Muse Spark 1.3, Meta’s coding and agent-focused model launched September 2, 2026 — the fourth release in the Muse Spark line in five months, following a rapid cadence of 1.1 in July and 1.2 in August — keeps the same pricing as its predecessor ($1.25 per million input tokens, $4.25 per million output) and adds a genuinely large 1M-token context window with near-perfect long-context retrieval. The catch sits inside Meta’s own headline comparison: several outlets, including a detailed breakdown from eesel AI, found that Meta’s testing methodology ran most of 1.3’s “biggest jump yet” scorecard in its gated max reasoning mode against 1.2’s already-available xhigh mode — a reasoning-tier change dressed up as a generational leap, not a clean apples-to-apples comparison.
“How much better is 1.3” isn’t answerable from Meta’s own materials alone
Meta’s official comparisons for Muse Spark 1.3 look impressive on their face — but multiple independent outlets covering the release, including reporting that traced Meta’s own testing methodology notes, found the company’s scorecard mostly compares 1.3 running in max mode against 1.2 running in xhigh mode. Max mode isn’t available to any customer yet — Meta says it’s still completing safety testing and will ship it “shortly.” That means a meaningful share of the advertised generational jump is really a comparison between two different reasoning-effort settings, not two different model generations under matched conditions. The efficiency claims that get repeated everywhere — about 20% fewer tool calls, about 25% fewer tokens than 1.2 — come from Meta’s own internal engineering comparisons, and as of this review, no independent lab has confirmed those specific figures.
Pricing has three tiers, and one of them costs your training data
The standard API tier didn’t move: $1.25 per million input tokens and $4.25 per million output, identical to Muse Spark 1.2, with an 88% discount on cached input. What’s new and worth knowing is a second, much cheaper tier:
| Tier | Price (per 1M tokens) | The actual trade |
|---|---|---|
| Standard | $1.25 input / $4.25 output | No data-use concessions required — this is the tier most business use should default to |
| Contributor | $0.10 input / $0.20 output | Roughly 10–20x cheaper, in exchange for letting Meta train on the data you submit |
| Max mode | Not yet priced or available | Currently in a limited partner preview only; Artificial Analysis lists no API provider offering it at all |
The contributor tier is a genuinely explicit, quantified version of a trade-off our piece on shadow AI for small business covers in softer terms for other vendors’ consumer tiers: here, Meta is putting an exact price on the difference between keeping your inputs private and letting them train the model, rather than burying it in a settings toggle most people never open. If you’re weighing the contributor tier for cost reasons, treat any client-confidential or NDA-covered material the same way our AI usage policy guide recommends — name which tier is approved for which kind of work, in writing, before a subcontractor or teammate defaults to the cheaper option without thinking about what it costs in data terms.
What’s genuinely real here
- A large, capable context window. 1M tokens, with near-perfect long-context retrieval scores (98.5 and 98.1 on the benchmarks Meta cites) — this is a real strength for anyone working with large codebases or long documents in a single session.
- Collaborative behavior improvements. Meta says the model asks clarifying questions when a request is ambiguous and requests confirmation before taking actions with irreversible consequences — a genuinely useful safety-relevant behavior for an agentic coding tool, if it holds up in practice as consistently as advertised.
- A real, if smaller, capability gain independently confirmed. Artificial Analysis’s own Intelligence Index tracks Muse Spark’s trajectory at 53 (version 1.1, July), 57 (1.2, August), and 61 for the currently available 1.3 xhigh configuration — a real four-point independent gain, even setting aside Meta’s own larger claims, and a reminder that comparing sticker numbers across releases still needs the same care our guide to free vs. paid AI tools recommends for any vendor’s pricing or performance claims. That score ties Muse Spark 1.3 with GPT-5.6 Sol (max), Grok 4.6 (high), and Claude Opus 5 (high), and sits behind Claude Fable 5.1 (max, 66) and Claude Opus 5 (max, 63).
The version being marketed isn’t the version you can actually use yet
This is the same access-gating pattern showing up across nearly every major model launch this month — our GPT-6 Astra review covers the same dynamic for OpenAI’s own launch. VentureBeat’s coverage confirms Artificial Analysis evaluated Muse Spark 1.3’s max mode only in a limited partner preview, and currently lists no API provider offering it at all. If a comparison chart you’re looking at cites Muse Spark 1.3’s strongest numbers, check which mode those numbers were measured on — the model you can actually deploy today is the xhigh configuration, not the one behind the flashiest marketing claims.
