Together AI launches Gemma-4-31B-it-Pearl endpoint with 25%+ discounted pricing

• Together AI shipped togethercompute's launch post for Gemma-4-31B-it-Pearl, a serverless endpoint priced at 25%+ below its usual serving economics because Pearl emissions are meant to offset part of inference cost.
• According to simran_s_arora's thread, this is the first endpoint that generates both tokens and a digital asset during use, with the official framing detailed in Together AI's blog post.
• togethercompute's feature post says the endpoint supports 32K context, configurable thinking, function calling, and JSON mode, while the model page lists it as available now.
• The mechanism, per Together's explanation and vipulved's launch note, is Pearl's "proof of useful work" system, which turns matrix multiplications already happening in training and inference into proofs that help secure Pearl Network.
• Together is also using the launch to telegraph a broader rollout: togethercompute says more Pearl-powered endpoints are planned, and customers may eventually receive a portion of the emissions directly.

You can read the official partnership post, inspect the live model page, and the weirdest part is right in Simran Arora's summary: the endpoint is pitched as producing both text output and a subsidy asset in the same flow. Together's launch post ties that subsidy to proofs generated from matrix multiplications already happening during inference. a later Together post adds the more practical details, including 32K context and function calling.

Together's core claim is not just cheaper inference. It is a different source for the discount.

According to Together's launch explanation, Pearl generates proofs from matrix multiplications that already occur during training and inference. Those proofs help secure Pearl Network, and the future value of Pearl emissions is supposed to subsidize endpoint cost.

That is why the announcement keeps describing this as an endpoint that produces both tokens and a digital asset. vipulved's launch note says the discount will move with Pearl spot prices, and Dan Fu's reaction framed the experiment as a new kind of inference tokenomics.

The shipped endpoint is Gemma-4-31B-it-Pearl, an instruction-tuned Gemma 4 31B checkpoint from Pearl Research Labs running on Together AI.

The concrete product details from togethercompute's feature list are straightforward:

• 32K context
• Configurable thinking
• Function calling
• JSON mode
• Serverless availability on Together AI

The launch card in Together's announcement also shows list pricing of $0.28 per 1M input tokens and $0.86 per 1M output tokens.

Together is treating this as the first Pearl-powered endpoint, not a one-off.

In that follow-up, togethercompute says the company plans to expand its Pearl-powered product portfolio and eventually offer customers a direct share of the currency emissions. That customer pass-through detail does not appear in the initial announcement, and it is the clearest sign this launch is also a test of a broader serving business model.