--- title: Open-weight models are eating proprietary AI's lunch,... description: Enterprise AI spending is shifting sharply toward open-weight models as companies optimize for cost. Uber, Pinterest, and AT&T report cost reductions of 52-80%... canonical: https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Open-weight models are eating proprietary AI's lunch, and the numbers are brutal | daily.dev og:description: Enterprise AI spending is shifting sharply toward open-weight models as companies optimize for cost. Uber, Pinterest, and AT&T report cost reductions of 52-80%... og:url: https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp og:image: https://api.daily.dev/og/posts/YR6Y50msp.png og:image:alt: Open-weight models are eating proprietary AI's lunch, and the numbers are brutal 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. # Open-weight models are eating proprietary AI's lunch, and the numbers are brutal **[Trends](https://daily.dev/sources/trends)** · 3 min read · 26 upvotes · 11 comments ## Summary Enterprise AI spending is shifting sharply toward open-weight models as companies optimize for cost. Uber, Pinterest, and AT&T report cost reductions of 52-80% by routing away from frontier proprietary models with minimal quality loss. OpenRouter data shows proprietary models' share of routed queries falling from roughly 60% to 25% within months. Anthropic's Claude Opus 5 is singled out as dramatically more expensive than alternatives like GPT-5.6 Luna and DeepSeek. Cloud providers benefit either way since they host inference regardless of which model wins, while OpenAI and Anthropic face eroding loyalty and a possible geopolitical risk of cost-sensitive users migrating to Chinese open-weight models. ## Content The numbers are getting hard to ignore. Uber cut AI session costs 52% by routing away from frontier models. AT&T cut bills 56% with LiteLLM routing, with quality dropping only 2%. Pinterest got cost per transaction under 8% of comparable closed models by fine-tuning open weights on proprietary data. These aren't scrappy startups optimizing for survival — they're large enterprises making deliberate architectural choices. The setup: Mozilla's 91-page report puts open-weight models roughly four months behind the frontier on capability. OpenRouter's data tells the rest of the story. Eight of its ten most-used models by August token volume were open-weight. Seven were Chinese-built. DeepSeek became the first open model to lead the platform in weekly requests. Qwen hit 942 million Hugging Face downloads — more than the next eight organizations combined. Here's the uncomfortable math for the closed-model incumbents: open models took ~20% of OpenRouter usage but only ~4% of model-layer revenue, because closed models cost roughly 6x more at ~90% capability parity. That gap is exactly what routing exploits. TrueFoundry ran 550 graded prompts through a Haiku/Sonnet/Opus router versus Claude Opus alone — 69% cheaper, 98% quality retained. On production workloads, 80% cheaper. The routing logic is simple enough to feel obvious in retrospect: classify each request as simple, medium, or complex, then send it to the appropriately sized model. Most requests are simple. You don't need Opus to fetch a stored PDF. One German B2B firm went further and pulled inference entirely in-house — three RTX PRO 6000 Blackwell GPUs running a quantized Qwen MoE model via vLLM, processing 69 billion tokens over 30 days. Cost: ~1,500 EUR/month versus an estimated 13,300 EUR/month at API rates. The caveat they're honest about: this only pencils out with sustained, non-spiky background load. Spiky workloads still favor APIs. OpenAI sits at ~$40B ARR, Anthropic at ~$65B — roughly 10x all Chinese AI model revenue combined, per Rhodium research. But that gap exists partly because Chinese models are mostly open-source, generating downloads instead of API calls. The question is whether enterprise routing behavior starts closing it. If AT&T's trajectory is representative, the routine enterprise usage that currently funds frontier model development is quietly migrating somewhere cheaper. Claude's specific problem, as one post put it bluntly: Haiku is already stretched, and Opus 5 is reportedly 100x more expensive than