Will open-weights models displace frontier models for most workloads?

Kieran Flanagan · Jul 11, 2025

Most companies will end up with a routing layer that automatically switches between open source, paid and reasoning models depending on the task

running expensive reasoning models like o3 on tasks a cheap open source model could handle is wasteful, so routing optimizes cost

Scope: 'at some point'

47:13 20Growth: The Death of Growth Teams? | How Hubspot Use AI to Triple Email Conversion | The Future of AI SEO | Why Prompt Engineering is the New Coding | What Every CMO Needs to Know About AI in 2025

Jonathan Siddharth · Dec 1, 2025

Whether closed or open models are better depends on the application: enterprises use a mix, with closed models easier to start with and open models chosen for cost and customizability

Turing sees enterprise demand for both, particularly in the small language model regime of 0.5B–10B parameters

Scope: about small language models between half a billion and 10 billion parameters

59:59 20VC: Scale, Surge, Turing, Mercor: Who Wins & Who Loses in Data Labelling | Is Revenue in Data Labelling Real or GMV? | Why 99% of Knowledge Work Will Go and What Happens Then? | Why SaaS is Dead in a World of AI with Jonathan Siddharth @ Turing

Matan Grinberg · Jun 13, 2026

Open-source models are an important counterbalance to frontier labs rather than merely a threat, because enterprises will discover most of their tasks don't require frontier intelligence and can be done faster and cheaper

Good resource allocation requires being able to sit anywhere on the cost/quality/speed trade-off

16:55 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Matan Grinberg · Jun 13, 2026

Spend is already concentrating on frontier models for the small set of key planning steps while open models handle the bulk of implementation tokens

If planning is the thing that determines the outcome, it's worth allocating a large budget to those few steps even though most tokens don't go there

Scope: most tokens still go to implementation; open models are typically really good at implementation

25:02 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Clay Bavor · Jul 4, 2026

There will be an assembly line where yesterday's frontier models become cheap open-weights fine-tunes for specific workloads, and companies will mix and match frontier and open models by task

GPT-4, which was frontier in early 2023, now costs roughly 1/300th per intelligence-equivalent token

10:48 20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Lin Qiao · Jul 20, 2026

The 'frontier' is not a single model but a company's own routing mechanism that decomposes a task across a highly intelligent expensive closed model for the hardest judgments and smaller customized open models for sub-agent work

Application companies deeply understand their own use case and own the evals, so they are best placed to build the decomposition

28:43 20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

Matt Murphy · Jul 27, 2026

The market is entering a stage where companies optimize across a tapestry of frontier model families, open source models and self-trained models rather than using one model

Even Anthropic itself ships a family of models (Sonnet, Opus) because one size doesn't fit all, and at scale companies allocate workloads across options

25:05 20VC: Leading Anthropic's First Ever Round | Will Open Source Threaten Anthropic's Business | Do Margins Matter in a World of AI | Why Triple, Triple, Double, Double is Not Good Enough Today | Why Series A is Hard Today with Matt Murphy @ Menlo

Zach Lloyd · Oct 17, 2025 · speculative

Coding may literally be solved at the model layer — given a codebase, prompt and enough context, a non-frontier model can make the right change.

The level of intelligence per token may already be good enough for coding tasks when paired with good context

Scope: 'may' / 'a crazy thing to say'; at the model layer specifically

41:54 20VC: The Startup Adding $1M ARR Every Week | Competing Against OpenAI's Codex and Claude Code: Who Wins | Why Gemini is Failing and GPT-5 Is Winning | Do Margins Matter in a World of AI | The Ugly Truth About AI Coding with Zach Lloyd, Warp

Joelle Pineau · Nov 3, 2025

There is large unmet demand for small, efficient models that run on one or two GPUs, not just ever-larger frontier models

During the peak LLM frenzy she pulled download stats and found RoBERTa, a small 2019 model, was still getting 20 million downloads a month — people want models they can actually run

53:42 20VC: Cohere's Chief AI Officer on Why Scaling Laws Will Continue | Whether You Can Buy Success in AI with Talent Acquisitions | The Future of Synthetic Data & What It Means for Models | Why AI Coding is Akin to Image Generation in 2015 with Joelle Pineau

Harry Stebbings · Jun 15, 2026 · hedged

We dramatically overestimate how important frontier models are for doing quite basic work

12:46 20VC: Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI | Why Dario Has Done a Disservice to AI with his Labour Replacement Messaging with Aravind Srinivas, Founder @ Perplexity

Mike Mignano · Jul 6, 2026

About 80% of non-coding enterprise tasks can be handled by non-frontier, open source models, while coding still requires frontier models

Tasks like summarization and generating docs and briefs don't need frontier capability, and open source models are catching up faster than they previously were

Scope: coding excepted

25:56 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Avishai Abrahami · Jul 13, 2026

For complex tasks like Base44's, running your own model is only roughly 1-30% cheaper than frontier models; the dramatic multiples of savings only apply to simpler tasks where basic open-source models suffice.

