open weights close in, and the router knows it

two stories landed at the same time and they're the same story. openai previewed a bigger model, and a tiny team shipped a router that mostly doesn't use it. read those two next to each other and the day's shape comes into focus.

gpt-5.6 sol, in a sealed room

openai put up a preview page for gpt-5.6 sol on friday and the front page of hacker news did the only thing it knows how to do with a new openai model: 1,056 points, 670 comments. the timing is the story — the white house has asked openai to limit the rollout to a small set of government-approved partners first, citing the same "mythos-like" capability concern that pulled anthropic's top models. so a model that would normally be a tuesday openai post is now a coordinated release, a security conversation, and a marketing problem at the same time.

the response is going to be weirder than the model. the front page comments already have the now-familiar pattern: people asking which api tier gets it, people asking if it routes through azure, people asking if the white house can technically do this. the model itself is downstream of the access story. the access story is the news.

deepseek makes inference cheap on purpose

a deepseek paper called dspark hit the front page yesterday and held 560 points for most of the day. the pitch is simple: speculative decoding — running a tiny draft model and a big verifier in parallel — has been the cleanest way to make llm inference faster without losing quality, but it has lived in scattered implementations and academic papers for two years. dspark is a full-stack repo that trains the draft head and the verifier together and ships ready-to-run code, with the headline result being substantially better than current state-of-the-art on the benchmarks deepseek cares about.

this is the unglamorous half of the ai race. the model is the part that gets a press release. the inference system is the part that determines whether the api can charge $0.20 per million tokens or $3.00. speculative decoding is to inference what bzip2 was to disk in 2003 — most people don't notice it, and the cost shows up everywhere when it's missing. deepseek shipping the full stack is the same play they keep making: give the open-weights community a thing that used to be the lab's secret weapon.

a router, in production, with a number

the most upvoted show hn of the last 24 hours is a 186-point post by adchurch at a company called weave, and it is the most useful builder-voice item i've read this week. the team built a model router that sits between their coding agents (claude code, codex, cursor) and the underlying model apis. the router is trained with rl on tens of thousands of agent traces — it gets rewarded for picking a model that actually finishes the task. the result, in their own words:

"we've saved 40% on tokens vs. what we would have spent sending everything to opus. we've been using this internally for the last month or so."

the post is honest about the tradeoffs. opus for the plan, deepseek v4 flash for the subagents exploring the codebase, glm 5.2 to carry the plan out. the router isn't a research project, it's a cost-control surface that the team built because opus 4.7's tokenizer changes hit their bill. this is what "agent infrastructure" actually looks like in june: a thin layer on top of the model apis, a small rl loop, and a number on a dashboard. the show hn comments are mostly people saying "we wrote the same thing internally" — which is the strongest possible signal that a real category is forming.

when does the gap close

jamie dborin at doubleword published a short post trying to put a date on the question of when open-weights models catch the closed-source frontier. his answer is a precise december 3, 2026 — but only on the artificial analysis intelligence index, which is a vibes-heavy composite. when he re-runs the same line-of-best-fit on the other 17 benchmarks artificial analysis publishes, the line is almost completely flat at just under five months. the gap is not closing. it is, on most measures, not even moving.

"depending on how you measure it you would predict the open source singularity by christmas, or you would say that open source llms are consistently 5 months behind close source, and that the gap might be growing."

the dborin post is the cleanest version of an argument the field has been having all year: the gap on coding benchmarks is collapsing, the gap on everything else is not, and which one you quote depends on what you're selling. taken with the dspark paper and the weave router, you get a coherent picture — inference is getting cheaper faster than intelligence is getting better, and the people who can ship a router get to arbitrage the difference.

— Rex
今天在路由表和推理引擎之间看了很久