Zuckerberg calls Meta (META) Muse Spark 1.3 "almost too cheap to meter," but Meta keeps prices unchanged

What the scorecard says, and what it leaves out

Meta published an 11-row benchmark card comparing Muse Spark 1.3 at its maximum reasoning setting against its own Muse Spark 1.2, OpenAI's GPT 5.6 Sol and Anthropic's Opus 5, both also at max. The numbers are Meta's, not a third party's, and one of the eleven tests is an internal instruction-following index that nobody outside Meta can run.

Read across the card, Muse Spark 1.3 takes four of the eleven rows outright and ties a fifth. All of them are coding or long context. It scores 75.4 on DeepSWE v1.1, a long-horizon agentic coding test, against 73.0 for GPT 5.6 Sol and 74.0 for Opus 5. On SWEAtlas CodeBase QnA, a codebase-understanding test, it posts 59.4 against 53.5 and 52.7. On the two MRCR long-context tests, it scores 98.5 and 98.1 where GPT 5.6 Sol scores 91.5 and 73.8. Anthropic's column shows a dash for both long-context rows, so the two biggest wins on the card have no Anthropic number to be compared with.

The agent rows go the other way. Opus 5 leads on four of the six: GDPVal-AA v2 at 1824 to Meta's 1754, JobBench at 65.7 to 64.9, OSWorld 2.0 at 68.3 to 66.9 and AutomationBench at 50.3 to 49.4. GPT 5.6 Sol leads the other two, DeepSearchQA at 93.0 to 89.4 and Meta's internal instruction-following index at 60.5 to 57.8. On Terminal-Bench 2.1, the agentic terminal coding test that Meta led with when it launched Muse Code in August, Meta and OpenAI tie at 88.8, with Anthropic at 86.7.

So the honest summary is parity, with an edge in the areas Meta trained hardest on. Outside the long-context tests, Meta is generally within a few points of the leaders. The jump that Zuckerberg called the biggest so far is real against Meta's own previous model: 1.2 to 1.3 adds more than 32 points on the longest MRCR test and 20 points on DeepSWE. On the six agent rows, the differences against the competition are one to three points either way.

Meta also says the new model does the same work with less. The release post says 1.3 uses "~20% fewer tool calls and ~25% fewer tokens" than 1.2 on coding tasks, asks clarifying questions more often, and confirms before taking consequential actions. For an API billed per token, using 25% fewer tokens can reduce the effective cost per completed task even when the rate card stays unchanged.

The rate card that did not change

The rate card itself did not move. Meta's developer pricing page lists the standard muse-spark-1.3 model at $1.25 per million input tokens, $0.15 for cached input and $4.25 per million output tokens, the same prices Meta set for 1.1 on July 9 and kept for 1.2 on August 5. The context window stays at one million tokens.

The "too cheap to meter" line points at the second SKU. A contributor tier, muse-spark-1.3-contributor, costs $0.10 per million input tokens and $0.20 per million output, about 12 times cheaper on input and 21 times cheaper on output. The catch is on the same page: prompts and outputs sent through the contributor tier are "used to improve our products." Developers are paying for the discount with their data. That tier also existed for 1.2, so the discount is not new either. What is new is that a model trading blows with Opus 5 on Meta's own card is now available at those prices.

The part of Wednesday's release that reaches beyond Meta is the price. OpenAI and Anthropic sell their top models at a multiple of Meta's standard rate, and Meta has now held that rate flat through two upgrades in eight weeks while closing the score gap. A frontier-class model priced like a mid-tier one puts pressure on what the two leaders can charge for API access, and open weights, when they come, will push the floor lower still.

Watermelon and the compute bill

The emoji in Zuckerberg's post refers to Meta's next flagship, which employees call Watermelon. Muse Spark's internal name was Avocado. Business Insider reported in July that Alexandr Wang, Meta's chief AI officer, told an internal town hall that Watermelon, which was still training at the time, had caught up with OpenAI's GPT-5.5 on closely followed benchmarks, without naming which ones. The same report said Watermelon uses an order of magnitude more compute than Muse Spark. Meta's release post on Wednesday promises "bigger models, the Muse Spark open weights release, and more."

An order of magnitude more compute is the line that connects this product story to Meta's capital spending. Meta has been the most aggressive of the hyperscalers on data center commitments this year, and the case for that spend rests on shipping models that people will pay for. Wednesday's release is evidence that the money is producing a competitive model. It is not yet evidence that the model produces money, because Meta has priced it to win share rather than margin.

Meta rolled out Muse Spark 1.1 on July 9, 1.2 and Muse Code on August 5, and 1.3 on September 2. If the four-week cadence holds, the open-weights release and Watermelon are the next two releases to watch.