In February 2024, two things happened in the same week, and the contrast between them tells the state of this debate better than any comparison chart.
Monday: Meta released Llama 2 under a license that permitted nearly unrestricted commercial use — the most capable downloadable model available at that moment. Tens of thousands of developers pulled it that week.
Wednesday: Anthropic published a long essay signed by Dario Amodei arguing, with technical detail, why his company doesn't publish the weights of the most capable models it trains — and why no one should. Same industry, same week, two opposite positions. Both reasoned. Both defensible.
That February week crystallized the question this evergreen tries to answer without propaganda: when does an open model make sense, when does a closed one, and why does a serious professional probably end up using both?
What "open" actually means
The word "open" in AI is messier than it is in traditional software. Worth unpacking.
A model has at least three components that can each be open or closed independently.
Weights — the file with the trained parameters. Llama publishes weights. Claude doesn't.
Code — the software that runs the model and the software that trained it. Llama publishes part of the inference code. The full training code for large models almost no one publishes.
Data — the corpus the model was trained on. Practically no one publishes the full training data, not even the most "open" projects. Copyright, privacy, and competitive reasons all apply.
On top of all that sits the license: what uses are legally permitted. And here's the trick a lot of people miss.
Llama 2 and Llama 3's license allows commercial use in most cases, but it has a clause saying that if your product has more than 700 million monthly active users you need a separate license from Meta. That's why the Open Source Initiative — the organization that formally defines what "open source" means — doesn't consider Llama open source in the strict sense. Llama is "open-weights with a restrictive license."
DeepSeek released R1 in January 2025 under a plain MIT license. Mistral publishes Mistral 7B and Mixtral under Apache 2.0. Those are open source in the traditional meaning of the term.
The distinction matters because if you're going to build a product on top of the model, the license is what defines your future legal position.
The players on the map
Worth mapping the landscape as of April 2026.
Meta (Llama). The biggest player in the open-with-restrictions camp. Llama 3 and Llama 4 are capable models, with versions up to 405 billion parameters. Meta has a strategic reason to play open: if the AI standards get set by Meta instead of OpenAI, Meta wins position. For Meta, Llama isn't a product — it's a platform play.
DeepSeek and Qwen (China). The genuinely-open players of 2025-2026. DeepSeek-R1 under MIT license showed that a Chinese model trained on a relatively modest budget could compete on reasoning with frontier closed models. Qwen, from Alibaba, is Apache 2.0. They're the models that most pushed the "if China publishes open and the US closes, who wins the race?" conversation.
Mistral (France). The interesting middle play. Publishes some open models (Mistral 7B, Mixtral) and some closed ones (Mistral Large). Their read is pragmatic: open to win mindshare and talent, closed to monetize frontier capability.
xAI (United States). An odd case: Elon Musk promised to open models and in March 2024 released the Grok-1 weights under Apache 2.0. Subsequent versions (Grok-2, Grok-3) weren't released openly. It sits as an intermediate gesture.
OpenAI, Anthropic, Google. The closed side. They publish technical papers, general architectures, sometimes performance figures. They don't publish weights. Each has its own public reasoning about why.
The pro-open argument
The case for open models rests on four concrete pillars.
No vendor lock-in. If your business depends on a closed model and the provider changes pricing, changes terms, or sunsets the model, your business is captured. With open, you have the model. If the provider vanishes, you keep operating.
Auditability. You can inspect how the model responds to specific prompts in a controlled environment. You can study its biases. You can verify what it says when you present your actual use case. With closed, you trust what the provider tells you it does.
Self-hosting for sensitive data. If your data legally can't leave your network (health, finance, defense, legal), an open model running on-premise is often the only technically viable option.
Low marginal cost at high volume. Once you've amortized the infrastructure, each additional inference costs close to zero. Closed APIs charge per token. For very high-volume use cases, the math favors open.
Distributed innovation. The community can fine-tune, quantize, optimize. Llama has thousands of specialized variants on Hugging Face. A closed model depends only on what the provider chooses to release.
The pro-closed argument
The case for closed models also has solid pillars. It isn't just marketing.
Concentrated economics for alignment research. Training a frontier model costs hundreds of millions of dollars. Serious alignment research on top costs tens of millions more. Open models rarely have that budget dedicated to alignment because the money comes back through usage, not through licensing. Claude invests in Constitutional AI partly because it can charge for the result.
