{"slug": "the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race", "title": "The smartest model lost — and it just redrew the 2026 AI race", "summary": "Claire Vo, founder of ChatPRD and host of the How I AI podcast, conducted a head-to-head comparison between OpenAI's GPT-5.6 lineup and Anthropic's Claude models, finding that the theoretically superior Claude Fable lost to the more collaborative GPT-5.6 Soul. Vo weighted her evaluation 70% on personal taste and 30% on machine judgment, concluding that a model's practical effectiveness and ability to collaborate are as important as raw intelligence. The results highlight a shift in AI model evaluation from pure intelligence to usability and taste.", "body_md": "The most interesting model comparison of 2026 isn't a benchmark table. It's a product exec quietly changing the question everyone asks about models — and getting a completely different ranking as a result.\n\nClaire Vo (founder of ChatPRD, host of the *How I AI* podcast) ran a head-to-head between OpenAI's new GPT-5.6 lineup (Soul / Terra / Luna) and Anthropic's Claude Fable and Sonnet. The result was an upset: **the most theoretically intelligent model, Claude Fable, lost to the one she could actually collaborate with, GPT-5.6 Soul.**\n\nHere's what that upset actually reveals.\n\nTired of vibe-checking, Vo built a real benchmark across the work she does every day: writing PRDs, prototyping apps, debugging multi-step code, and talking to an agent. Scoring had two layers — an LLM-as-judge (she picked the harshest judge, GPT-5.5) and her own hand-graded \"taste test,\" where she clicked through every artifact and wrote notes.\n\nThen the key move: she weighted the final score **70% her taste / 30% the machine.** \"It's my show. I trust my own taste more.\"\n\nThat's the first insight. Benchmarks are getting more rigorous, but the final call is still human taste. The point of blind testing isn't to *replace* taste — it's to force it to be **honest**. Cover the labels, react to the work itself, then put your judgment back at the center.\n\nOn raw intelligence, Fable is elite. But Vo's verdict is the sharpest line on models I've seen this year:\n\nFable is theoretically hyper-intelligent. Soul is practically effective.\n\nShe describes Fable as \"an engineer who has never met a human.\" Precise to the point of pedantry — it scores every risk, hardens every edge. In one case it hardened a tool-calling loop so tightly that **only one specific model could run it at all.** It optimized itself into a corner.\n\nSoul's edge was the opposite: it gets *out* of its own head. Same stuck problem — she moved it to Codex, said \"stop arguing, just do what you think is right,\" and it fixed it in one shot. Not perfect, but shipped.\n\nHer framing is a manager's framing, and it's the heart of the whole review: the hardest colleague isn't the one who's wrong. It's the one who is theoretically brilliant but can't actually get anything done — can't see the forest for the trees, too stuck in their own head.\n\nUnderneath this sits something bigger: **whether you can work with a model is part of the model's capability — not a nice-to-have.**\n\nVo's biggest complaint about Fable was that she \"couldn't talk to it.\" Its output reads \"for agents, by agents\" — nearly unreadable to a human. A model you can't collaborate with fluently has a high collaboration cost, no matter its IQ. Soul's highest praise from her: \"it writes like a normal person.\" (For genuine *personality* in an agent voice, she still reaches for Sonnet.)\n\nWe used to discuss \"intelligence\" and \"usability\" separately. They've merged. A model you can't build alongside doesn't turn into productivity.\n\nVo's design bar is brutal — 50 written reactions, 14 of them flat \"garbage.\" The line she draws is clear. She **rewards** unique, opinionated, functional design. She **punishes** one thing: slop. Gradients, emoji placeholders, em-dashes, the same dark-mode monospace dashboard everyone ships.\n\nShe even named an aesthetic: the \"editorial\" look — beige background, burnt orange, italic serif. Recognizable, and ranked at the bottom. Not because it's ugly, but because it's **too familiar, with no point of view.** (Fun tell: Soul quietly loves a forest green — \"woodland elegance.\" Model aesthetics are becoming fingerprints.)\n\nThe insight: **when generation is nearly free, \"having no point of view\" becomes the defect.** Taste is turning into the ability to *refuse slop.*\n\nGPT-5.6 isn't one model, it's three: **Soul** (frontier), **Terra** (balanced), **Luna** (cheap, high-volume). Pricing is aggressive — Soul at \\$5/\\$30 per million tokens, roughly half of Fable's \\$10/\\$50.\n\nAnd there's no single \"best model,\" only the right horse per task:\n\n| Task | Her pick |\n|---|---|\n| Prototypes |\nSoul — most functional, real point of view |\n| Clean, direct PRDs |\nTerra — \"down to earth\" |\n| Agentic voice (talks like a human) | Sonnet |\n| Complex multi-step debug |\nSonnet (per the LLM judge) |\n\nModel selection is quietly becoming a craft in itself.\n\nThe last thing she showed points at the real shift. She wired GPT-5.6 + Codex to Chrome (`@Chrome`\n\n) and had it burn through ~500 LinkedIn messages, cut a long talk into 5 social clips, fill out forms, test web apps.\n\nThe model stopped being \"give me some text\" and became \"**go do this thing end to end.**\" It's growing into an operator — which rhymes with a broader shift: AI value is sliding from *delivering a tool* to *delivering an outcome.*\n\nPull it together and this review isn't really about \"OpenAI beat Anthropic this round.\" It's about this:\n\n**In 2026, the model race changed axes.** When every frontier model is already smart enough, differentiation moves off \"how smart\" and onto a harder-to-measure cluster: collaborability, practical effectiveness, taste-alignment, and the absence of slop.\n\nAnd don't miss the last detail — the final judge wasn't GPT-5.5 or any LLM. It was a person, weighting her own taste 70/30 over the machine. The stronger models get, the further *forward* human judgment moves.\n\nSo next time you see a benchmark table, maybe ask a different question. Not \"which model scores higher,\" but: **\"which model can I actually build real user value with?\"**\n\n*Based on the How I AI podcast by Claire Vo: \"GPT-5.6 Sol vs. Claude Fable.\"*", "url": "https://wpnews.pro/news/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race", "canonical_source": "https://dev.to/hunter_g_50e2ec233acd07b5/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race-3623", "published_at": "2026-07-09 21:00:57+00:00", "updated_at": "2026-07-09 21:05:24.444417+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-products", "ai-research", "ai-agents"], "entities": ["Claire Vo", "ChatPRD", "OpenAI", "GPT-5.6", "Anthropic", "Claude Fable", "Claude Sonnet", "Codex"], "alternates": {"html": "https://wpnews.pro/news/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race", "markdown": "https://wpnews.pro/news/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race.md", "text": "https://wpnews.pro/news/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race.txt", "jsonld": "https://wpnews.pro/news/the-smartest-model-lost-and-it-just-redrew-the-2026-ai-race.jsonld"}}