Liquid AI released a new model generation Thursday morning. By the afternoon a 600-million-parameter multimodal variant was running entirely inside a browser tab on WebGPU, no install step. Anthropic shipped Haiku 5.5 the same day and reset the price floor on its cheap tier. Two releases pointing the same direction: the work at the bottom of the stack keeps getting cheaper and keeps moving closer to the client.
The clearest convergence of the week is that the cheap decision tier now has vendors competing for it. OpenAI's Decisions API went live, and Simon Willison had a plugin wrapping it for his llm command line tool at version 0.1a0 within hours. Willison's read on Haiku 5.5 was that the previous Haiku had been badly showing its age against current small models, which is the same observation from the pricing side. Put those beside the open-weight experiments of the last fortnight and a product category exists that did not exist a month ago: small, fast, structured calls that route, approve, classify, and moderate. The operator question is whether the routing layer in your own system is something you build or something you now buy, and that question was not live in September.
The second pattern sits underneath the first. An omni model small enough to run client-side means transcription, image question answering, and intent classification can come off the API bill and out of the privacy surface together, which is a different kind of saving than a cheaper token. Alongside it, a 19,969-parameter language model went up on Hugging Face, 81 kilobytes of weights, still producing coherent short stories. Neither of those is a production system. Together they are a measurement: the capacity narrow tasks actually require is being bounded from below, in public, by people who publish the weights. Liquid AI has published no benchmarks, so quality stays unverified until somebody runs their own evaluations, and that is the honest state of most of what shipped.
Third, the plumbing builders hand-rolled is being claimed by the companies that own the primitives underneath it. Docker published its own agent harness, notable for provenance rather than capability, because agent sandboxing from the people who own the sandbox primitive is a different proposition from container isolation you wrote yourself on a Friday. OpenAI paired GPT-6 with Intelligent UI, its pitch for model-generated interfaces in place of chat transcripts, which puts the tool-call-to-component rendering layer teams have been maintaining into commodity territory. Jensen Huang, talking to Tom's Hardware at CES, put the vendor side of it plainly: "You sell a chip one time, but when you build software, you maintain it forever." The vendors are taking on that maintenance because the layer is worth owning. I have watched this absorption run before. In 1999 every shop I knew maintained its own session handling and its own connection pooling, and then the application servers shipped both, and the teams who had built theirs well got about six months of credit for it.
Fourth, on the local-inference side, the constraint people are engineering around has moved from capacity to bandwidth. A pending llama.cpp patch caches expert-routed weights on the GPU while they live in host memory, which is the hot path for anyone running a large sparse model on consumer cards, and a second patch adds speculative decoding. Meanwhile a salvaged-hardware build is running GLM-5.3-Flash at 384K context and roughly 90 tokens per second on two retired mining cards carrying 64GB of high-bandwidth memory between them. The purchasing logic there is genuinely different. Memory bandwidth bought the long context; the card count did not. Monday's piece argued that local AI is a premises problem, and bandwidth-first is the part of that argument that got cheaper.
The fifth pattern governs the other four. Strata, the engine claiming a 125-billion-parameter sparse model at around 100 tokens per second on a single 4090, finally got put through real coding work by a reviewer who framed the verdict as deliberately unpopular. The coverage around it has shifted from throughput figures to hands-on skepticism, which is usually where an inference engine either earns adoption or stalls. Liquid AI shipped weights without benchmarks. Haiku 5.5 changes a routing decision nobody can actually make without re-running their own evaluations against their own traffic. Every significant release landed ahead of its evidence, and the gating function is whoever gets around to testing it. Tuesday's edition asked who publishes your plan B; the adjacent question is who runs your benchmarks, and for most small teams the answer is still nobody.
The two most informative numbers of the week are both small: 600 million parameters running in a browser tab with no install step, and 19,969 parameters in 81 kilobytes still writing stories. The floor of what a model has to be in order to be useful dropped twice in a day, in public, and Anthropic repriced the tier sitting directly above it the same afternoon.