Cheap open-source models work great on smaller, basic tasks, but a complicated model needed for something like Base44 won't reach that level of cost advantage.

Scope: depends on the task; achievable savings on complex tasks more like 5-10%

17:04 20VC: Wix's Founder on What Wall St Gets Wrong About AI and Wix | Will Base44 Win the Vibe Coding Wars | The Truth About the Economics of Vibe-Coding | The Buyback Disaster: Lessons Learned with Avishai Abrahami

Jeff Seibert · Nov 22, 2023

The Apple/Android analogy holds for AI: a vertically integrated proprietary model will likely stay the most advanced, but an open or more open equivalent will be a very close second and just as good for many use cases

Full vertical integration lets the closed player control all the variables and stay best, but that advantage doesn't preclude a good-enough open alternative

Scope: doesn't claim the dynamic is different from past open-vs-closed history

22:42 20VC: Why OpenAI Will Become an Infrastructure Play, Why Apple Will Win in an AI World, Why Google is the Most Vulnerable Incumbent, Will LLMs Be Commoditised, Which Startups Are Thin vs Thick Wrappers on Top of LLMs with Jeff Seibert, Founder @ Digits

Andrew Feldman · May 26, 2026

The best open source models are made by Chinese companies and are exceptionally good, though still not quite as good as the closed source frontier models

Kimi K2, DeepSeek, Qwen and GLM are extraordinarily good models and have been easy to adopt and run fast on Cerebras

Scope: not quite at closed-source frontier quality

38:22 20VC: Cerebras CEO on the Future of Data Centres, Token Costs and Memory | We are Not in an Infra Bubble & Dario Got a Bad Deal with Elon for Compute | Should US Companies Sell to China & Why Most Layoffs are AI Washed with Andrew Feldman

Matan Grinberg · Jun 13, 2026

Engineers systematically overestimate how much of their work requires a frontier model, out of ego about the difficulty of their own work; open models can usually handle it

He experienced this bias himself when first switching over — assumed an open model couldn't handle his work and it could

Scope: based on personal experience and observed dynamic

17:33 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founder with Matan Grinberg, Founder @ Factory

Mike Mignano · Jul 6, 2026

More smart teams will shift toward open-source models as the open ecosystem becomes genuinely competitive

Startups and teams go where the incentives are, and the open 'Rebel Alliance' now has a fighting chance

26:33 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Harry Stebbings · Jul 20, 2026

Open source models have accelerated sharply in the last three months, reaching roughly 90% of frontier capability at around 15x better cost effectiveness

Scope: 'not quite comparable' to frontier models; cost-effectiveness figure credited to Chamath

12:48 20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

Anish Acharya · Feb 9, 2026

We are not at a point in the cycle where companies choose models primarily for cost optimization, so closed models remain advantaged over open source

Companies are optimizing for maximizing the direction of their ambition and their ability to fulfill it rather than taking cost out, and closed models are still slightly ahead on capability

Scope: exceptions exist for idiosyncratic product-quality reasons; applies to most cases, not all

59:44 20VC: Is SaaS Dead in a World of AI | Do Margins Matter Anymore | Is Triple, Triple, Double, Double Dead Today? | Who Wins the Dev Market: Cursor or Claude Code | Why We Are Not in an AI Bubble with Anish Acharya @ a16z

Matt Murphy · Jul 27, 2026 · hedged

Open source models fine-tuned on a company's own data will not be powerful enough to displace frontier models like Anthropic's

Anthropic's models are uniquely performant and intelligent, so an open-source substitute can only cover part of what you're doing

Scope: open source will still be functional and positive for some portion of workloads