Operational accountability. When something goes wrong with Claude, there's a company to hold accountable. There are contracts, SLAs, indemnifications. With an open model running on your server, you're the accountable party. For regulated applications (health, legal, financial), that difference is large.
Revocability. If the provider detects a misuse mode, it can update the model, add mitigations, even revoke access to specific accounts. Once weights are out on the internet, there's no way back.
Frontier technical quality. Closed models tend to be ahead on general capability benchmarks because they have more compute and more proprietary data. The gap narrows every year, but as of April 2026 it still exists in several domains.
Anthropic's position in particular is honest about this: its argument isn't "closing is better for everyone." Its argument is "for us, as a company whose mission is safety, closing the weights is part of the operational commitment." That stance has integrity whether or not you agree with it.
The capture risk
There's a counter-argument to pro-closed worth naming without melodrama.
If the market concentrates into three or four closed companies setting prices, usage terms, and access, AI becomes infrastructure controlled by oligopoly. It isn't a wild hypothesis — it's what happened with cloud compute.
In that scenario, open models stop being just a technical option and become systemic health. They're the guarantee that if closed providers turn abusive, there's an alternative. That guarantee has value even if you don't use it.
That's why Meta, DeepSeek, and Mistral publishing open isn't just generosity — it's competitive pressure that benefits the rest of the market. Closed providers also lower prices and improve terms when there are viable open alternatives.
When do I use one, when the other?
Getting practical, an honest guide by user profile.
Solo professional or small business. Closed. Claude, ChatGPT, or Gemini on subscription. The cost of engineers to maintain an open model on-premise exceeds the savings.
Mid-size company with a technical team and moderate compliance. Mostly closed for quality, with some open experimentation if there's a specific use case (domain fine-tuning, massive batch processing).
Company with strict compliance (health, banking, defense). Open on-premise is typically required for sensitive data, closed as a complementary tool for non-sensitive tasks.
Data consulting with reputational risk. Frontier closed by default (reliability). Open as Plan B for cases where the client demands data residency.
Product built on top of the model. Depends on volume. At low volume, closed for simplicity. At high volume with tight margins, open and fine-tuned.
To close, and to keep going
The "open vs closed" question is old in software and old in hardware. Every time a new technology shows up, the debate gets run again. The historical answer is almost always the same: both coexist, each one settles into the niche where its trade-offs pay off best.
The same is happening in AI. The closed models dominate in frontier quality and in professional use with accountability. The open ones dominate in control, customization, and cost at scale. They'll keep coexisting.
The only bad answer is zealotry. Whoever tells you "the closed models are evil because they hoard knowledge" or "the open ones are irresponsible because they release dangerous capability" is flattening a debate that has real nuance on both sides.
Which constraint weighs most in your work today: immediate quality, data control, or cost at volume?
If you want to understand the player that pushed the open frontier the most over the past year, read Meta and Llama — the open bet. If you want the full competitive map of the industry, The AI race.
Picture two identical cars. Same engine, same horsepower, same price tag. They take you to the same place at the same speed.
One comes with the hood welded shut at the factory and a warranty sticker on the windshield. If something fails, you call the dealer and they fix it. The other one comes with the hood wide open, every part labeled, and the glove box holds the full engine schematic.
Both cars are good. The question isn't which one is better. The question is: what do you do the day something breaks?
The same question, applied to AI
That's the open-vs-closed model conversation.
Closed models — Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini — work like the sealed-hood car. You open the web app, type, get an answer. You don't see how it works inside. You can't copy it. You can't modify it. If something goes wrong, you call the vendor.
Open models — Meta's Llama, DeepSeek, Mistral — work like the open-hood car with the schematic in the glove box. The company that trained it publishes the "weights" (a large file with all the numbers that make the model run). Anyone with a powerful-enough computer can download it, run it on their own machine, and even modify it.
What each one wins
The closed models win on convenience and average quality. You open the site, pay the subscription, and within thirty seconds you're working. You don't need technical knowledge. And the top-performing models of the moment tend to be closed.
The open models win on control. If you're a clinic and your patient data legally can't leave your network, a closed model doesn't work — every question travels to a third-party server. An open model runs inside your own network. Your data never goes anywhere.
They also win on cost at scale. Using a closed API charges you per word. An open model you pay for once (the infrastructure) and then use as much as you want.
An uncomfortable truth
"Open" doesn't always mean the same thing.
