Can System One models like Jev help choose better visuals?

When we were building our Tesserae (now Liqa Labs) presentation generator in early 2025, we invested a couple of months in researching LLMs as classifiers. At the time, LLMs were poor at creating McKinsey-grade slides, in large part because they struggled with spatial reasoning and with drawing creative business visualisations.

We therefore built our own logic for those two tasks, and let the LLM take on two slide-building tasks it was good at: determining which business illustrations are best suited for the user’s prompt (then drawing these illustrations deterministically with our logic), and populating the illustrations Tesserae drew with summarised text content. The bet seemed to pay off: the slides this Tesserae + mainstream lab hybrid model could produce were better - as per user feedback and usage metrics - than what ChatGPT/Claude could produce out of the box throughout 2025 and Q1 of 2026.

Today, slide building is all but a solved problem by mainstream LLMs. However, for a slight tradeoff in visual flexibility, this “classifier-deterministic drawing-populate text” approach allows us to get good results with smaller, faster, and cheaper models. The dollar-cost-per-slide of this approach is 10x less than building a slide from scratch in Sonnet 5 and 25x less than Opus 5. As such, we still have a small number of larger customers using Tesserae’s old hybrid approach. As an important bonus, given we set up each customer’s brand template in advance for the deterministic drawing step, the slides this approach produces are consistently in the customer’s branding.

This is why we were particularly excited when Jev came out. The cheaper LLMs were working well as classifiers, but there was clear scope to drive cost and latency down for that step that LLMs didn’t seem inherently built for.

We benchmarked Jev against four lightweight LLMs on the key classification step of our slide-building approach: picking the business illustration that best fits a slide. Every model saw the user’s prompt for the slide and the same descriptions of 35 illustration types (bar chart, process, timeline, quotes and so on) as part of the system prompt, and had to pick one. We ran 80 labelled cases, with a roughly 50/50 split between synthetic prompts and anonymised prompts from real users - three times each. A pick scored 1 if it matched our label, 0.5 if it was an acceptable alternative, and 0 otherwise. The scores in the table below are averaged over the three runs. Each LLM ran at its two lowest reasoning settings, with its answer restricted to the 35 options.

Two examples:

• “Share what customers told us in interviews: the CFO of Acme said ‘We closed the books two days faster’, the Head of FP&A at Globex said ‘Finally, one source of truth’...” - expected: quotes.
• “Can you make me a chart that shows how long it is light and how long it is dark during a winter day in Denmark?” - expected: pie chart, with a stacked bar, stacked column or column chart also acceptable.

What stood out:

• Jev was the fastest and, with Claude Haiku 5.5, the most accurate (compared to non-system one models). Its median latency was under half the fastest LLM’s, and it tied Claude Haiku 5.5 for the top score overall, though Haiku 5.5 did better on real user prompts (81.7% against 79.4%). The first part was expected, but the second came as a surprise - we were expecting a noticeable tradeoff between cost and accuracy vs LLMs.
  • Intuitive, non-verified explanation: Jev is built and advertised as a “System One” model (following Daniel Kahneman’s split between the brain’s “intuitive and fast” System 1 and “deliberate and slow” System 2 in his well-known book Thinking, Fast and Slow). Picking the right type of visual to convey a complex message is arguably an intuitive choice.
• Jev was not the cheapest model. With prompt caching, gpt-6-luna was 3-4x cheaper than Jev, though both cost well under a dollar per 1,000 calls. In addition, the accuracy difference was minimal. Without caching, gpt-6-luna would cost about $0.26 per 1,000 calls, twice Jev’s price. Claude Haiku 5.5 matched Jev on accuracy at about half its price ($0.07 against $0.13 per 1,000 calls). Against Claude Haiku 4.5 the gap is large: Jev was 24-40x cheaper, 4-19x faster and more accurate.
• Jev was the most consistent. Its answer changed between runs on 2 of the 80 cases, against 3-11 for the LLMs (7-11 excluding Haiku 5.5).
• Jev’s confidence is usable. Alongside each pick, Jev returns a probability for every one of the 35 options. When it was right, its pick’s probability averaged 88%, against 54% when it was wrong. If we had accepted either of its top two picks, it would have scored 97.7%. So a low-confidence pick could trigger a second option or a fallback to a larger model.
• More reasoning barely helped for most LLMs, except for Haiku 5.5. gpt-6-luna scored almost the same at both settings, and Gemini did worse with more thinking. Claude Haiku 4.5 gained 4 points from thinking, at 5x the latency. Claude Haiku 5.5 gained almost 8 points from adaptive thinking, with no change in median latency.
• Real user prompts are where models differ. Every model scored 93-99% on synthetic prompts, but only 67-82% on real ones. The misses were often shared: the word “pipeline” pulled almost every model, Jev included, towards a funnel.

A week after we published this piece, OpenAI released the Decisions API, widely reported as its answer to Jev. We ran it through the same benchmark on 7 October 2026, under the same conditions as the other models, and added it to the table above.

What stood out:

• Accuracy is a tie with Jev, but the misses differ. On real user prompts, Jev picked our exact label more often (74% of the time, against 67%), but Decisions was outright wrong half as often (7%, against 16%), settling for an acceptable alternative instead.
• It was the most consistent model we tested. It gave the same answer in all three runs for every one of the 80 cases.
• Its confidence is about as usable as Jev’s. When it was right, its pick’s probability averaged 87%, against 46% when it was wrong. Accepting either of its top two picks would have scored 94.4%, against Jev’s 97.7%.
• Price is where it loses. Decisions bills input only, at $0.10 per million tokens, but gives no discount for cached input. Our 35 descriptions, about 2,500 tokens, are resent with every call, so it costs about the same as gpt-6-luna without caching: twice Jev’s price, and about 7x gpt-6-luna with caching.

In late 2024, the major labs released a feature that, in our opinion, was under-advertised: structured outputs. Structured outputs allow us to define a strict schema that the model response always adheres to. One can easily see how we can use structured outputs to turn an LLM into a classifier, but we can in fact use structured outputs to define much richer output schemas.

This came particularly handy for the slide-building approach above, because in addition to simply picking the right illustrations for a prompt, we could:

• Define a tree of rules for the output, per client (e.g. only use numbered bullets or icon-shaped bullets for a list, never normal bullets) with minimal tokens, that would always be respected
• Populate the content of each classified illustration in the same pass
• Restrict certain classifier options in certain cases (e.g. a column chart can have max 20 data points)

However, we always struggled when using structured outputs:

• The major labs do document them, but in a cursory manner, and we often encountered undocumented errors when using them.
  • Example: At one point our schema became too big for some models, and the error message didn’t mention this. The max schema size was buried deep in deprecated documentation.
• They output in JSON, which is token-inefficient compared to newer proposed paradigms like TOON.
• Subfeatures of structured outputs sometimes disappear, or at least don’t get documented, with the release of newer models.

OpenAI, Anthropic, and Google all present the blockers above to different extents.

Given the negligible discrepancy in accuracy between Jev and cheaper LLMs, and the minor cost difference, we’ll still continue to use small LLMs with structured outputs for this case. However, the race is on: System One models will continue to get better, faster, and cheaper, while major labs don’t seem to be investing much in structured outputs. We believe this presents a clear opportunity for the major labs.