0:00 20VC: Leading Anthropic's First Ever Round | Will Open Source Threaten Anthropic's Business | Do Margins Matter in a World of AI | Why Triple, Triple, Double, Double is Not Good Enough Today | Why Series A is Hard Today with Matt Murphy @ Menlo

Matt Murphy · Jul 27, 2026

Open source models will not displace frontier models like Anthropic for the large majority of enterprise workloads; companies will run a mix rather than shifting ~96% to open source

Anthropic's models are distinctively performant, and application companies are seeing that using Anthropic increases customer retention, revenue and engagement even though open source is cheaper

Scope: allows that some API calls don't need frontier-level functionality; illustrative split like 50% Anthropic / 50% open source

23:43 20VC: Leading Anthropic's First Ever Round | Will Open Source Threaten Anthropic's Business | Do Margins Matter in a World of AI | Why Triple, Triple, Double, Double is Not Good Enough Today | Why Series A is Hard Today with Matt Murphy @ Menlo

Brendan Foody · Jun 1, 2026 · hedged

OpenAI and Anthropic are still incredible investments despite open source competition, because total demand will grow by perhaps four or five orders of magnitude in five years

Demand growth of many orders of magnitude outweighs the competitive pressure from distilled open source models

Scope: order-of-magnitude demand figure given loosely

49:41 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Clay Bavor · Jul 4, 2026

As open-weights models improve, the set of tasks worth pointing a frontier model at shrinks, but the ceiling on demand for frontier intelligence is still far higher than people imagine

Intelligence that can work around the clock to invent, build and discover opens uses that are hard to get your mind around

Scope: concedes the open-model-improvement point

12:39 20VC: Open Models vs Frontier Models: Who Actually Wins? | The $100,000 Token Budget Every Engineer Will Need | Why Forward-Deployed Engineers Are the Future of Enterprise AI with Clay Bavor, Co-Founder of Sierra

Jerry Murdock · Aug 22, 2026

If frontier labs execute against their goals of continuous and ultimately lifelong learning, demand for frontier models will never cease and will keep growing over the next decade

Frontier companies are committed to continuous and lifelong learning in their models, which sustains their differentiation

Scope: conditional on execution over the next decade; allows for short-term disruptions lasting three months to a year where economics appear to level out

15:27 20VC: The AI Bubble Will Burst: Half the Neoclouds Will Die | China: Should We Ban Chip Exports & Be Fearful of Chinese Open-Source | Mag7: Who Dies and Who Thrives: Why Meta is Meh and Microsoft is Mega

Harry Stebbings · Feb 9, 2026

Jason Lemkin says this will be the year of true substitution of AI products based on price — the shift from 'it works' to 'it works but it's too expensive'

Lemkin uses ElevenLabs for voice in one of his games and finds it too expensive despite loving it

60:50 20VC: Is SaaS Dead in a World of AI | Do Margins Matter Anymore | Is Triple, Triple, Double, Double Dead Today? | Who Wins the Dev Market: Cursor or Claude Code | Why We Are Not in an AI Bubble with Anish Acharya @ a16z

Harry Stebbings · Apr 11, 2026

Jason Lemkin argues this will be the year of substitution, where buyers try a great but expensive product like ElevenLabs and then swap to a cheaper alternative delivering 80% of it

Jason Lemkin used ElevenLabs for a game's voice, loved it, and found costs escalated as usage grew

31:42 20Sales: ElevenLabs: Why We Set a 20x Sales Quota | How to Structure Sales Compensation Plans | Customer Success: 'Total BS' or Growth Engine? | Building an AI Sales Machine: What Tools & Tactics Must CROs Adopt Today with Carles Reina

Richard Socher · Apr 18, 2025

Open source models have caught up enough that closed foundation models can no longer claim super-unique technology, and sophisticated users will increasingly use open source

It follows the database pattern: Oracle mattered when a really good database was powerful, and only truly enormous-scale users still need Oracle while everyone else uses simpler alternatives

Scope: applies to sophisticated users; very large scale users may still need the frontier vendor

54:03 20VC: Foundation Models: Who Wins & Who Loses | How Economies and Labour Markets Need to Change in a World of AI | China vs the US in an AI Race: What You Need to Know | Rich Socher, Founder @ You.com

Harry Stebbings · Jun 20, 2026

If most mundane workloads migrate to open source or older models and frontier labs are used only for the hardest tasks, the frontier labs' core business is much smaller than people assume