GPT-5.6 Luna or DeepSeek. There's no cheap tier doing the boring work. That's not a capability problem. It's a product gap that routing infrastructure is filling for them, just not in their favor. ## Questions this post answers ### How much did companies save by switching from proprietary AI models to open-weight alternatives? Reported enterprise savings from switching to open-weight models range widely: Uber cut cost per AI session by 52% by routing away from frontier models, Pinterest got cost per transaction under 8% of comparable closed-model pricing, and AT&T reported up to 80% savings using open-weight alternatives via LiteLLM routing, with only about a 2% quality drop. _Teams weighing model costs against quality tradeoffs can track real-world routing benchmarks like these on daily.dev._ ### How has the market share of proprietary AI models changed relative to open-weight models in routed queries? Proprietary models' share of routed queries on OpenRouter dropped from roughly 60% to 25% within a few months, as routing systems increasingly send requests to the cheapest capable model, which is often an open-weight option rather than a frontier proprietary model. _Anyone deciding which models to route traffic to can follow shifting adoption trends like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@ristotoldsep** · 6 upvotes > Proprietary models had better enjoy those margins while they last. Open weights brought a calculator. **@devgenx** · 5 upvotes > The best thing for the progress of LLMs will be the crash and burn of Anthropic and OpenAI. **@andruhudz** · 1 upvotes > Hope that in a nearest future everyone could have personal offline LLM model which would easily solve routine tasks **@ray73864** · 0 upvotes > Since getting my Strix Halo (MinisForum MS-S1 Max), and ever since the release of Qwen3.8-Flash-Next and what the community is doing with their Strix-Halo fork of llama.cpp, I've been pushing more and more of my stuff to local. > > Sure it might not be fast compared to the likes of ChatGPT and Claude, but compared to session usage on Claude, it's free (not including the obvious incurred costs of the hardware, etc...). > > > I've gone from 100% with Claude (and buying more usage credits because I kept getting dangerously close to my weekly limit before rollover) to around 70% Claude, 30% local. > > As I... **@toddyholiday** · 0 upvotes > We have AGI at home: [https://mulgadc.com/blog/we-have-agi-at-home](https://mulgadc.com/blog/we-have-agi-at-home) ## Similar posts on daily.dev - [Open Weights Is All You Need](https://daily.dev/posts/open-weights-is-all-you-need-8gjhbnwkm) · Kilo Blog · 1 upvotes · 0 comments - [Chinese AI Models Went From 2% to 45% of Global Developer Traffic in 12 Months. Nobody in Silicon Valley Is Talking About It.](https://daily.dev/posts/chinese-ai-models-went-from-2-to-45-of-global-developer-traffic-in-12-months-nobody-in-silicon-va-vs6otv2ry) · Medium · 0 upvotes · 0 comments --- Tags: [#open-source](https://daily.dev/tags/open-source), [#claude](https://daily.dev/tags/claude), [#deepseek](https://daily.dev/tags/deepseek), [#ai-gateway](https://daily.dev/tags/ai-gateway) [View this post on daily.dev](https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp) ```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":"Open-weight models are eating proprietary AI's lunch, and the numbers are brutal","url":"https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp"},"datePublished":"2026-09-11T06:56:56.932Z","dateModified":"2026-09-18T06:55:29.407Z","description":"Enterprise AI spending is shifting sharply toward open-weight models as companies optimize for cost. Uber, Pinterest, and AT&T report cost reductions of 52-80%...","image":"https://pbs.twimg.com/media/HR6rooObwAAvbuX.png","thumbnailUrl":"https://pbs.twimg.com/media/HR6rooObwAAvbuX.png","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":11,"discussionUrl":"https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":26},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":11}],"keywords":"open-source,claude,deepseek,ai-gateway","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":"Open-weight models are eating proprietary AI's lunch, and the numbers are