing detail, including the Muse Code plans, see our Muse Wonder guide.
Benchmarks: How Big Is the Jump?
Artificial Analysis (Independent, Max Effort)
| Test | Spark 1.3 | Wonder | Wonder Full |
|---|---|---|---|
| Intelligence Index v4.3 | 48 | 52 | 56 |
| Terminal-Bench 4.0 | 33% | 52% | 57% |
| AutomationBench-AA | 58% | 68% | 71% |
Plain Wonder's headline gain is modest at four index points, but the agentic rows move a lot. Terminal work improves by nearly 20 points, and automation tasks by 10. Wonder Full improved on Spark 1.3 across all ten of the lab's index evaluations. Knowledge recall remains a weak area for the whole Muse family, partly offset by low hallucination rates.

Meta's Own Charts
Meta published a plot comparing Spark 1.3 with Wonder Full at several reasoning settings. Its claim is that Wonder Full does better on unfamiliar tasks than Spark's trend line predicts, which Meta attributes to a larger base model. That is a vendor analysis without independent replication yet. Third-party reviewers have also pointed out that earlier Meta tables mixed reasoning settings between Spark versions, so compare like with like. Our Muse Spark 1.3 deep dive covers how the smaller model behaves at each effort level.
Decision Guide: Which Tasks Go Where
| Task type | Route to | Reason |
|---|---|---|
| Prototypes, synthetic data, public-data bulk jobs | Spark Contributor | Lowest price, training trade-off acceptable |
| Production chat, classification, short tool calls | Spark 1.3 Standard | Cheap, fast, private |
| Long refactors, research across many files | Wonder | Biggest agentic gains |
| Assistants that need memory across sessions | Wonder | Built around Meta's Active Memory |
| Heavy terminal coding on a flat monthly fee | Wonder Full via Muse Code | Highest Muse scores, no token billing |
Build a Spark-First Cascade
- Send every request to Spark 1.3 by default.
- Escalate to Wonder when tests fail, a tool call errors twice, or the task is tagged long-horizon.
- Log how often escalation fires and what each escalated task costs.
- Review weekly. If more than a third of jobs escalate, Wonder may be the better default for that workload.
Because both models share one request format, the switch is a single field. Keep prompts identical so the comparison stays honest.
When to Skip Both
If you need the highest independent scores rather than the best Meta option, look outside the family. Our Wonder and Opus 5.5 head-to-head and the Spark versus Gemini matchup cover the main rivals.
Access and Data-Use Differences
- Status. Wonder is labelled preview, so limits and behavior can shift. Spark 1.3 is the version Meta recommends for new work.
- Countries. Meta says availability for both varies by country and is shown during account onboarding.
- Training on your data. Wonder is offered solely on Meta's Standard tier, which Meta says keeps your content out of training. Spark offers both modes, so check which model ID your code calls.
- Muse Code default. Muse Code subscriptions start in Contributor mode, which you can switch off in settings.
- Age. Both require users aged 18 or over.
The original Muse Spark review covers how Meta's API launched and where access began.
Verdict: Spark by Default, Wonder by Escalation
Muse Spark Pros:
- Far lower token prices, plus a near-free Contributor tier
- Cheaper per completed task in independent testing
- Three released versions and Meta's recommended default
Muse Spark Cons:
- Four to eight index points behind the Wonder models
- Weaker on terminal and automation work
- Audio on 1.3 is flagged by Meta as incomplete
Muse Wonder Pros:
- Large gains on agentic coding and automation
- Standard-only mode keeps your data out of training
- Drop-in swap from Spark
Muse Wonder Cons:
- Roughly 3x the cost per task in independent testing
- Preview status
- Best variant, Wonder Full, is not available on the API
Default to Spark 1.3 and escalate to Wonder for the jobs where a failed attempt costs you more than the tokens.
Run Muse Spark as an Always-On Agent on MyClaw

MyClaw hosts OpenClaw and Hermes Agent and supports Muse Spark as a model choice, so the default tier from this guide can go to work without running servers. Wonder is not available on MyClaw, so treat it as an escalation through Meta's own API. For other model picks, see our OpenClaw model guide.
- Spin up a MyClaw workspace and select OpenClaw or Hermes Agent to run it.
- Choose Muse Spark as the agent's model in MyClaw, connect your chat app for replies, and enable just the integrations the task calls for.
- Hand it one recurring job, then compare its run log and token use after a week to see whether any steps truly need Wonder.
Skip the Setup, Run Best-in-Class AI Agents Now
Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.