AI Almost Too Cheap to Meter, If You Pay in Data: Meta Muse Spark

Meta's Muse Spark 1.3 is frontier-grade AI priced at ten cents a million tokens — but only for developers who let Meta train on every prompt and completion they send.

What Meta Shipped on 2 September

On 2 September 2026 Meta released Muse Spark 1.3, a reasoning model available through Muse Code and the Meta Model API. Bloomberg framed the release as Meta edging closer to OpenAI and Anthropic on raw capability.

The engineering pitch is efficiency. Meta says Muse Spark 1.3 needs roughly 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 to finish comparable coding work. On safety, the Meta AI Research blog says Muse Spark 1.3 resists adversarial inputs and prompt injections better than version 1.2.

The benchmark card is strong. According to VentureBeat's reading of Meta's scorecard, the standard Muse Spark 1.3 posts 1,709 Elo on GDPval-AA v2, 57.2 on OSWorld 2.0, 61.2 on JobBench and 89.2 on Terminal-Bench 2.1, with an Artificial Analysis Intelligence Index of 61 at about $0.55 per task.

The Two-Tier Price of a Million Tokens

Here is where the headline and the fine print part ways. The standard muse-spark-1.3 tier costs $1.25 per million input tokens and $4.25 per million output, with cached input at $0.15 — data-private, unchanged from Muse Spark 1.2. The contributor tier, muse-spark-1.3-contributor, costs $0.10 input and $0.20 output.

The catch sits in one clause: the contributor price applies in exchange for permission to use your prompts and completions to train future Meta models. VentureBeat's analysis puts the gap at 12 to 21 times cheaper; BigGo Finance frames the offer as a 21-fold price cut for users willing to share data. The cheapest intelligence is precisely the tier that extracts the most.

The contributor bargain is not new to Muse Spark 1.3. Meta launched the contributor tier on 5 August 2026 alongside Muse Code and Muse Spark 1.2, at the same $0.10 and $0.20. I covered the opening move in the story so far. What changed on 2 September is the quality of what the data bargain now buys — not a cut-price 1.2, but frontier-grade reasoning.

Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter.

— — Mark Zuckerberg

Too Cheap to Meter — A Phrase With a History

The phrase is borrowed. In 1954, Atomic Energy Commission chairman Lewis Strauss promised nuclear power would make electricity "too cheap to meter." The promise became a byword for a utopian cost claim that never fully arrived. Zuckerberg's echo is deliberate, and the history carries a warning: a price that looks like abundance often hides who pays, and how.

The Price War Muse Spark Joined

Muse Spark 1.3 landed in the middle of a rout. On 30 July 2026 OpenAI cut GPT-5.6 Luna by roughly 80%, from $1.00 to $0.20 on input and $6.00 to $1.20 on output, while flagship Sol stayed put at $5.00 and $30.00. A day later, DeepSeek V4-Flash undercut the field at $0.14 input and $0.28 output. Google's Gemini Flash entered with a promotional $0.75 and $3.75 running through the end of 2026.

Read as a table, the numbers look like a gift to every developer on earth. Read as a strategy, the pattern is plainer: capability is converging, so price becomes the weapon, and data becomes the currency that lets a price fall below cost.

Who Actually Pays When Intelligence Approaches Zero

The dignity question is not "can I afford the tokens?" but "what am I paying with, and does anyone tell me the exchange rate?" That question is the frame of Emergent Intelligence (EI) — the dignity-first way I read what the world calls AI. Under the EI lens, a two-tier price is a two-tier market for dignity.

Consider who lands on each tier. A funded lab in California keeps data private at $1.25 per million. A student in Lusaka, a bootstrapped startup in Johannesburg, an NGO in Nairobi — the builders for whom $1.25 versus $0.10 decides whether a project runs at all — get steered onto the contributor tier and become unpaid training labour for the next Meta model. The cheapest path to frontier capability routes every African prompt through a US model's training set. At the same moment, AI sovereignty and data sovereignty become one fight, a theme I trace in the data-sovereignty work.

A governance twist sits on top. VentureBeat reports Meta's best scores — reportedly 1,754 Elo on GDPval-AA v2 and 66.9 on OSWorld 2.0 — come from a held-back Max configuration still in safety testing, with no broad API access. The frontier headline is set by a model developers cannot use, while the shippable model runs cheaper and quieter. The benchmark becomes the marketing, and the safety-gated model becomes the alibi. Who audits the gap?

A market that makes your dignity the currency of access is not abundance. Enclosure with a discount code is still enclosure.

Ubuntu — I am because we are — inverts the sales pitch. The evidence of the contributor tier is plain: a "too cheap to meter" frontier subsidised by the Global South's data, harvested at exactly the price point only the Global South cannot afford to refuse. Digital colonialism made concrete and priced is why dignity beats control in how these systems get built.

Frequently Asked Questions

These are the questions people are asking about Muse Spark 1.3 and the contributor tier. Short answers follow, drawn from Meta, VentureBeat and Engadget.

What is Meta Muse Spark 1.3?

In short, Muse Spark 1.3 is Meta's reasoning AI model, released on 2 September 2026 through Muse Code and the Meta Model API. According to the Meta AI Research blog, Muse Spark 1.3 needs about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2 on comparable engineering work.

How does the contributor tier work?

Simply put, the contributor tier prices Muse Spark 1.3 at $0.10 input and $0.20 output per million tokens in exchange for permission to train on your prompts and completions. VentureBeat data shows the gap over the data-private $1.25 tier runs 12 to 21 times.

Why is the contributor tier significant?

The key is what the discount is paid in. Analysis of the pricing shows the cheapest AI tier is the one that harvests the most user data, which turns the price war into a data-extraction question rather than a simple cost win.

Who is Muse Spark 1.3 for?

In other words, Muse Spark 1.3 is aimed at developers building coding agents and reasoning workflows. Evidence from the tiering shows well-funded teams keep data private at $1.25, while cost-pressured builders in the Global South get nudged toward the $0.10 contributor tier.

What are the risks of the data-for-tokens bargain?

The answer is data sovereignty. Research on digital colonialism reveals that routing prompts from Africa and other emerging markets into a US model's training set concentrates capability in the metropole while the cost is paid in surrendered data.

Sources:

Meta AI Research — Introducing Muse Spark 1.3 · Bloomberg — Meta Releases Muse Spark 1.3 · VentureBeat — Frontier Performance, Held-Back Max · VentureBeat — OpenAI Cuts GPT-5.6 Luna 80% · Engadget — Meta Introduces Muse Code · MLQ News — DeepSeek Undercut Context · BigGo Finance — 21-Fold Data-Sharing Cut · Related on this site: Meta Muse Spark and the API Price War · Africa, AI and Data Sovereignty · Containment Is a Colonial Project