brutal"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp","comment":[{"@type":"Comment","text":"Proprietary models had better enjoy those margins while they last. Open weights brought a calculator.","datePublished":"2026-09-11T09:04:12.917Z","url":"https://daily.dev/posts/YR6Y50msp#c-Jke4ZKX6c","author":{"@type":"Person","name":"Risto Tõldsep","url":"https://daily.dev/ristotoldsep","image":"https://lh3.googleusercontent.com/a/ACg8ocLDWc6mZn0JwNmXXw6WY0L_HJ6pegzRttooC5VgtXESHSj3MNYx=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":6}},{"@type":"Comment","text":"The best thing for the progress of LLMs will be the crash and burn of Anthropic and OpenAI.","datePublished":"2026-09-11T17:49:54.494Z","url":"https://daily.dev/posts/YR6Y50msp#c-qShgy41qb","author":{"@type":"Person","name":"Mike","url":"https://daily.dev/devgenx","image":"https://media.daily.dev/image/upload/s--PrWOWtSq--/f_auto/v1789486214/avatars/avatar_ytCVp6dQicn9P0yKfhQCI?_a=BAMAMicg0"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":5}},{"@type":"Comment","text":"Hope that in a nearest future everyone could have personal offline LLM model which would easily solve routine tasks","datePublished":"2026-09-13T19:56:39.191Z","url":"https://daily.dev/posts/YR6Y50msp#c-CzvIZebcG","author":{"@type":"Person","name":"andru hudz","url":"https://daily.dev/andruhudz","image":"https://lh3.googleusercontent.com/a/ACg8ocL---snYLZfyby0hGSZRNE0HBZxFUaV8F3aRsSy8J8sc6V6Ew=s96-c"},"interactionStatistic":{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1}},{"@type":"Comment","text":"Since getting my Strix Halo (MinisForum MS-S1 Max), and ever since the release of Qwen3.8-Flash-Next and what the community is doing with their Strix-Halo fork of llama.cpp, I’ve been pushing more and more of my stuff to local.\nSure it might not be fast compared to the likes of ChatGPT and Claude, but compared to session usage on Claude, it’s free (not including the obvious incurred costs of the hardware, etc…).\nI’ve gone from 100% with Claude (and buying more usage credits because I kept getting dangerously close to my weekly limit before rollover) to around 70% Claude, 30% local.\nAs I do more and more testing, and as the community improves the Halo situation, I can see that becoming 50/50 or even 60/40 in favour of local.","datePublished":"2026-09-11T21:58:51.749Z","url":"https://daily.dev/posts/YR6Y50msp#c-55b3OjRHm","author":{"@type":"Person","name":"Ray","url":"https://daily.dev/ray73864","image":"https://avatars.githubusercontent.com/u/1343103?v=4"}},{"@type":"Comment","text":"We have AGI at home: https://mulgadc.com/blog/we-have-agi-at-home","datePublished":"2026-09-17T03:42:11.259Z","url":"https://daily.dev/posts/YR6Y50msp#c-g39cTYX1O","author":{"@type":"Person","name":"Tom Newton","url":"https://daily.dev/toddyholiday","image":"https://media.daily.dev/image/upload/s--ysbIZMDT--/f_auto/v1788752382/avatars/avatar_laWs9srWzbRzSuTxv3qhW?_a=BAMAMicg0"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/open-weight-models-are-eating-proprietary-ai-s-lunch-and-the-numbers-are-brutal-yr6y50msp#faq","mainEntity":[{"@type":"Question","name":"How much did companies save by switching from proprietary AI models to open-weight alternatives?","acceptedAnswer":{"@type":"Answer","text":"Reported enterprise savings from switching to open-weight models range widely: Uber cut cost per AI session by 52% by routing away from frontier models, Pinterest got cost per transaction under 8% of comparable closed-model pricing, and AT&T reported up to 80% savings using open-weight alternatives via LiteLLM routing, with only about a 2% quality drop. Teams weighing model costs against quality tradeoffs can track real-world routing benchmarks like these on daily.dev."}},{"@type":"Question","name":"How has the market share of proprietary AI models changed relative to open-weight models in routed queries?","acceptedAnswer":{"@type":"Answer","text":"Proprietary models' share of routed queries on OpenRouter dropped from roughly 60% to 25% within a few months, as routing systems increasingly send requests to the cheapest capable model, which is often an open-weight option rather than a frontier proprietary model. Anyone deciding which models to route traffic to can follow shifting adoption trends like this on daily.dev."}}]} ```