Meta publishes Llama's weights, but its license says if your product has more than 700 million monthly active users, you can't use it for free. That's a commercial restriction. Many engineers don't consider that "pure open source."
DeepSeek, on the other hand, publishes its models under an MIT license — the most permissive license around. That's full open source.
xAI released Grok-1's weights in March 2024. Mistral ships both open and closed models. The picture is a spectrum, not a switch.
And Claude — why closed?
Anthropic has a public thesis on why it doesn't publish Claude's weights. Here it is in my own words.
Once the weights are out in the wild, you can't take them back. If you discover someone using them for harm, you can't revoke access. You can't patch the alignment without training a new model. There's no one to hold accountable when something goes wrong.
For Anthropic, closing the weights isn't secrecy. It's part of the operational safety commitment. It's a defensible, honest stance. Not everyone agrees, and the debate is legitimate.
What to take away
Three things worth holding onto:
- There's no universal winner. Open and closed models solve different problems. A mature professional usually ends up using both.
- "Open" is a spectrum, not a binary label. Llama is open-with-restrictions; DeepSeek is fully open; Grok-1 is somewhere in between. Read the license before betting your product on a model.
- For most readers of this blog — professionals without an infrastructure team — a closed model like Claude is the sensible default. The day your work needs data that can't leave your network, then you look at open.
In February 2024, two things happened in the same week, and the contrast between them tells the state of this debate better than any comparison chart.
Monday: Meta released Llama 2 under a license that permitted nearly unrestricted commercial use — the most capable downloadable model available at that moment. Tens of thousands of developers pulled it that week.
Wednesday: Anthropic published a long essay signed by Dario Amodei arguing, with technical detail, why his company doesn't publish the weights of the most capable models it trains — and why no one should. Same industry, same week, two opposite positions. Both reasoned. Both defensible.
That February week crystallized the question this evergreen tries to answer without propaganda: when does an open model make sense, when does a closed one, and why does a serious professional probably end up using both?
What "open" actually means
The word "open" in AI is messier than it is in traditional software. Worth unpacking.
A model has at least three components that can each be open or closed independently.
Weights — the file with the trained parameters. Llama publishes weights. Claude doesn't.
Code — the software that runs the model and the software that trained it. Llama publishes part of the inference code. The full training code for large models almost no one publishes.
Data — the corpus the model was trained on. Practically no one publishes the full training data, not even the most "open" projects. Copyright, privacy, and competitive reasons all apply.
On top of all that sits the license: what uses are legally permitted. And here's the trick a lot of people miss.
Llama 2 and Llama 3's license allows commercial use in most cases, but it has a clause saying that if your product has more than 700 million monthly active users you need a separate license from Meta. That's why the Open Source Initiative — the organization that formally defines what "open source" means — doesn't consider Llama open source in the strict sense. Llama is "open-weights with a restrictive license."
DeepSeek released R1 in January 2025 under a plain MIT license. Mistral publishes Mistral 7B and Mixtral under Apache 2.0. Those are open source in the traditional meaning of the term.
The distinction matters because if you're going to build a product on top of the model, the license is what defines your future legal position.
The players on the map
Worth mapping the landscape as of April 2026.
Meta (Llama). The biggest player in the open-with-restrictions camp. Llama 3 and Llama 4 are capable models, with versions up to 405 billion parameters. Meta has a strategic reason to play open: if the AI standards get set by Meta instead of OpenAI, Meta wins position. For Meta, Llama isn't a product — it's a platform play.
DeepSeek and Qwen (China). The genuinely-open players of 2025-2026. DeepSeek-R1 under MIT license showed that a Chinese model trained on a relatively modest budget could compete on reasoning with frontier closed models. Qwen, from Alibaba, is Apache 2.0. They're the models that most pushed the "if China publishes open and the US closes, who wins the race?" conversation.
Mistral (France). The interesting middle play. Publishes some open models (Mistral 7B, Mixtral) and some closed ones (Mistral Large). Their read is pragmatic: open to win mindshare and talent, closed to monetize frontier capability.
xAI (United States). An odd case: Elon Musk promised to open models and in March 2024 released the Grok-1 weights under Apache 2.0. Subsequent versions (Grok-2, Grok-3) weren't released openly. It sits as an intermediate gesture.
OpenAI, Anthropic, Google. The closed side. They publish technical papers, general architectures, sometimes performance figures. They don't publish weights. Each has its own public reasoning about why.
