--- title: "Decision models" is just classifiers with a new name tag description: A Twitter/X debate erupted over Perplexity's new open-weights model being labeled a 'decision model.' Harrison Chase praised it for cheap, typed-answer tasks... canonical: https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: "Decision models" is just classifiers with a new name tag | daily.dev og:description: A Twitter/X debate erupted over Perplexity's new open-weights model being labeled a 'decision model.' Harrison Chase praised it for cheap, typed-answer tasks... og:url: https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr og:image: https://api.daily.dev/og/posts/q4lbic5Pr.png og:image:alt: "Decision models" is just classifiers with a new name tag og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # "Decision models" is just classifiers with a new name tag **[Trends](https://daily.dev/sources/trends)** · 3 min read · 1 upvotes · 0 comments ## Summary A Twitter/X debate erupted over Perplexity's new open-weights model being labeled a 'decision model.' Harrison Chase praised it for cheap, typed-answer tasks like routing and approvals inside agent harnesses. Antirez countered that calling classifiers 'decision models' misrepresents them, since they lack the chain-of-thought reasoning of LLMs and 'decide a lot less.' Sayak/Merve dryly noted this is really just zero-shot classification rebranded. The piece sides with antirez on naming but agrees with Chase that cheap classifiers are useful for low-stakes routing tasks, not high-stakes approvals. ## Content OpenAI's Decisions API just hit public beta, and within days the category went from a TypeSafe curiosity called Jev to a crowded shelf. Perplexity, Cloudflare, Amazon, Microsoft and Liquid AI all have one. The argument is already about the name. ## The naming fight @antirez doesn't hold back: "Calling classifiers 'decision models' is a crime against machine learning." His case is that these models decide *less* than LLMs. They are either small LLMs that aren't allowed to think, with the last-position logits mapped to classes, or BERT-style encoders that compress the input and project a label. No chain of thought, no serial processing. @mervenoyann is less angry but just as dry: "rebranding of zero-shot classification models as decision models wasn't on my 2026 bingo card, but happy people finally discovered them." ## The product, in one line The pitch is a model that skips text and returns a number or a category. OpenAI says its API, running on GPT-6 Luna at $0.10 per 1M input tokens, is up to 10x faster than going through the Responses API. @hwchase17 frames it as "cheap typed answers for the small calls inside a harness (routing, approvals, judging), big model for the rest." ## Who's winning, depending on who you ask Everyone, according to their own benchmarks. - Microsoft says Microsoft-Decision-1, post-trained on Qwen3.5-9B, had the highest accuracy across 36 benchmarks and is 35x faster than GPT-6 Sol. - Perplexity claims pplx-decider-v1-27b beats Jev overall across 11 benchmarks, even though it lost six of them individually. - Liquid AI says d1-3B tops the Decision Index among sub-10B models, with edge latency under 50ms. - Cloudflare cut Clef-flash to $0.038 per M input tokens, now cheaper than Jev, though the hosted context window shrank from 64k to 24k. The open-weights camp has a real structural edge: Perplexity, Cloudflare and Amazon ship weights you can self-host, while OpenAI's is hosted only. ## Practitioners @ThePrimeagen ran OpenAI's model against omarchy's QA results with the same prompt and image. It was "generally much faster" and "much more accurate at finding things to click based on screen shots." @simonw shipped an LLM plugin for it within days and called it a "Jev-clone." The DIY crowd is louder. @jaredpalmer says a $200 Kev-4B fine-tune can "mog every other decision model" and published a skill to do it on Modal. Unsloth says you can train one on 3GB VRAM, and a Qwen3.5 0.8B run went from 37% to 65% accuracy in 60 steps. The awkward question for the vendors: if a $200 fine-tune is competitive, what exactly are the hosted APIs selling? ## Questions this post answers ### What is a decision model in the context of AI agents? A decision model is a rebranded term for small classifier models used inside agent harnesses to make cheap, fast, typed decisions like routing, approvals, and judging, instead of invoking a large LLM for every call. These models are either small LLMs restricted to a single forward pass using last-token logits, or BERT-style encoders that project input meaning onto a class, both lacking chain-of-thought reasoning. _Developers weighing cheap classifiers against reasoning models for agent routing can follow this debate on daily.dev._ ### Why do some argue that decision models decide less than LLMs? Critics argue decision models decide less than LLMs because they skip chain-of-thought reasoning entirely, relying on a single forward pass rather than serial, step-by-step processing. Antirez called labeling classifiers as 'decision models' a 'crime against machine learning' since they compress input and pick a class without the deliberation an LLM performs before answering. _Teams deciding how much reasoning to delegate to lightweight models can track this discussion on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 19 comments across x (as of 2026-10-10). **TL;DR:** Reaction to Microsoft's small decision/scoring model built on Qwen3.5-9B is mostly curious and cautiously interested, with recurring worries about calibration and what happens when it's confidently wrong. **Sentiment:** 20% positive · 55% mixed · 25% skeptical **The case for** - Single-pass scoring on a small 9B model is seen as a cheap, fast fit for gating every step in agent workflows. - Some see this pattern of small post-trained models for one job as the practical direction for automation. - The low per-token cost is highlighted as making it feasible to run a scorer on every agent step. **The pushback** - Several worry that a fast but confidently wrong score is worse than no score at all, especially in incident response. - Questions about calibration versus purpose-built decision models suggest skepticism about reliability. - Some note the notable choice of building on Chinese open-weight Qwen rather than OpenAI's own models undercuts the narrative of it being a proprietary breakthrough. **By community** - x (mixed): Replies mix genuine curiosity about latency and cost gains with recurring concern about calibration and wrong-but-confident decisions. **Hottest debate:** Whether a fast single-pass scorer is actually trustworthy enough for high-stakes gating versus just being a cheap latency win. **Open questions** - Does the 35x speed advantage hold once the model is rebased onto MAI or OpenAI's own base models? - How well-calibrated is the model's confidence compared to purpose-built decision models? - What's the real median latency on non-trivial decision trees, rather than headline benchmark numbers? **Highlights** > @testingcatalog A Qwen3.5-9B post-train is interesting: in our test, a decision model retrofitted from an LLM gave probabilities very differently from purpose-built ones. Hope someone runs Decision-1 through calibration checks: > — [0xAlphana on x](https://x.com/0xAlphana/status/2108656104884572245) > @testingcatalog Single-pass scoring on a 9B base is the interesting part. Cheap enough to run on every output as a gate. Curious how well calibrated it is, though. A judge that's confident and wrong is worse than no judge. > — [TestedAITools on x](https://x.com/TestedAITools/status/2108660323599131094) > @testingcatalog Agent loops hit a "proceed, retry, or escalate" gate on nearly every step, and today each one costs a full LLM call. A single-pass scorer at $0.042 per million input tokens makes it cheap enough to gate every step. > — [DoDataThings on x](https://x.com/DoDataThings/status/2108686461499478143) > @testingcatalog Single-pass scoring makes the decision fast, but in incident response a fast wrong call costs more than the score. What's the guardrail when it picks the wrong action? > — [kashif\_manzer on x](https://x.com/kashif_manzer/status/2108654954642895227) > @testingcatalog useful bit isn't decision-making vs LLMs - it's the latency gap. when an agent picks which of 12 outreach variants to ship, shaving the choose step to sub-second means 200+ live experiments/week. what's the real median latency on a non-trivial decision tree, not the p99 number? > — [faraz0z on x](https://x.com/faraz0z/status/2108693813002649628) **Source threads** - [x](https://x.com/danielhanchen/status/2108569476186902823) · 0 points · 0 comments - [x](https://x.com/testingcatalog/status/2108647491033821668) · 1 points · 19 comments ## Similar posts on daily.dev - [As AI agents grow, vendors carve out decision-making as a separate model layer](https://daily.dev/posts/as-ai-agents-grow-vendors-carve-out-decision-making-as-a-separate-model-layer-s16prsxwi) · InfoWorld · 0 upvotes · 0 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#classification](https://daily.dev/tags/classification), [#perplexity](https://daily.dev/tags/perplexity) [View this post on daily.dev](https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"\"Decision models\" is just classifiers with a new name tag","url":"https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr"},"datePublished":"2026-10-03T05:40:34.685Z","dateModified":"2026-10-10T00:00:05.654Z","description":"A Twitter/X debate erupted over Perplexity's new open-weights model being labeled a 'decision model.' Harrison Chase praised it for cheap, typed-answer tasks...","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,classification,perplexity","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"\"Decision models\" is just classifiers with a new name tag"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/decision-models-is-just-classifiers-with-a-new-name-tag-q4lbic5pr#faq","mainEntity":[{"@type":"Question","name":"What is a decision model in the context of AI agents?","acceptedAnswer":{"@type":"Answer","text":"A decision model is a rebranded term for small classifier models used inside agent harnesses to make cheap, fast, typed decisions like routing, approvals, and judging, instead of invoking a large LLM for every call. These models are either small LLMs restricted to a single forward pass using last-token logits, or BERT-style encoders that project input meaning onto a class, both lacking chain-of-thought reasoning. Developers weighing cheap classifiers against reasoning models for agent routing can follow this debate on daily.dev."}},{"@type":"Question","name":"Why do some argue that decision models decide less than LLMs?","acceptedAnswer":{"@type":"Answer","text":"Critics argue decision models decide less than LLMs because they skip chain-of-thought reasoning entirely, relying on a single forward pass rather than serial, step-by-step processing. Antirez called labeling classifiers as 'decision models' a 'crime against machine learning' since they compress input and pick a class without the deliberation an LLM performs before answering. Teams deciding how much reasoning to delegate to lightweight models can track this discussion on daily.dev."}}]} ```