Muse Spark 1.3 Contributor Launches at $0.10/$0.20: Meta's Cheapest Coding Model, Paid for in Data

Meta released Muse Spark 1.3 Contributor on September 2, and the surprise is not the model — it is the price tag. Muse Spark 1.3 Contributor is the data-sharing SKU of the new Muse Spark 1.3 checkpoint, listed at $0.10 per million input tokens and $0.20 per million output tokens against $1.25 and $4.25 for the standard Muse Spark 1.3: a roughly 12× discount on input and 21× on output that buys Meta the right to train on your prompts and completions. Alexandr Wang, who leads Meta's AI work, called the pricing "aggressive," said a "meaningful double-digit percentage" of developers already pick the Contributor tier that Meta introduced alongside Muse Spark 1.2 in August, and framed the release as groundwork for personal agents that work on a user's behalf around the clock. The model went live in the Meta Model API and inside Muse Code, Meta's terminal coding agent, on day one; a "max reasoning" variant of the checkpoint is still in safety testing. Teams that route coding workloads through an API should read this as a price war move dressed as a model launch — and the Contributor SKU is the weapon.

A release, and the tier that changes its meaning

Muse Spark 1.3 is the fourth release in the Muse Spark line since April, and the fastest cadence Meta has ever held on a flagship: the original Muse Spark on April 8, Muse Spark 1.1 on July 9 (Meta's first paid API model), the coding-focused Muse Spark 1.2 on August 5, and now 1.3 a month later. Meta kept 1.3's standard pricing identical to 1.2's, which is itself notable — the company is holding API prices flat while claiming a large jump in coding capability. The Contributor tier is where Meta's pricing strategy actually lives: rather than discount the standard SKU, Meta ships the same checkpoint twice, once at full price with no training-data clause and once near cost in exchange for usage rights. That two-SKU structure, first shipped with 1.2 in August, is now the default for the family.

• Checkpoint — Muse Spark 1.3, a closed-weight multimodal reasoning model (text, image, video, audio and PDF in; text out).

• Context — 1,048,576 input tokens, 131,072 output ceiling, unchanged from 1.2.

• Standard SKU — $1.25 in, $4.25 out, $0.15 cached input. No training-data clause.

• Contributor SKU — $0.10 in, $0.20 out, $0.002 cached input. Prompts and completions may be used to train Meta models.

• Reasoning modes — standard modes live at launch; "max reasoning" held back pending safety testing.

• Where it runs — the Meta Model API and Muse Code, with the Contributor SKU also distributed through the platforms Meta has signed for it.

The "same checkpoint, two contracts" structure matters for one practical reason: model quality is not the variable between the SKUs. A developer who pays $4.25 per million output tokens and a developer who pays $0.20 are calling the same weights. The decision between them is entirely about the data clause and the operational limits that come with it — which is why the Contributor option deserves coverage as a product of its own rather than a footnote on the 1.3 launch.

What Meta claims the new checkpoint changes

Because 1.2 was itself a coding release, the 1.3 story is mostly "better at the same job." All of the following are Meta's own numbers from the launch, not yet reproduced by an independent evaluator:

• Terminal-Bench 2.1 — 88.8, which Meta says ties the best result on the board, up from 82.9 it reported for Muse Spark 1.2 on the same benchmark.

• DeepSWE v1.1 — 75.4, which Meta calls a state-of-the-art result on the long-horizon agentic benchmark, ahead of the 73.0 and 74.0 it cites for two frontier rivals.

• Long-context retention — 98.5 on MRCR at 256K–512K and 98.1 at 512K–1M, both vendor-reported.

• Efficiency — Meta's engineers report roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on the same tasks, which is the claim that most directly feeds a cost comparison.

The efficiency claims matter even if the benchmark scores do not survive contact with an independent harness, because they compound with the Contributor price. The one independent anchor available today is the previous generation: on Artificial Analysis, the newest Muse Spark entry is still Muse Spark 1.2, holding an Intelligence Index of 57 with a measured Terminal-Bench 2.1 score around 80.1 — below Meta's own 82.9 for that model. No independent run of Muse Spark 1.3 has been published as of this writing, and Artificial Analysis has not yet listed the 1.3 checkpoint. Treat the 88.8 and 75.4 figures as Meta's claims until an arena or a third-party harness confirms them.

