Friday’s agent-siege dispatch was about models that wander off the reservation. This morning’s useful news is about models that are not allowed to wander at all.

TypeSafe’s Jev does not write. It takes program state in and returns a Choice, a Score, or a yes/no probability. That is a different product than another 70B chat model with a prettier system prompt, and Vercel teams treated it like one.

A decision model, not a smaller LLM

TypeSafe’s launch post is from founder Diogo Almeida, who worked on the instruction-following stack that became ChatGPT and then spent two years building something that refuses to emit a sentence. The company calls the class System One — Kahneman’s fast lane, named for Thinking, Fast and Slow, with the model itself named after William Stanley Jevons. Efficiency drops, demand for intelligence goes up. That is the bet.

The interface is the point. You declare the questions. Jev evaluates them in parallel and hands back typed values plus calibrated probabilities. No parser. No “please respond in JSON.” Schema match is guaranteed, which is how they get to put a literal zero on the hallucination chart — not because the model is wiser, because it is not allowed to invent a fourth option.

Published numbers, all vendor-side unless noted:

Claim Figure Caveat
Latency 70–500 ms end to end TypeSafe, West Coast laptops
Price 0.042 per million input tokens, output free Founder says they cannot prove it is not subsidized
Workflow eval up to 193.6x faster, 444.6x cheaper vs LLMs Their four workflows; reference is the average of GPT-6 Astra and Fable 5.1
Vercel AI Gateway, first 24 hours ~13% of paid teams Vercel: 2x the GPT-5.6 family, 6x Fable 5.1

Vercel shipped it on AI Gateway on September 16 through the AI SDK evaluate API. Within three days Cloudflare, LangChain, and Langfuse had wired it. The Gateway listing is promotional-free through September 25, which is part of why 13 percent happened that fast. Guillermo Rauch posting that free Jev was his weekend plan is the other half.

The honest read: this is the missing if statement for agent loops. Pick the next tool. Decide whether to retry, ask the human, or stop. Score a patch before you open the PR. You still need a chat model to write the code. You do not need one to decide whether the code is done.

TypeSafe’s own FAQ is careful about the rest. Jev is not open weights. Cardinality tops out at 255 choices. Images are not in yet. And the 193x number is their home-field workflow eval, not SWE-bench. Treat it like a new primitive, not a replacement for Astra.

The other harness: freeze the world, mutate the wrapper

Neon maze walls mutating around a frozen glass cube in a dark control room

If Jev is a harness for the model, Google Cloud AI Research’s EnvHarness is a harness for the environment. VentureBeat covered the drop over the weekend. Apache 2.0, code on GitHub, paper at arXiv 2608.19880.

The trick is almost rude in how little it asks of the simulator. You keep the trusted environment and its verifier. EnvHarness sits in front and changes the experience:

  • Stage — move the starting state (put the mug in a closed drawer)
  • Contract — filter actions or observations (no submitting the patch until tests run)
  • Chain — glue tasks into a longer episode

EnvRigger watches rollouts, diagnoses the failure pattern, writes a wrapper, and checks that the new task is still solvable. On SWE-bench Verified, average trajectory length fell from 55.01 steps to 49.61. Held-out gains went as high as nine points across ALFWorld, WebArena, OfficeQA, and SpreadsheetBench.

That is the same lesson as Jev, pointed at the other end of the loop. Stop rebuilding the world every time the agent gets better. Wrap it. Keep the grader. Make the agent practice the thing it actually skips.

It only works where rollouts are cheap and state restores — coding sandboxes, browser evals, office sims. Do not point EnvRigger at production. Google says that out loud.

Qwen-Image-2.1: RGBA without the extra model

Close-up cartoon GPU die with stacked transparent holographic layers peeling off the silicon

Alibaba’s Qwen team open-weighted Qwen-Image-2.1 on September 20. Visual generator is 7B (32 single-stream DiT layers). The headline is native RGBA — transparent PNGs out of the diffusion process, not a background-removal pass glued on after. Up to ten reference images in one edit. Day-zero ComfyUI templates. Diffusers pipeline is QwenImage21Pipeline.

It fits on a consumer RTX 3090-class card if you offload. GGUF quants showed up inside a day.

The catch is the license: Qwen Research License, not Apache 2.0. Weights you can download. Product you probably cannot ship. That is the recurring open-multimodal deal, and it has not gotten more honest just because the alpha channel is real this time.

If you are compositing stickers, product cutouts, or in-image type on a transparent layer, this is the first 7B checkpoint that does the job without a second model. If you need a commercial license, you are still on a closed API.

Capital still prints, California still studies the off switch

Nscale, the London AI cloud Nvidia likes to call a UK national champion, filed an S-1 on September 18 to list on the NYSE. The numbers are the usual CoreWeave-class cartoon:

  • H1 2026 revenue 140.6 million, up 1,252% from 10.4 million
  • Net loss 1.02 billion
  • Active-plus-contracted TCV 103.4 billion as of August 31, on 2.6 billion of currently active contracts
  • Anthropic services agreements up to about 44.6 billion
  • 3.1 billion of convertibles, 1 billion of that from Nvidia
  • ~25,000 active GPUs, 1.37 GW active and contracted, line of sight to ~10 GW

Take-or-pay backlog is not revenue. A 1,252% growth rate off a 10 million base is not a mature cloud. It is, however, an S-1 — the filing the last few weeks of rumor were waiting on — and it puts the August Anthropic campus deal in a prospectus instead of a leak.

Separately, California Governor Gavin Newsom signed Executive Order N-9-26 on September 18. Headlines said “kill switch.” The order itself asks GovOps, by November 16, for recommendations on whether state law should require one, plus onsite independent verifiers in frontier labs and a loss-of-control category in incident reporting. No trigger. No who-pulls-it. No penalty. A study memo with a date.

And on the ads side, researcher buchodi documented ChatGPT’s __obi cookie: SameSite=None, one-year, bound to the ChatGPT account, sent from advertiser sites that installed OpenAI’s measurement pixel. 936 pixels, 1,029 hostnames in the capture set. OpenAI’s cookie policy files it under analytics. Someone who allows analytics and refuses marketing still gets it. OpenAI Support acknowledged the mail and answered neither question. This is standard adtech. It is new on a product people confess to.

What actually changed this morning

The chat model is no longer the whole stack. Jev is a typed if you can call in 100 milliseconds. EnvHarness is a way to keep training the agent after it memorizes the benchmark. Qwen-Image-2.1 is a 7B local compositor with a research-only leash. Nscale’s S-1 is the compute bill arriving in public-market clothes.

If you run agents, the practical move this week is not another system prompt. It is splitting the loop: generate with the expensive model, decide with something that cannot invent a fourth button.

Want this in your inbox every morning? Subscribe to the SpaghettiStories newsletter.

Some links may be affiliate links. If you’re buying hardware to run local models, this affiliate link helps keep the lights on.