The pro-open argument
The case for open models rests on four concrete pillars.
No vendor lock-in. If your business depends on a closed model and the provider changes pricing, changes terms, or sunsets the model, your business is captured. With open, you have the model. If the provider vanishes, you keep operating.
Auditability. You can inspect how the model responds to specific prompts in a controlled environment. You can study its biases. You can verify what it says when you present your actual use case. With closed, you trust what the provider tells you it does.
Self-hosting for sensitive data. If your data legally can't leave your network (health, finance, defense, legal), an open model running on-premise is often the only technically viable option.
Low marginal cost at high volume. Once you've amortized the infrastructure, each additional inference costs close to zero. Closed APIs charge per token. For very high-volume use cases, the math favors open.
Distributed innovation. The community can fine-tune, quantize, optimize. Llama has thousands of specialized variants on Hugging Face. A closed model depends only on what the provider chooses to release.
The pro-closed argument
The case for closed models also has solid pillars. It isn't just marketing.
Concentrated economics for alignment research. Training a frontier model costs hundreds of millions of dollars. Serious alignment research on top costs tens of millions more. Open models rarely have that budget dedicated to alignment because the money comes back through usage, not through licensing. Claude invests in Constitutional AI partly because it can charge for the result.
Operational accountability. When something goes wrong with Claude, there's a company to hold accountable. There are contracts, SLAs, indemnifications. With an open model running on your server, you're the accountable party. For regulated applications (health, legal, financial), that difference is large.
Revocability. If the provider detects a misuse mode, it can update the model, add mitigations, even revoke access to specific accounts. Once weights are out on the internet, there's no way back.
Frontier technical quality. Closed models tend to be ahead on general capability benchmarks because they have more compute and more proprietary data. The gap narrows every year, but as of April 2026 it still exists in several domains.
Anthropic's position in particular is honest about this: its argument isn't "closing is better for everyone." Its argument is "for us, as a company whose mission is safety, closing the weights is part of the operational commitment." That stance has integrity whether or not you agree with it.
The capture risk
There's a counter-argument to pro-closed worth naming without melodrama.
If the market concentrates into three or four closed companies setting prices, usage terms, and access, AI becomes infrastructure controlled by oligopoly. It isn't a wild hypothesis — it's what happened with cloud compute.
In that scenario, open models stop being just a technical option and become systemic health. They're the guarantee that if closed providers turn abusive, there's an alternative. That guarantee has value even if you don't use it.
That's why Meta, DeepSeek, and Mistral publishing open isn't just generosity — it's competitive pressure that benefits the rest of the market. Closed providers also lower prices and improve terms when there are viable open alternatives.
When do I use one, when the other?
Getting practical, an honest guide by user profile.
Solo professional or small business. Closed. Claude, ChatGPT, or Gemini on subscription. The cost of engineers to maintain an open model on-premise exceeds the savings.
Mid-size company with a technical team and moderate compliance. Mostly closed for quality, with some open experimentation if there's a specific use case (domain fine-tuning, massive batch processing).
Company with strict compliance (health, banking, defense). Open on-premise is typically required for sensitive data, closed as a complementary tool for non-sensitive tasks.
Data consulting with reputational risk. Frontier closed by default (reliability). Open as Plan B for cases where the client demands data residency.
Product built on top of the model. Depends on volume. At low volume, closed for simplicity. At high volume with tight margins, open and fine-tuned.
To close, and to keep going
The "open vs closed" question is old in software and old in hardware. Every time a new technology shows up, the debate gets run again. The historical answer is almost always the same: both coexist, each one settles into the niche where its trade-offs pay off best.
The same is happening in AI. The closed models dominate in frontier quality and in professional use with accountability. The open ones dominate in control, customization, and cost at scale. They'll keep coexisting.
The only bad answer is zealotry. Whoever tells you "the closed models are evil because they hoard knowledge" or "the open ones are irresponsible because they release dangerous capability" is flattening a debate that has real nuance on both sides.
Which constraint weighs most in your work today: immediate quality, data control, or cost at volume?
If you want to understand the player that pushed the open frontier the most over the past year, read Meta and Llama — the open bet. If you want the full competitive map of the industry, The AI race.
The phrase "open source model" applied to LLMs has a definitional problem worth meeting head-on. In traditional software, "open source" has a formal definition — the Open Source Initiative's Open Source Definition — specifying ten precise criteria about licensing, redistribution, use, and modification. A program is open source or it isn't, by those criteria.