Usage that can be pushed to cheaper models will be, shrinking the addressable market for frontier providers

27:55 20VC: Why Remote Work is White Collar Fraud | Why Revenge and Patriotism are the Best Founder Traits | Two Questions Every Founder Needs to Ask | The Wild Story of Raising $1BN from Masa Son in an Hour Long Meeting with Ryan Peterson, Founder @ Flexport

Carles Reina · Apr 11, 2026

Open source voice models will not commoditize ElevenLabs, even if they reach equivalent quality

Open source models aren't built for enterprise scale and still require maintenance, operationalization and orchestration; companies that try lose three months and come back

32:12 Enterprise scale and orchestration prevent commoditization

Harry Stebbings · Feb 9, 2026

Far more companies use open-source models than admit it publicly, and willingness to use open models is higher than ever

59:24 Open model usage is underreported and rising

Anish Acharya · Feb 9, 2026

We will not see a mindset shift toward price-based substitution of AI products this year

As models get better, downstream players' ability to productize new capabilities and raise prices has outstripped cost increases; nobody downgrades to a cheaper older model — the capability jump sparks imagination about doing more rather than doing the same thing cheaper

61:37 Capability gains outrun cost so nobody downgrades

Jeff Seibert · Nov 22, 2023

LLMs will be commoditised

Market forces create immense energy for an open source equivalent, Meta appears highly motivated to open source its work, and many want to run and tune models themselves; nothing in tech history that was hard and expensive stayed that way for long

21:56 Market competition and open source pressure will commoditize llms entirely

Anton Osika · Aug 18, 2025 · hedged

The best models will always be closed, but open models may end up being what most people choose because of flexibility and the surrounding ecosystem

Open models offer maximum flexibility and an open ecosystem even if they are not the frontier

59:00 Open models win adoption via flexibility and ecosystem despite closed frontier lead

Zach Lloyd · Oct 17, 2025

Frontier models will not be open source, because the analogy to open source software breaks down — someone has to spend enormous capital to make a frontier model competitive, and there is no economic incentive to give that away.

Open source software works because hobbyist developers donate time; frontier models require huge money, not donated labor

40:38 Frontier economics preclude open sourcing unlike hobbyist software

Anastasios Angelopoulos · Aug 3, 2026

Models are not yet commoditized if you look only at closed source — that layer would be an oligopoly — but rapidly improving Chinese open source models are what is driving the commoditization question.

The closed-source layer is still owned by a very small group of companies, while Chinese open models have improved fast enough to beat the best closed American models on an important subset of tasks like front-end web development.

5:16 Chinese open models are the source of commoditization pressure

Jerry Murdock · Aug 22, 2026

Open source models and ASIC chips will be a tsunami of their own, driven by the fact that frontier models can't be customized by users and cost far more per token

You are not allowed to customize the big Anthropic or OpenAI frontier models, which opens the door to tunable open source; and a double-digit-dollar-per-token frontier model versus a ten-cent open source token is a big enough gap to drive massive adoption even though not all tokens are equal

13:07 Customization lockout plus token cost drives open adoption

Richard Socher · Aug 18, 2023

Open source models will take over a lot of use cases and LLMs will become increasingly commoditized, leaving only a few foundational model companies

University researchers' careers depend on models they can analyze, improve and publish on — they won't run their research agendas on closed APIs, so many very smart people will collaborate to build fully open foundational models and layer improvements on releases like Llama 2

20:48 Academic research needs drive open model collaboration and llm commoditization

Richard Socher · Aug 18, 2023 · hedged

The argument that open ecosystems' lack of alignment means closed will win is not decisive — open models like Llama 2 are already incredibly good and progressing faster than prior open releases

Coordination has succeeded before when there was enough interest, excitement and a strong structure — Wikipedia shows global collaboration with layered trust and editing hurdles can work

23:03 Open models can match alignment and quality despite decentralized development

Christian Kleinerman · Sep 22, 2023 · hedged

Open weights will drive research and innovation, but commercial solutions will largely be hosted cloud services, so open-vs-closed may not matter much to customers

As with open source, what ultimately wins is who has the best answers or best-integrated service for customers

31:40 Open weights drive research but hosted cloud services win commercially so open vs closed matters little to customers

Mike Mignano · Jul 6, 2026 · hedged

In a plateaued, commoditized model market, enterprises and individuals optimize for cost — leaning on open weight and open source models and routing layers to get the most intelligence per dollar rather than always defaulting to the most powerful model

