Two fast, cheap models aimed at agent workloads arrived on October 19, 2026, and both reached the front page of Hacker News the same day. They take different technical routes to the same claim: that latency and cost per step, not peak intelligence, now limit what agents can do.
Mercury 3: does text diffusion scale
Inception calls Mercury 3 "the most capable diffusion-based LLM released to date". Instead of emitting one token at a time, a diffusion LM refines many positions in parallel, which is where the speed comes from. The company says it has "scaled Mercury more than 40x in total parameters", with benchmark intelligence up 3x from Mercury 1 (mini), and that "The scaling laws that drove progress in autoregressive models hold for diffusion too." Its stated numbers:
- Speed: 1,068 tokens per second on NVIDIA GPUs, described as 4-8x faster than GPT-6 Luna and Claude Haiku 5.5, which it says Mercury 3 matches or beats on quality.
- Context: 1M tokens, up from 260K for Mercury 2.5.
- Price: $0.20 per million input and $0.70 per million output tokens, half that at launch; OpenRouter lists
inception/mercury-3at the $0.10/$0.35 launch rate with a 1M window. - Agent test: on 50 live Mind2Web navigation tasks in Inception's own web-agent harness, Mercury 3 "completed 50% of them, significantly more than GPT-6.1-Sol", using the same token volume and step count.
It is available through Inception's API, OpenRouter, Baseten and Vercel, and is OpenAI-API compatible.
MiniMax M3.1 Flash: the anonymous model revealed
MiniMax says "Space Bunny", which "processed more than 200 trillion tokens while operating anonymously", was an early version of M3.1 Flash. The post claims online inference "nearly three times faster" and agentic usage cost "70% lower than its predecessor", through "more aggressive sparsity", with output speeds of 200+ tokens per second and a P50 of 170+. Pricing is "$0.15 input · $0.60 output · $0.015 cached input" per million tokens. Reported scores: 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82% on CyberGym for vulnerability discovery. The model is on MiniMax's API and coding tools; "Model weights will be released later," as will a larger Pro model. OpenRouter so far lists only a free variant, so the paid price above comes from MiniMax alone.
What remains uncertain. Neither company publishes parameter counts, and every comparison is self-run: Inception's Mind2Web result is 50 tasks in its own harness, and MiniMax does not say which predecessor its 70% saving is measured against. For reference, OpenRouter lists the earlier MiniMax M3 at $0.30/$1.20, double the Flash list price. Mercury 3's quality claim against Haiku 5.5 and GPT-6 Luna has no independent evaluation yet.
Analysis: The interesting split is architectural. MiniMax is squeezing a conventional sparse autoregressive model; Inception is betting that diffusion's parallel decoding can now carry real reasoning, and a 40x scale-up with a 1M context is the first serious test of that bet. For builders, both are candidates for the high-volume inner loop -- tool selection, page navigation, log triage -- with a frontier model reserved for planning and review. Two cautions apply. Diffusion models may behave differently from autoregressive ones on long structured outputs and streaming, so test schema adherence before swapping one in. And a model scoring 82% on CyberGym that will soon ship open weights is a dual-use release worth watching when those weights land.
On October 19 Inception launched Mercury 3, a diffusion LLM it says is 40x larger than Mercury 1, runs at 1,068 tokens/s with a 1M context and costs $0.10/$0.35 per million tokens at launch, and MiniMax launched M3.1 Flash at $0.15/$0.60, revealing that the anonymous 'Space Bunny' model was its early version; all benchmarks are vendor-run and M3.1 Flash weights are promised later.