Weights stay closed, and Meta hasn’t said whether that changes
Muse Spark 1.3 is closed — Meta hasn’t published a parameter count, and self-hosting isn’t currently offered. This is a real shift in posture from Meta’s Llama-era reputation for open releases, though it’s worth being precise about what’s actually confirmed: Meta has not stated whether Muse Spark weights will ever be released, rather than ruling it out outright. If open weights specifically matter to your use case, Meta’s own Llama line and its smaller Muse Glimmer model remain the open options in its current lineup — Muse Spark itself is not one of them, for now.
Rankings disagree, too — check what list you’re actually reading
Third-party trackers don’t even agree on where Muse Spark 1.3 ranks: one lists it at #6 of 636 tracked models overall, another at #13 of 435, and a coding-specific tracker places it at #9 of 208 for code generation specifically. These aren’t contradictions so much as a reminder that “#6 overall” and “#13 overall” are only meaningful relative to the specific list and model count each tracker maintains — treat any single ranking claim, including ones in this review, as tied to its source rather than as an absolute fact about the model.
Who should actually use this, and who should wait
Teams already building on Meta’s Muse Code agent or the Meta Model API, and anyone whose workflow benefits from very large context windows, have a real reason to try Muse Spark 1.3 today at the standard tier. Cost-sensitive teams considering the contributor tier should weigh the 10–20x discount specifically against what kind of data they’d be submitting — fine for non-sensitive experimentation, a real liability question for anything touching client or regulated data. Anyone waiting specifically for the numbers in Meta’s own marketing materials should hold off: those come from max mode, which isn’t shippable to any customer yet. And if you’re choosing between this and a same-week competitor rather than in isolation, our Muse Spark vs Gemini 3.8 Flash comparison covers how sharply independent benchmarks disagree on that specific matchup, and our Cursor vs Claude Code vs Copilot piece is worth reading if the tool wrapped around the model matters more to your workflow than the model itself, a point also covered from a different angle in our coding-agent harness guide.
Is Muse Spark 1.3 actually better than 1.2?
Independently, yes by a real but modest margin — Artificial Analysis’s Intelligence Index moved from 57 to 61 for the currently available configuration. Meta’s own larger “biggest jump” claim mixes in a comparison against its own gated max mode, which isn’t available to anyone yet, so the full scale of Meta’s advertised improvement isn’t independently verifiable at this point.
How much does Muse Spark 1.3 cost?
The standard tier is $1.25 per million input tokens and $4.25 per million output, unchanged from Muse Spark 1.2. A separate “contributor” tier runs roughly 10–20x cheaper but requires allowing Meta to train on the data you submit.
What’s the “contributor” pricing tier, and is it worth it?
It’s a much cheaper API tier — $0.10 input and $0.20 output per million tokens — in exchange for letting Meta use your submitted data for training. It’s a reasonable trade for non-sensitive experimentation, and a real liability question for anything involving client-confidential or regulated data, the same way any free or discounted AI tier’s training defaults deserve a second look before business use.
Can I access Muse Spark 1.3’s best benchmark configuration right now?
No. The max reasoning mode behind Meta’s strongest cited numbers is still in limited partner preview and completing safety testing; Artificial Analysis currently lists no API provider offering it. The version available today through Muse Code and the Meta Model API runs in xhigh mode.
How does Muse Spark 1.3 compare to Gemini 3.8 Flash or Claude?
It’s genuinely close and contested — independent benchmarks disagree sharply on Muse Spark 1.3 versus Gemini 3.8 Flash specifically, covered in detail in our separate comparison. Against Claude, Muse Spark 1.3’s currently available configuration sits behind Claude Fable 5.1 and Claude Opus 5’s top configurations on Artificial Analysis’s Intelligence Index.
Is Muse Spark 1.3 available everywhere yet?
It’s rolling out through Muse Code and the Meta Model API. Some regional rollout inconsistencies have been reported informally, though these aren’t independently confirmed — if availability in your region matters, check directly with Meta’s own API status rather than assuming immediate global access.
Shurah is the founder of AI Tools Daily, tracking pricing, licensing and policy changes across AI tools so readers can make decisions without wading through marketing claims themselves.
