Google copied Jev's API, matched OpenAI's price, and beat Anthropic to market. The niche research category is now a vendor fight over the small, constant choices agents make thousands of times a day.
The most revealing detail in Google's launch of Gemini Decision 1 is not the model. It is the vocabulary. According to Simon Willison's write-up, the new API is "almost identical to Jev". It supports the same three question types, and it "even reuses Jev's term 'Noul' for a yes/no question."
When the largest company in the field borrows a startup's jargon, the category has stopped being an experiment. It has become a standard.
That changes the framing most people still carry, that decision models are a curiosity for researchers studying classifier-style AI. Latent Space called November 11 "D-Day", and the label fits. On a single conference day, Google launched a production decision model, beating Anthropic to it. Cloudflare showed which parts of a build failure a classifier could handle. TypeSafe AI paraded production customers.
For anyone running an agent, this matters in a specific way. A lot of what an agent does all day is not writing. It is choosing. Is this email urgent? Which tool fits this request? Does this pull request pass the checklist? Those small calls now have three competing vendors, a price floor measured in fractions of a cent, and a shared interface. The question is no longer whether to delegate them. It is to whom, and at what point on the autonomy dial.
Google copying Jev's API is the clearest sign the category has standardized
Start with what Gemini Decision 1 actually is. Per Willison's API notes, it answers three kinds of questions:
- Yes/no (Jev's "Noul")
- Choice between options
- Rubric scoring against defined criteria
That is the whole surface. No essays, no chat, no open-ended generation. A decision model takes a question and some context and returns a constrained answer.
What stands out is that Google did not invent its own shape for this. It matched Jev's question types and Jev's terminology. In most software markets, that kind of convergence is what happens just before a layer gets treated as interchangeable. Think of how cloud storage APIs drifted toward one familiar pattern, so that switching providers became a configuration change rather than a rewrite. The pattern here resembles that moment.
Meanwhile, the Latent Space recap of the day frames Google's move as "a BIG surprise" and notes that Sundar and Logan announced it personally. That is not how a company ships a side project. Putting the CEO on a narrow, classifier-style product signals that Google sees decision-making as a slot in the agent stack worth owning.
For agent users, standardization is good news. If three vendors accept the same kinds of questions, your agent's decision step stops being locked to one provider. The skill that asks "should I escalate this ticket?" can, in principle, route that question to whichever model is cheapest or most accurate this month. On a Wardley map, this is the move from custom-built toward product, and the next stop on that axis is commodity.
Free output and sub-dime input make delegation almost costless
Pricing is where the market shape gets sharp. Willison reports that Gemini Decision 1 is priced the same as OpenAI's GPT-6 Luna decision model: 10 cents per million input tokens, and output is free. Jev remains the cheapest of the three at 4.2 cents per million.
Output being free is the telling part. A decision model's answer is tiny: a yes, a no, a pick from a list, a score. Charging nothing for it tells you vendors expect the cost of these products to live almost entirely in the context you send them. The bill scales with how much your agent reads before it decides, not with what it says back.
That has a practical consequence for anyone watching their agent costs. Routing, triage, and approval checks are among the most frequent things an agent does. If each one used to go through a full general-purpose model, every small "should I?" carried the price of a big "write me this." Moving those calls to a decision model at 4.2 to 10 cents per million input tokens turns them into rounding errors on a monthly bill.
There is a less obvious effect too. When a choice is nearly free, agents can afford to ask more of them. A workflow that once checked one gate before taking an action can check five: is this safe, is this in scope, is this the right tool, does it meet the rubric, does a human need to see it. Cheap decisions do not just lower costs. They make more careful agents economically viable.
Jev's lower price, set against two giants matched at 10 cents, also sets up a familiar fight. The incumbent-style vendors are competing on breadth and brand. The specialist is competing on cost. Which of those wins depends on the next point: what kinds of inputs each model can actually decide about.
Google's wedge is multimodality, not price
Google did not undercut anyone on price. It matched OpenAI and stayed above Jev. Its differentiation is what the model can look at. According to Willison's comparison, the three products split cleanly:
- Jev: text and JSON only
- OpenAI's GPT-6 Luna decision model: text and images
- Gemini Decision 1: fully multimodal, starting with text and extending beyond images
This is a sensible wedge for Google. Its general models have long been pitched as natively multimodal, and a decision model that can judge a screenshot, a chart, or a recording opens use cases the text-only option cannot touch. Picture an agent that watches a dashboard and decides whether a spike is worth waking someone up for, or one that reviews a product photo against a listing rubric.
For agent users, this splits the buying decision in two. If your agent's choices are mostly about text (emails, tickets, documents, structured data), Jev's text-and-JSON scope covers the job and costs less than half as much per input token. If your agent needs to judge what it sees, the choice narrows to OpenAI or Google, and Google covers more input types at the same price as OpenAI.
The pattern resembles classic market segmentation more than a winner-take-all race. A specialist owns the high-volume, text-only tier on cost. The large platforms compete for the richer, more expensive inputs where their broader model families carry an advantage. Whether that holds depends on whether Jev expands beyond text, or whether the giants decide to cut price to squeeze the low end.
Decision models let agents move up the autonomy dial one narrow choice at a time
The most useful signal from AI Engineer was not Google's launch. It was Cloudflare. The Latent Space recap notes that Cloudflare "showed which parts of a build failure a classifier could handle." The phrasing matters: which parts. Not the whole failure. Not "the agent fixes builds now." Specific sub-decisions inside a messy process.
