Proprietary vs. open source LLMs in 2026: an enterprise decision guide

In March 2026, the price per million tokens of an LLM ranges from $0.014 (DeepSeek V3.2) to $30 (GPT-5.4 Pro). A range of 2,142×. That's not a price difference. It's a market where the same product category has versions that cost as much as a coffee and versions that cost as much as dinner for four.

Meanwhile, the performance gap between the best open source models and the best proprietary ones has narrowed to roughly 10%. On benchmarks like GPQA, SWE-bench and Chatbot Arena, open models like DeepSeek and Kimi K2.5 compete in the top range. It's no longer "good for open source". It's good, period.

This changes the nature of the decision. When the performance difference was 40-50%, choosing proprietary was obvious if you could afford it. When it's 10%, the decision stops being technical and becomes strategic: cost, privacy, regulation, control, iteration speed, dependency.

This article is a decision framework for CTOs and tech leads who need to make this call -- or who made it 6 months ago and the market has shifted under their feet. No commercial agenda: the data is the data.

Cost per million tokens (May 2026)

A summary of the public prices on the most-used hosted APIs and the amortized self-hosted range. The figures were updated on April 28, 2026 and are consistent with the per-model detail in the following sections.

Public prices as of April 28, 2026. The self-hosted row reflects the amortized cost per own GPU at reasonable utilization. It doesn't include private rates or volume discounts. The detail by model, context and benchmarks is developed in the following sections.

The real landscape in March 2026

Let's forget marketing rankings. This is the data from Chatbot Arena, the most reliable benchmark because it's based on blind human evaluation, not on benchmarks that providers optimize to pass.

The 5 proprietary models leading

1. Claude Opus 4.6 (Anthropic): #1 globally with 1,504 Elo. 1M-token context, 128K-token output. $5/$25 per million tokens. The most capable model in complex reasoning and coding, with the widest context window in production.
2. Gemini 3.1 Pro (Google): #2 globally with 1,500 Elo and 94.3% on GPQA (the highest recorded on that benchmark). Clear dominance in scientific reasoning and multimodal analysis.
3. GPT-5.4 (OpenAI): Launched on March 5, 2026 with 1M context. Two-tier pricing: $2.50/$10 (standard) and $30/$180 (Pro). The Pro tier is, by far, the most expensive model on the market.
4. Grok 4.20 (xAI): #4 with 1,493 Elo. Differentiator: a native multi-agent system that runs 4 simultaneous agents in parallel.
5. Gemini 3 Pro (Google): #5 with 1,485 Elo. The best multimodal model in the world, with native capabilities in image, video and audio.

The open source models that changed the conversation

• Llama 4 Scout/Maverick (Meta): 17B active parameters (MoE), 10M tokens of context -- the longest of any model, open or closed. $0.19/MTok on self-hosting.
• DeepSeek V3.2: Comparable to GPT-5. The V3.2-Speciale variant surpasses GPT-5 on several tasks. Trained for $6M (against the ~$100M estimated for GPT-4). API at $0.014/MTok. That's not a typo.
• Mistral Small 4: 119B total parameters, 6B active (MoE). Launched in March 2026. Unifies reasoning, multimodal and coding. The most serious European approach.
• Qwen 3.5 (Alibaba): February 2026, Apache 2.0 license. Strong in multilingual and reasoning.
• Kimi K2.5: HumanEval 99.0 -- the highest score recorded on that coding benchmark. For pure code generation, it's the best model there is, open or closed.

A couple of things jump out of this table. The Elo difference between Claude Opus 4.6 (1,504) and Kimi K2.5 (1,447) is 57 points. On SWE-bench, the difference is 4 percentage points. On GPQA, the difference is more pronounced (91.3% vs 87.6%), which suggests complex scientific reasoning is still terrain where proprietary models keep an edge. But on HumanEval (pure coding), Kimi K2.5 has the best absolute score with 99.0. The terrain is far more nuanced than any "X is better than Y" can capture.

The decision matrix: 4 axes that matter

The question "proprietary or open source" is incomplete. The real decision has 4 dimensions that interact with each other. You can't optimize all 4 at once: you have to decide which one weighs most for each specific use case.

Axis 1: Performance

Proprietary models lead in complex reasoning, agentic coding and multi-step tasks. Open source has closed the gap to ~10% on most general tasks, and in pure coding (HumanEval) they already surpass the proprietary ones.

Does your use case require 100% of the model's capability, or is 90% enough? For most enterprise applications -- RAG, classification, extraction, summarization, content generation -- the 10% difference doesn't justify a 100× price difference.