No equivalent consensus yet exists for AI models. And that isn't accidental — it reflects that a trained model is technically different from a program. There are at least three artifacts that can each be open or closed: the weights (the model proper), the code (training and inference), and the data (the training corpus). A model can have open weights and closed code, or open code and closed data, or any intermediate combination. The Open Source Initiative published in late 2024 a draft "Open Source AI Definition" attempting to formalize this, but its adoption is far from universal.
That conceptual ambiguity is the starting point of any serious expert analysis of the topic.
The real openness spectrum
An honest technical map of the state as of April 2026 requires replacing the "open vs closed" dichotomy with a spectrum of at least five levels.
Level 1 — Fully closed. Private weights, private code, private data. Access only through API. Claude, GPT-4o, Gemini sit here. The only public elements are the architecture at a conceptual level and whatever performance metrics the provider chooses to disclose.
Level 2 — Open-weights with restrictive license. Public weights, license that limits commercial use above specific thresholds (users, revenue, jurisdictions). Llama 2/3/4 sit here. The Llama Community License v2 restricts use for products with more than 700M MAU and has acceptable-use clauses Meta can legally enforce.
Level 3 — Open-weights with standard permissive license. Public weights under Apache 2.0, MIT, or similar. Mistral 7B, Mixtral, Qwen 2, DeepSeek-R1 sit here. Commercial use without restrictions beyond the standard ones (attribution, no endorsement, no liability).
Level 4 — Open-weights + inference code + partial training recipes. The most transparent Hugging Face projects, some Allen AI (OLMo) releases, Mistral in its most open releases. They publish enough for a technical team to approximately reproduce the process.
Level 5 — Fully open (weights, code, data). Very rare at frontier scale. BLOOM (2022), Allen AI's OLMo 2 (2024) are the most cited examples. None is a frontier model competitive with GPT-4o or Claude Opus. The reason is economic and legal — frontier-scale training data includes materials for which rights clarification is incomplete.
The relevant distinction for competitive analysis isn't "open vs closed" but "which level of the spectrum each actor operates at, why, and what implications that has."
Amodei's thesis on not publishing weights
In his early-2024 essay, Dario Amodei articulated Anthropic's position with technical precision worth analyzing because it's the strongest argument on the closed side.
The argument has three components.
Component 1 — Irreversibility. Once weights are publicly available, the capability they represent can't be uninstalled from the world. If it's later discovered that the model has a dangerous capability undetected before release (e.g., effective assistance in pathogen design, explosives synthesis, cyberattacks), the provider no longer has any mechanism to mitigate. A closed model can be updated, restricted, or pulled. An open model is already in thousands of copies distributed globally.
Component 2 — Alignment amortization. Frontier alignment research — mechanistic interpretability, constitutional AI, systematic red-teaming — is expensive work requiring dedicated compute and top-level researchers. That cost is amortized for the provider through model usage (API, subscriptions). Publishing weights gives away the product of the alignment work but not the cost incurred. At the margin, that reduces the economic incentive to invest in alignment, because every dollar invested benefits competitors at no cost.
Component 3 — Legal accountability. A closed provider can be sued if its model causes harm. That creates aligned incentives: the company with legal skin in the game invests more in harm mitigation than one that publishes weights and walks away. Current legal structure doesn't cleanly transfer liability from an open model to its deployer.
The honest critique of this argument is that it rests on an empirical assumption difficult to verify: that the expected marginal harm of publishing frontier weights exceeds the expected marginal benefit in innovation and decentralization. That calculation doesn't have a consensus answer in the community. Serious people argue in both directions with partial data.
The technical counter-argument from the open side
The open side also has a body of serious technical arguments worth presenting rigorously.
Democratized interpretability. Research on mechanistic interpretability progresses much faster when the academic community has access to weights. Work on sparse autoencoders, probing, feature attribution depends on being able to interact with weights directly. If only Anthropic, OpenAI, and Google can investigate the internals of frontier models, the academic field is structurally subordinated to three companies' research agendas. That's bad for general progress.
Independent safety validation. External red-teaming and independent audit of behavior are only possible in depth on public weights. The claim "our model is safe" made by the provider isn't falsifiable without access to the artifact. Open models let third parties make verifiable claims about risks and mitigations.
