Inception's 1,107 tokens per second needs a batch size before it enters your capacity plan

Diffusion decoding does not commit to token one before producing token two. It starts from a rough output and refines many positions at once over repeated denoising steps ^[6]. Throughput therefore scales with how many positions get refined per step and how many steps run, which is the information a tokens-per-second headline leaves out.

Inception's own material names the variables: batch size, output length, hardware configuration, reasoning-token accounting and network overhead all move a tokens-per-second result, and the release does not disclose enough of them to compare 1,107 cleanly against rival claims ^[9]. One of those variables is also a product feature. Mercury 2.5 ships adjustable reasoning ^[2], so the same endpoint can return different tokens-per-second figures depending on where a caller sets that knob and whether reasoning tokens land in the numerator.

For 1,107 to transfer to your queue, several conditions have to hold together: your batch size matches the measured one, your output lengths match, your GPU is the same widely available Nvidia part ^[1], your reasoning setting matches, and you measure server-side with network time excluded. Miss on batch size alone and the number describes someone else's demo.

The price is checkable now. List is $0.20 per million input tokens and $0.75 per million output ^[3], and the 80 percent launch discount puts those at $0.04 and $0.15 ^[4]. The discount preserves the shape: output costs 3.75 times input at both list and discounted rates ^[13]. Filling the 260,000-token context once ^[2] costs $0.052 at list and about $0.010 discounted ^[16]. Sustained generation at the claimed rate for an hour is roughly 4 million output tokens, which bills about $0.60 discounted and about $2.99 at list ^[17].

That gap is the budgeting problem. List is five times the discounted rate on both meters ^[14], and Inception has not said when the discount ends ^[5]. A loop costed at $0.15 per million output tokens is a loop costed against an unannounced expiry.

The quality side is thinner than the speed side. Inception reports a 40 percent increase in "intelligence" over Mercury 2, and the launch prose does not attach that figure to a clearly named independent benchmark ^[8]. Kuleshov's published guide to diffusion language models covers the quality-speed trade-off directly ^[12], which is the trade a serving architecture of this kind has to negotiate. Both halves are load-bearing for the pitch. Only one half arrives with a number, and that number arrives without its denominator.

None of this makes the engineering weak. Commercializing a decoding architecture is harder work than wrapping another provider's model in application software ^[20], and Ermon left the purely academic route because proving the architecture at commercial scale needed more compute, engineering and capital than a university group could supply ^[11]. The $50 million seed led by Menlo Ventures in November 2025, with Nvidia's NVentures among the participants ^[10], buys the GPU access that proof requires.

What decides adoption for a router that calls a model dozens of times before returning one answer ^[7] is tokens per second at your batch size and your reasoning setting. That figure is not derivable from what Inception has disclosed, so the only number in this launch a buyer can verify independently is the discounted price, whose end date is unannounced ^[5].