The price war nobody announced
Four labs are within a couple of points of each other on the same coding benchmark, and one of them has stopped pretending capability is the differentiator. That admission is the story.
Muse Spark 1.2 scores 82.9 percent on Terminal-Bench 2.1 inside Muse Code — ahead of GPT-5.6 Terra and Grok 4.5, behind Opus 5 at max effort in Claude Code. Meta says openly that it is differentiating on price rather than capability.
What a two-point spread means
At the top of a saturating benchmark, a two-point gap is inside the noise of harness configuration, effort settings and prompt scaffolding. Meta's own framing concedes this by comparing against a competitor at max effort, which is a way of saying the comparison depends on how hard you let the other model try.
When four organisations spending tens of billions between them arrive at the same score, the score has stopped discriminating. It has become a qualification threshold — necessary to be in the conversation, insufficient to win it.
Where the differentiation moved
Into the harness. Muse Spark 1.2 was trained alongside Muse Code, and Meta argues the coupling itself is a source of coding performance. Muse Code launches its own sub-agents to work across large repositories in parallel, which is a systems design decision, not a model capability.
And into context economics. Poolside's Laguna S 2.1 mixes sliding-window and global attention at 3:1 and quantises the KV cache to FP8 to make a natively trained million-token window affordable. That is not a smarter model. It is a cheaper one at the length that matters.
The uncomfortable part for everyone
A market that competes on price with parity capability is a commodity market, and commodity markets do not support the capital structures currently financing frontier labs. The training run that produced parity cost the same as the one that would have produced a lead.
Meta can afford that outcome — it has other businesses. The question worth watching for the rest of the year is which labs cannot, and what they do about it.
VentureBeat — Meta enters the AI coding wars with Muse Spark 1.2 and Muse Code → · CNBC — Meta debuts Muse Code to take on Anthropic and OpenAI → · Hugging Face — poolside/Laguna-S-2.1-FP8 model card →