Buyers start weighing the tradeoff between intelligence and token spend

18:01 Plateau commoditizes models so buyers optimize intelligence per dollar

Guillermo Rauch · Oct 6, 2023 · hedged

Whether open or closed AI systems win over the next decade is genuinely undetermined — there is real evidence on both sides

Open models have an entire ecosystem betting researchers, chip budgets, documentation and user-space frameworks on them getting better, but a multi-year technology lead in a closed system may be impossible to overcome

47:25 Outcome is genuinely undetermined real evidence exists on both sides

Arvind Jain · Jul 11, 2026

The majority of enterprise workloads will run on open source models within three years

18:02 Majority of enterprise workloads run open source within three years

Vince Hankes · May 3, 2023

Founders are increasingly choosing hosted model APIs over open source models because open source forces them to solve infrastructure scalability and reliability problems themselves

It mirrors cloud ten years ago — a bank might reasonably run its own infrastructure, but you would never have told a startup to do so when Amazon was doing it for them; resource-constrained companies should be shipping customer value instead

19:10 Infra burden of open source pushes builders toward hosted frontier apis

Roman Chernin · Jun 8, 2026

The decisive property of open models is that they are tunable and trainable, not that they are open source — a post-trained specialized model can outperform the best universal frontier model on a specific use case

Once you know your use case and have customer data, you don't need a best-in-the-world universal model; you need one specialized to your case

7:21 Tunability not openness is the decisive property

Sam Altman · Apr 15, 2024

The open-source versus managed-service debate is a detail that misses the bigger picture: intelligence is shifting from a scarce resource requiring many smart people to something one person can access abundantly and cheaply.

Today doing anything intelligence-intensive requires assembling many smart people across the whole stack; the revolution is that this becomes accessible to individuals.

17:24 Open vs closed debate is secondary to intelligence becoming individually accessible

Lin Qiao · Jul 20, 2026 · hedged

Frontier-model price cuts will not close the gap with open weights, because open models have effectively zero acquisition cost while frontier labs must recoup R&D

Open models released by companies like NVIDIA cost nothing to use, whereas frontier labs carry fundamental R&D costs they must recover, and closed general-purpose models can't be customized

17:42 Frontier cannot price match zero acquisition cost open weights

Lin Qiao · Jul 20, 2026

Comparing Fireworks' price to cheaper competitors like Together is not apples-to-apples, because Fireworks sells customized models optimized for quality rather than commoditized off-the-shelf serving

The majority of Fireworks traffic is customized models where quality — both model quality for the specific use case and inference quality — is the primary optimization target

45:17 Customization and quality keep serving from commoditizing

Bret Taylor · Oct 2, 2024

He has changed his mind over the past year on how quickly open source foundation models would become impactful — Zuckerberg accelerated both the timing and the quality well beyond his original thesis.

His March-2023 thesis was that frontier models would be financed by hyperscalers with a meaningful open source model or two emerging eventually, on the pattern of Postgres/MySQL, Google adopting Linux, or Facebook adopting MySQL and Memcache; Llama 3.1 arrived far sooner and better than that.

34:30 Meta accelerated open source frontier model quality and adoption timeline beyond expectations

Matan Grinberg · Jun 13, 2026

The 10-20% of tokens that require frontier models are the most valuable ones because they are the planning and decision-making tokens

It mirrors how human orgs work: most hours go to gathering data and implementing, but a select few hours of leadership decision-making determine the fate of the company — and those decision-makers are paid the most

24:20 Planning tokens are where frontier value concentrates

Jake Saper · Mar 10, 2025 · hedged

The core bet in Together.ai is that open source LLMs become a dominant part of the enterprise market — dominant meaning a large share like 20%, not necessarily a majority

If open source ends up at 1% of the market rather than 20%, the outcome for the company looks entirely different; the trend has moved positively but remains partly unresolved

41:36 Open source becomes a large minority like 20 percent of enterprise llm market not necessarily majority

Harry Stebbings · Mar 10, 2025

Even if open source only captures 10% of enterprise LLM usage, that is still a very interesting market because the addressable base is every company in the world

Most enterprises are less intelligent and less adventurous than we assume and will stick to core providers, but 10% of the entire universe of companies is still enormous

42:17 Even a 10 percent open source share is huge given the universal enterprise base