Avoids infrastructural power concentration. If three companies control the world's frontier models, AI becomes infrastructure with all the problematic dynamics that implies (regulatory capture, rent-seeking, systemic fragility). Capable open models are the only structural way to keep the field competitive long-term.
Technological sovereignty. For countries and organizations outside the US-UK-rest-of-world triangle, depending exclusively on foreign closed providers is a strategic risk. The European Union has pushed this explicitly; China has its own open ecosystem for analogous reasons. DeepSeek and Qwen are partly national-sovereignty moves dressed as technical releases.
An expert should hold both sets of arguments in mind and acknowledge that the debate isn't resolved.
Comparative economics as of April 2026
Honest numerical modeling of the economic trade-off for real use cases.
Case A — Startup with 10 million requests/month, average tokens 500 in / 1000 out. Closed API (Claude Sonnet): ~$4,500/month. Self-host of Llama 3 70B: $3K amortized capex + ~$2,500/month compute + $500/month operations. Break-even at 3 months. But the technical setup and operational risk mean most startups at this scale stay with the closed API. The decision isn't purely economic.
Case B — Mid-size company at 100M requests/month, moderate compliance. Closed API: ~$45K/month. Self-host: $15K capex + ~$18K/month compute + $5K/month operations. Meaningful structural savings, but requires a dedicated technical team (2-3 specialized engineers, ~$30K/month fully loaded). Net-net, break-even depends on the exact case.
Case C — Fortune 500 with strict data residency. The pure cost calculation is irrelevant because compliance makes closed API not legally viable for certain workloads. Open on-premise is effectively mandatory.
Case D — Heavy fine-tuning use. Customizing an open model for a specific domain costs on the order of $50K-$200K in compute and expertise. A closed model with limited fine-tuning (when available) costs similarly but with less control over the outcome. For cases where the domain is central to the product (medicine, law, proprietary code), open fine-tuned wins structurally.
Case E — Rapid prototype / MVP. Closed API wins without debate on time-to-value. Spinning up open-source infrastructure to validate an idea is an operational anti-pattern.
The Chinese market as a structural variable
An expert analysis can't skip China's role as a force shifting the equilibrium.
DeepSeek-R1 published under MIT in January 2025 was a market event because it demonstrated three things simultaneously. One: that Chinese models could compete on reasoning capability with Western frontier models. Two: that they could do so with reported training budgets on the order of $5-10M (a fraction of what OpenAI spends), though the figures are disputed. Three: that they were willing to publish them under a maximally permissive license.
That combination changed the strategic calculus. Western companies that had decided to close on safety considerations saw that Chinese openness was happening anyway — and that DeepSeek's weights were circulating globally on Hugging Face. The argument "if we close, the capability doesn't proliferate" lost empirical force.
Alibaba's Qwen plays a similar game under Apache license. The strategic question "is the next generation of competitive open models going to come from the US or from China?" today has an uncomfortable answer for the Western ecosystem.
Editorial thesis
I'll close with a thesis that goes past reporting and into evaluation.
The question "open or closed?" is badly framed because it presupposes exclusivity. The right question for a mature professional in 2026 is "which is the right bundle of models for my work?"
The typical high-sophistication bundle has three layers. A frontier closed model (Claude Opus or equivalent) for high-stakes work where reliability and accountability are absolute priority. A self-hostable open model (Llama 4, Mistral Large, or DeepSeek depending on license) running on-premise or in private cloud for data that can't leave the network or for massive volume with controlled marginal cost. A fine-tuneable open model (Mistral 7B, Qwen, smaller Llama) for domain-specific tasks where fine customization beats generalist quality.
The existence of that three-layer stack is the strongest argument against the "one will win" view. Each layer has a set of properties the others can't replicate without destroying their own trade-offs. The frontier closed can't turn open without losing the alignment economics. The permissive open can't guarantee legal accountability. The small fine-tuneable can't compete in general quality with the large closed.
The healthy ecosystem of the near future is one where the three layers coexist, each performing an irreducible function. A professional operating in a single layer is operating with an incomplete tool.
This blog's stance is therefore honestly pro-Claude but not exclusivist. Claude is the default for most readers because it's the most reliable tool for producing work your reputation depends on. And at the same time, knowing how to deploy an open model when the case requires it is part of professional competence over the next few years. It isn't incompatibility — it's technical craft.
Do you have an experiment running today with some open model, even a small one, so you aren't stuck depending only on closed APIs the day the problem requires it?
