LLMs Think Too Much for Jobs That Don't Require Thinking

Ask a frontier model to choose between two options and it does exactly what it was built to do: generate a chain of reasoning to arrive at one bit of information. It will usually be correct. It will also spend up to two seconds before producing a single token, and bill for every token of thinking on the way there. The question teams ask is whether the model is smart enough for the job. It is. That is the wrong question.

TypeSafe AI just gave that question a name. The company's launch of Jev introduces what it calls System One models, a new model class built explicitly around Daniel Kahneman's distinction between fast, automatic cognition and slow, deliberate reasoning. It is a bold metaphor to build a product category on, and borrowing it correctly requires knowing what Kahneman actually meant, not the flattened version that shows up in conference slides.

What Kahneman Actually Meant by System 1 and System 2

Daniel Kahneman did not invent the terms System 1 and System 2. Psychologists Keith Stanovich and Richard West coined them in their 2000 paper on individual differences in reasoning, describing two distinct modes of cognitive processing. Kahneman adopted the terminology and, through Thinking, Fast and Slow, made it the shorthand an entire industry now reaches for when it wants to sound rigorous about AI architecture. Most of that industry has never read the distinction closely enough to use it correctly.

System 1 is fast, automatic, and effortless. It is pattern-matching against prior experience, not deliberation. You do not consciously reason your way to recognizing a face or catching a thrown ball. System 2 is slow, deliberate, and effortful, the mode you engage for multiplying two three-digit numbers or evaluating a legal contract. Neither system is a brain region. Both are metaphors for what neuroscientists call implicit versus explicit cognition, and the distinction between them was itself grounded in Tversky and Kahneman's 1974 paper on the heuristics people use under uncertainty, work that predates the System 1/System 2 label by 26 years.

The reason this matters for AI infrastructure is not metaphorical decoration. It is architectural. A decision that is bounded, repeatable, and pattern-matched against precedent is a System 1 job. A decision that requires generating a novel chain of reasoning across unfamiliar territory is a System 2 job. Software has spent two years building exactly one kind of model, and routing System 1 work through it anyway.

Why Every LLM Is a System 2 Architecture, Even the Small Ones

Here is the distinction most vendor comparisons miss: a smaller, faster LLM is not a System 1 model. It is System 2 with a shorter chain. Autoregressive generation, predicting the next token conditioned on every token before it, is deliberation by construction. Making that chain shorter or running it on faster silicon does not change what kind of cognitive act is happening. It is still sequential construction of a novel string, one token at a time, and every token depends on the ones before it, which is exactly why it cannot be parallelized away.

This shows up directly in the numbers. Frontier reasoning models sit at 0.5 to 2 seconds of time-to-first-token, the delay before generation even starts, and specialized low-latency providers like Groq's custom LPU silicon get that floor down to roughly 700 milliseconds, still generating text sequentially after that point, just faster hardware doing the same sequential job. None of that latency comes from the model being under-optimized. It comes from the architecture requiring a full pass through the network for every single token of output, and time-to-first-token is only the floor, the total climbs further once the model has to generate more than a few tokens of reasoning.

A faster LLM is not a different kind of thinking. It is the same deliberation, timed with a stopwatch instead of a calendar.

TypeSafe AI's announcement of Jev, the first model in what it calls the "System One" class, is explicit about targeting the other half of the problem: decisions with a fixed answer space, evaluated fast enough to sit inline in a workflow rather than alongside one. The company claims 70 to 500 millisecond end-to-end response times and structured, typed outputs with calibrated confidence scores, generated through parallel sampling rather than sequential token generation. Those are vendor claims, not independently reproduced benchmarks, and the difference matters, which is a point this article returns to.

That floor matters because a routing or classification call never stops at the first token, it has to finish generating before a downstream system can act on it. TypeSafe's own announcement characterizes typical LLM workflow latencies at 3 to 329 seconds end-to-end, against Jev's claimed 70 to 500 milliseconds for the same kind of decision. That comparison comes entirely from the vendor making the claim, and it measures a different thing than the time-to-first-token chart above, but the direction of both numbers agrees: the wait starts before generation and gets worse from there.

What Jev Actually Changes, Mechanically

Strip away the launch-post language and three concrete architectural claims remain. First, a training method TypeSafe calls Reinforcement Learning for Calibrated Decisions, aimed specifically at making the model's stated confidence track its actual accuracy, rather than optimizing purely for the reward signal that produces confident-sounding text regardless of correctness. Second, parallel sampling instead of sequential generation, which is the mechanical reason the latency claims are even plausible. Third, output constrained to typed structured values rather than strings that require a parsing and validation step afterward.

That third point is not cosmetic. A large body of recent work on structured output has found that format constraints and content accuracy are not the same problem, and conflating them is where a lot of production incidents originate.

The price gap tracks the architecture gap. Frontier models charge for every token of the reasoning they generate on the way to an answer, System One's pricing charges for a single typed value instead. TypeSafe prices Jev's input tokens at $0.042 per million, against $0.30 per million for Google's cheapest current frontier model and $5 per million for Anthropic's flagship. That is a vendor-set price for an early-access product being compared against published, current rack rates for established ones, not an apples-to-apples benchmark, but the gap is wide enough that even a generous margin of error leaves it intact.

The structured-output research bears this out with numbers worth sitting with. A 2026 benchmark study found that LLMs constrained to a JSON schema hit structural validity, meaning the output actually parses as valid JSON matching the schema, above 96 percent of the time. But value accuracy, meaning the content inside that valid structure was actually correct, ranged from 0.693 to 0.830 across models, and the rate of getting a fully correct response collapsed to as low as 0.376. The model can format a wrong answer as flawlessly as a right one, and a downstream system checking only for schema compliance will never catch the difference.