This maps directly onto what we call the Autonomy Spectrum: agent deployments sit somewhere between copilot and full autonomy, and most failures come from deploying at the wrong point. General-purpose agents often fail because they are handed a whole job at full autonomy when only some of its steps are actually reliable.
Decision models offer a different path. Break the job into choices. Find the ones a classifier handles well, like "is this a flaky test or a real regression?" or "which team owns this failure?" Hand those to a cheap, constrained model. Keep the open-ended work (writing the fix, deciding whether to ship) closer to the copilot end, with a human in the loop.
Meanwhile, TypeSafe AI used the same stage to show off production customers, and the recap describes a speaker's return to talk about "the Jevolution" alongside notable production customers, in a context that reportedly involves crossing $100M ARR. Read together, these suggest the category has moved past demos into real workloads with real revenue behind them.
For someone configuring an agent, the takeaway is practical. The best near-term use of decision models is not making your agent more autonomous across the board. It is making specific gates more reliable, so you can grant autonomy to exactly the steps that earned it.
Meanwhile, Google's general model pricing shows why a stable decision layer matters
One day after Gemini Decision 1, Google shipped something else. Willison reports that Google launched Gemini 4 Hydrogen, "the lowest cost member of the Gemini 4 family", and that it is "actually not too far off the price of its much bigger siblings, Carbon and Argon." He calls the pricing "really confusing."
The details back him up. Carbon and Argon are the same price, and both are scheduled to double on December 1st. Google says Argon "delivers Gemini 4 intelligence with faster responses and lower token spend per turn compared to Gemini 4 Carbon", so the same sticker price can produce different real bills depending on how many tokens each turn burns.
Put these two launches side by side and a contrast appears. Google's general-purpose lineup is getting harder to reason about: overlapping tiers, a coming price jump, and cost that depends on per-turn behavior. Its decision model has one simple price, matched to a competitor, with free output. The pattern resembles a company keeping its premium layer complex and flexible while making the narrow layer cheap and predictable.
That is the logic of Commoditize Your Complement. If decisions become nearly free and interchangeable, more agent activity flows toward the general models where Google can still charge, and where prices can double on a date it chooses. The decision layer becomes the cheap on-ramp.
For agent users, the lesson is to separate the two kinds of spending. Your agent's frequent small choices can sit on a stable, cheap decision model. Its heavier reasoning and writing will ride the more volatile general-model pricing. Knowing which calls fall in which bucket is now a real cost lever, especially with a December price change on the calendar.
The harness, not the model, will decide who benefits from this price war
If three vendors offer near-identical decision APIs at cents per million tokens, the models themselves start to look like supply. That brings us to The Harness Hypothesis: the value in AI is not in the model but in the harness that connects the model to the world.
The early evidence of harnesses absorbing Google is already in the plumbing. Langfuse's v4.80.0 release adds Google AI Studio as a gateway provider. That is the kind of quiet change that makes switching vendors routine. A gateway sits between your agent and the model vendors, so the agent asks one place and the gateway decides where the request goes. Willison's own command-line tool shipped an llm-gemini release the day Hydrogen launched, another sign the tooling layer moves as fast as the vendors do.
Here is where Aggregation Theory comes in. Whichever layer owns the user relationship captures the value, and in agent land that is usually the harness: OpenClaw, Claude Managed Agents, or whatever runtime you configure daily. When decision models are interchangeable, the harness that routes each question to the cheapest adequate model, logs the result, and lets you audit the choice gets to treat Jev, OpenAI, and Google as competing suppliers.
The absence of Anthropic sharpens this. The Latent Space recap explicitly frames Google as having "beat Anthropic to launching a production decision model", which implies Anthropic was expected to arrive. If it does, there will be four suppliers for one standardized interface. For anthropic agent infrastructure in particular, the question will be whether a decision model is bundled into Claude Managed Agents as a native advantage or offered as one more interchangeable endpoint.
Either way, the direction of travel favors users. The vendor fight compresses the price of choosing. The harness turns that compression into cheaper, more careful agents. And the one thing worth watching is whether your harness lets you pick the supplier, or quietly picks for you.
/Figures
| Model | Input price (per million tokens) | Output price | Inputs supported |
|---|---|---|---|
| Jev | 4.2 cents | Not specified | Text and JSON |
| OpenAI GPT-6 Luna decision model | 10 cents | Free | Text and images |
| Gemini Decision 1 | 10 cents | Free | Fully multimodal |
/Sources
/Key Takeaways
- Gemini Decision 1 copies Jev's three question types and even its 'Noul' terminology, a strong sign decision models have standardized into a shared interface.
- Pricing has converged: Google and OpenAI's GPT-6 Luna charge 10 cents per million input tokens with free output, while Jev undercuts both at 4.2 cents.
- Google's differentiator is full multimodality; Jev is text and JSON only, and OpenAI handles text and images.
- The best near-term use is hardening specific agent gates (triage, routing, approvals), as Cloudflare's build-failure demo suggests, rather than granting blanket autonomy.
- Gemini 4's confusing tiers and a December 1 price doubling for Carbon and Argon make a cheap, stable decision layer more valuable by contrast.
- As decision models commoditize, the harness that routes between suppliers captures the value; check whether yours lets you choose.