The Contributor fine print is where the deal gets real

The gap between $0.20 and $4.25 output tokens is real, and the terms attached to it are equally real. Meta's Contributor program, as published with the August 1.2 release, pairs the discounted rate with constraints that change how a serious team can use it:

• Data use — prompts and completions sent through the Contributor SKU may be used to train Meta's models. That includes code, and for an agentic coder, effectively the contents of the repositories it is pointed at. This is the entire reason for the discount.

• Rate limits — the Contributor tier runs far tighter than standard: 60 requests per minute and 2.1M tokens per minute, against 3,000 requests and 4M tokens per minute on the standard SKU. That is a 50× gap in request throughput, and it rules out the Contributor tier for bursty CI or large parallel eval runs.

• Region limits — Contributor availability is restricted to a subset of regions, and a payment method is required even though the marginal cost of a session can be cents.

• Add-ons are not discounted — web-search grounding costs $2.50 per 1,000 queries on both tiers, so a retrieval-heavy agent sees a smaller percentage saving than the token rates imply.

• Reasoning bills as output — any reasoning tokens count at the output rate, so the 21× output discount is the number that dominates an agentic workload.

What $0.20 output tokens do to an agent bill

Take a representative agentic-coding task: 200,000 input tokens of context and repository state, and 40,000 output tokens including reasoning and tool calls. On the standard Muse Spark 1.3 SKU that is $1.25 × 0.2 plus $4.25 × 0.04, or about $0.42 a task. On Muse Spark 1.3 Contributor the same arithmetic is $0.10 × 0.2 plus $0.20 × 0.04, or about $0.03 — roughly a fifteenth of the price before the 25%-fewer-tokens efficiency gain is even counted. Scale to a team running a few thousand agentic tasks a month and the Contributor SKU turns a five-figure token bill into a three-figure one. That is the offer. The fine print above is the cost, and for many teams the relevant question is not whether the saving is real but whether the data clause and the 60-requests-per-minute ceiling are acceptable.

How you'd actually run it today

Muse Spark 1.3 Contributor is not yet routed by every platform that carried Muse Spark 1.2, so the honest access picture is a split one. On the Meta side, the Contributor SKU is available directly through the Meta Model API — that is the vendor's own endpoint, and for the data-sharing tier it is where the terms are enforced. On our side, the Muse Spark family's routable member today is the standard Muse Spark 1.2, served on OrcaRouter at Meta's list price of $1.25 in and $4.25 out with 0% markup — one API key, no second contract, and automatic failover if the upstream provider is congested. When a routed provider adds the Muse Spark 1.3 family, both SKUs would appear at list price the same day, because OrcaRouter passes the provider rate through untouched: a vendor price cut, or a Contributor tier, shows up in the invoice immediately rather than after a renegotiation. That pass-through is the practical reason to keep an eye on where 1.3 lands rather than assume the current split is permanent.

The next two weeks

Four things will decide whether Muse Spark 1.3 Contributor is a genuine inflection or a marketing cadence. First, whether an independent harness confirms the 88.8 Terminal-Bench and 75.4 DeepSWE claims, since the whole "frontier at a discount" framing rests on them. Second, when the held-back "max reasoning" variant ships, and whether it is also offered on the Contributor tier. Third, how Meta's rivals respond on price — a serious coder at $0.20 output tokens pressures every premium coding model above it, and the response will show up within weeks. Fourth, whether the community's early skepticism — some reviewers called 1.3 iterative, "squeezing toothpaste" after a month-long gap — gives way to measured third-party results. Until those land, the rational position is that Muse Spark 1.3 Contributor is an unusually cheap way to run a very capable coding model, with the entire discount priced in a data clause that deserves the same scrutiny as the benchmark chart.