The Reasoning Tax Nobody Line-Items

This is the pattern, not an isolated misjudgment. A routing or classification step built on a reasoning LLM keeps working, accuracy stays fine, and nothing breaks. The cost shows up as a rounding error, not an incident: a token-spend line that never gets questioned because it isn't the biggest one on the invoice, and a response-time SLA that's marginally worse than it should be for reasons nobody traces back to one architectural choice. A System 2 pipeline running a System 1 job doesn't fail loudly enough to get caught.

Gartner's June 2025 research predicts that more than 40 percent of agentic AI projects will be canceled before the end of 2027, attributing the failures to escalating costs and unclear business value, not model capability. Separately, MIT's Project NANDA found that 95 percent of generative AI pilots at surveyed companies failed to reach production or deliver measurable P&L impact. These are different studies measuring different things, agentic project cancellation versus pilot-to-production conversion, but they rhyme: the failure mode is economic and architectural, not a shortage of model intelligence.

Gartner: more than 40% of agentic AI projects will be canceled by the end of 2027. MIT Project NANDA: 95% of generative AI pilots failed to reach production or deliver measurable return. Different studies, different populations, different definitions of failure, neither one a verdict on model intelligence.

Where This Actually Fits in a Real Stack

The honest framing is not "replace your LLM." Most of what a production AI system does still needs one: drafting, summarizing, reasoning across ambiguous context, handling a case nobody anticipated. The question worth asking per decision point in a pipeline is narrower: does this specific step have a fixed, enumerable set of possible answers, and does it happen often enough that shaving two seconds and four cents off each instance compounds into something real.

The two questions that actually decide it

If the answer space is open-ended, a System 1 model cannot help, because it was never built to generate novel content. If the answer space is fixed but the decision runs once a week, the latency and cost savings are real but marginal, not worth a new vendor integration. The case for a System One model shows up specifically at the intersection: bounded answers, high volume.

Routing, classification, intent detection, entity extraction, fraud scoring, content moderation triage: all bounded, all high-volume, all currently running through System 2 architectures in most production stacks because until this year there was no serious System 1 alternative to reach for. Every one of those categories shows up inside the workflow types TypeSafe explicitly targets with Jev, and every one of them is a step, not a whole pipeline, which is the distinction the next section depends on.

What This Doesn't Solve

Jev is an early-access product from a single vendor, and every performance number in TypeSafe's own announcement is self-reported against reference models the company chose. The company's post acknowledges this directly, noting its evaluations weren't independently constructed and some claims need long-term validation. That acknowledgment should be taken at face value and applied by the reader too: nothing here has been reproduced by a third party yet.

The category is also narrower than the launch framing suggests. A model built for calibrated decisions over a fixed answer space cannot draft a memo, cannot reason through a contract clause it hasn't seen the shape of before, and cannot handle the genuinely novel case that shows up in every real production system eventually. System 1 architectures fail exactly where System 1 cognition fails in people: confidently, on the unfamiliar case that looks like a familiar one. That is precisely why the calibration claim matters more than the speed claim. A fast wrong answer with a confident-sounding score is worse than a slow one, and the entire value of the category collapses if the calibration doesn't hold under real traffic.

This is also a research question with genuine open findings on the constrained-decoding side. The same structured-output research cited above found that forcing outputs into a rigid format can itself degrade a model's underlying reasoning quality. Whether TypeSafe's architecture avoids that tradeoff or simply moved it somewhere less visible is not yet answerable from outside the company.

The teams that will get real value out of this category are not the ones replacing their LLM. They are the ones who can look at a pipeline and say, with precision, which three steps in it were never reasoning tasks to begin with.

Sources

1. TypeSafe AI - Introducing System One Models and Jev (2026) - the product announcement this article evaluates; latency, cost, and calibration claims are self-reported by the vendor.
2. Stanovich & West - Individual Differences in Reasoning: Implications for the Rationality Debate (2000) - the paper that coined "System 1" and "System 2" before Kahneman popularized the terms.
3. Tversky & Kahneman - Judgment Under Uncertainty: Heuristics and Biases, Science (1974) - the foundational research on heuristic-driven judgment underlying the later System 1/2 framework.
4. Geng et al. - JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models, arXiv (2025) - benchmarks constrained decoding across ~10,000 real-world JSON schemas.
5. A Multi-Source Benchmark for Evaluating Structured Output Quality in Large Language Models, arXiv (2026) - the source of the structural-validity-versus-value-accuracy gap cited in this article.
6. Mind the Confidence Gap: Overconfidence, Calibration, and Distractor Effects in Large Language Models, arXiv (2025) - documents the gap between LLM verbalized confidence and actual accuracy.
7. Gartner - Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (2025) - attributes cancellations to cost and unclear value, not model capability.
8. MIT Project NANDA - The GenAI Divide: State of AI in Business 2025 - 95% of generative AI pilots failed to reach production or deliver measurable ROI, based on 52 executive interviews and 153 leader surveys.
9. Anthropic - Claude Opus 5 API Pricing (2026) - $5/MTok input, $25/MTok output, used as the frontier-model cost anchor.
10. Google - Gemini API Pricing (2026) - Gemini 3.5 Flash-Lite at $0.30/MTok input, used as the cheapest frontier-model cost anchor.
11. Inworld AI - Fastest LLM Inference APIs in 2026: TTFT and Throughput Guide - time-to-first-token benchmarks across inference providers.