Price terms as published
Each line is the vendor’s own wording, split from the price capture checked on 2026-10-15. Nothing here is rounded, converted or re-worded — the full sentences stay in the review below. Lines that carry only a price continue the plan named on the line above, in the vendor’s own order.
| # | Plan / term | Checked |
|---|---|---|
| 1 | Standard $1.25/M in | 2026-10-15 |
| 2 | $4.25/M out | 2026-10-15 |
| 3 | $0.15/M cached | 2026-10-15 |
| 4 | Contributor $0.10/$0.20/M | 2026-10-15 |
The short version
Muse Spark 1.3 is Meta’s frontier model for agentic and coding work, released September 2, 2026 and served through Muse Code and Meta Model API. It reads text, image, video and PDF, writes text, and carries a 1,048,576-token context window with no long-context premium - you pay the same rate whether the window is empty or full.
Everything below was read off Meta’s own documentation on 2026-09-29. This is an official-source review: we did not run the model, and every benchmark figure below is attributed to whoever measured it.
Pricing, straight from the rate-limits page
| Tier | Model IDs | Cached input | Input | Output |
|---|---|---|---|---|
| Standard | muse-spark-1.3, -1.2, -1.1 |
$0.15 | $1.25 | $4.25 |
| Contributor | muse-spark-1.3-contributor, -1.2-contributor |
$0.002 | $0.10 | $0.20 |
All figures are per 1M tokens. The three Standard versions share one price - 1.3 costs exactly what 1.2 and 1.1 cost, which is the main reason it is a straight upgrade rather than a decision.
The Contributor tier is the number most write-ups skip. It is roughly 12x cheaper on input and 21x on output, and the trade is explicit in Meta’s own words: on Contributor, your prompts and completions are used to train future Meta models. Standard tier data is never used for training. For prototyping and integration tests where that is acceptable, Contributor is the cheapest frontier-adjacent token pricing in Meta’s lineup.
Two extras worth knowing: web search grounding is $2.50 per 1,000 queries on top of token cost, and rate limits are per team, not per API key - 3,000 RPM / 4M TPM on Standard, 100 RPM / 3M TPM on Contributor.
What changed in 1.3, and one thing that did not
Meta’s framing is longer-horizon work: holding many constraints across a long thread, asking for clarification when a prompt is ambiguous, confirming before consequential actions, and noticing when it is stuck instead of inventing an answer. On multitasking, Meta specifically claims better mapping of incoming prompts to the right task inside a messy single thread, including when you interrupt or redirect a prior request.
The regression is concrete and documented on the models page: audio understanding in 1.3 is “currently not fully supported,” and response quality on audio requests may be degraded. Meta’s own recommendation is to use Muse Spark 1.2 for audio, or Muse Voice Transcribe ($0.18/hour) for dedicated speech-to-text. If your agent pipeline ingests audio, 1.3 is not the version to move to.
Benchmarks, with the measuring stick shown
These are third-party measurements from Artificial Analysis (September 2, 2026), not Meta’s own scorecard and not ours:
| Benchmark | Muse Spark 1.2 | 1.3 (xhigh) | 1.3 (max) |
|---|---|---|---|
| AA Intelligence Index | 57 | 61 | 62 |
| Tau3-Bench Banking | 35% | 47% | 52% |
| Terminal-Bench 2.1 | 80% | 85% | 86% |
| GDPval-AA v2 (Elo) | 1,615 | 1,709 | 1,754 |
| GPQA Diamond | 90% | 94% | 94% |
| CritPt (scientific) | 18% | 26% | ~25% |
| AA-LCR | 83% | 79% | 79% |
xhigh ties GPT-5.6 Sol (max) and Grok 4.6 (high) at 61; the max variant’s 52% on Tau3-Bench Banking was the top score on that evaluation at launch. Two numbers went backwards: AA-LCR fell 4 points for both variants, and AA-Omniscience accuracy fell 3 points for xhigh. Artificial Analysis attributes the omniscience drop to a higher abstention rate - the model declines more often when it is unsure, which also pulled its hallucination rate down. For agent work that is usually the trade you want, but it does mean fewer straight answers.
By the same measurement, xhigh costs $0.55 per Intelligence Index task against $0.95 for GPT-5.6 Sol (max) and $1.23 for Claude Opus 5 (high), at an index score of 59 or above no model was cheaper.
Where the second-hand coverage is now out of date
Worth flagging because most of the coverage still contradicts the docs: the Artificial Analysis
launch piece reports that the max variant was in limited preview with pricing unannounced, and
that 1.3 shipped in two variants. Both statements were true on September 2. As of 2026-09-29
Meta’s models page documents muse-spark-1.3 as supporting all reasoning effort levels
including "max", on the Standard tier only, sharing the single $1.25/$4.25 standard rate.
One model ID, one price, no preview gate.
Who should use it
- Building an agent on a budget - Standard tier at $1.25/$4.25 with a 1M window and no long-context surcharge is the whole pitch. The cache rate of $0.15 makes long-running agent loops materially cheaper than the headline output number suggests.
- Prototyping where training consent is fine - Contributor at $0.10/$0.20 is the cheapest way to run real agent traces at volume.
- Piping in audio - stay on 1.2. This is the one workload where the newest version is the wrong choice, on Meta’s own documentation.
- Comparing against closed frontier tiers - read the AA table above with its regressions included; the roundup posts that quote only the index score skip the part where 1.3 got worse.
The catch
1.3 is the fourth Muse Spark release in five months, and Meta shipped 1.1, 1.2 and 1.3 inside ten weeks. Expect the next version quickly, and expect this page to need a re-check sooner than most.
Pricing, model IDs, context window, rate limits and the audio limitation were read from Meta’s model API docs and the rate-limits page on 2026-09-29 and re-read on 2026-10-06 and again on 2026-10-15 with every price unchanged — $1.25 / $4.25 / $0.15 standard, $0.10 / $0.20 / $0.002 on Contributor, plus the $2.50 per 1,000 web-search grounding fee and $0.01 per Muse Image. Benchmark figures are Artificial Analysis measurements dated 2026-09-02. We have not run benchmarks ourselves and do not score this model.