For agents that execute multi-stage workflows, it changes. Each step accumulates error. A model that's 10% less accurate at each step can be 40% less accurate at the end of a 5-step workflow. There, the proprietary model's marginal performance justifies the cost.

Axis 2: Cost

The 2,142× range between DeepSeek ($0.014/MTok) and GPT-5.4 Pro ($30/MTok) is the largest price dispersion in the history of LLMs. The math that matters:

• Low volume (<100M tokens/month): Proprietary APIs are probably cheaper than self-hosting. No infrastructure to maintain, no MLOps team, no idle GPUs.
• High volume (>500M-1B tokens/month): Self-hosting open source models starts to be profitable.
• Medium volume (100M-500M tokens/month): Gray zone. It depends on the cost of your infra team and how much control you need.

A real case: a company that processed 100% of requests with proprietary models at $50,000/month rerouted 60% of requests (the simplest ones) to open source models. Result: $27,000/month. A 46% saving without degrading quality.

Axis 3: Privacy and data sovereignty

67% of companies are already actively working to avoid AI vendor lock-in. And it's not paranoia: it's regulation.

The EU AI Act has a deadline of August 2, 2026 for high-risk AI systems. Fines reach up to €35 million or 7% of global revenue. If your application processes health, finance, HR data or decisions that affect people, this applies to you directly.

With open source models deployed on-premise, your data never leaves your infrastructure. The OpenEuroLLM project (a European consortium with centers in Spain, a €7 billion budget) will have first versions in mid-2026.

Axis 4: Latency and availability

For interactive applications (chatbots, coding assistants), latency matters. For batch processing, throughput matters. MoE models like Mistral Small 4 (6B active of 119B total) and Llama 4 Scout (17B active) offer performance comparable to much larger models with the latency of small ones.

When open source is the right answer

It's not the answer for everything. But for these cases, it's the clear answer.

RAG and processing of sensitive data

If your enterprise RAG system processes confidential internal documentation, contracts, customer data or financial information, you want that data to stay inside your infrastructure.

Recommended model: Llama 4 Scout for RAG with long documents (10M tokens of context). DeepSeek V3.2 as a low-cost alternative if volume is high.

High volume (>100M tokens/day)

When you process more than 100M tokens a day -- classification, mass extraction, log analysis -- the bill with proprietary APIs scales linearly. With self-hosting, the cost is fixed regardless of volume.

Recommended model: DeepSeek V3.2 for general tasks. Mistral Small 4 if you need multimodal + efficient reasoning.

Fine-tuning for specialized domains

Open source models allow fine-tuning without restrictions and without incremental inference cost. With LoRA on Together AI: $0.48/MTok. If you have domain data (your stack's code, specialized documentation, sector terminology), fine-tuning gives you a model that's yours.

Recommended model: Qwen 3.5 (Apache 2.0) or Llama 4 Maverick as a base.

Sovereignty and EU compliance

For sectors regulated by the AI Act or that need to prove data never leaves the EU, self-hosting on European infrastructure is the most direct route.

Experimentation and fast iteration

For prototyping and assessing viability, DeepSeek at $0.014/MTok removes the economic barrier. Your team can iterate over 50 prompt variations for the cost of a coffee.

When proprietary is still the best option

The 10% performance gap matters more in some cases than in others.

Complex, multi-step reasoning

For tasks that require deep reasoning -- legal analysis, complex technical diagnosis, strategic planning -- the top proprietary models keep a measurable edge. On GPQA, the difference between the best proprietary (94.3%) and the best open source (87.6%) is almost 7 points.

In agentic workflows where each step depends on the previous one, those 7 points compound. An agent that plans, executes and verifies in 5 steps needs each step to be as accurate as possible.

Agentic coding and autonomous agents

On SWE-bench, Claude Opus 4.6 (80.8%) and Gemini 3.1 Pro (80.6%) lead by ~4 points over the best open source. For teams using coding agents with autonomy, that margin matters. Although Kimi K2.5 with 99.0 on HumanEval shows that for pure code generation, open source is already equal or superior. The difference lies in understanding a complete codebase and acting on it.

Startups with low volume

If you process fewer than 50M tokens/month, the cost of self-hosting isn't justified. GPT-5.4 Standard at $2.50/MTok or Claude Opus 4.6 at $5/MTok are marginal costs for low volumes. Zero infrastructure, zero maintenance, time-to-market in days.

Maximum quality with no margin for error

If the cost of an error is high -- medical diagnoses, legal advice, financial decisions -- paying the premium for the most capable model is an insurance premium. The difference between $5/MTok and $0.014/MTok is irrelevant when an error can cost orders of magnitude more.

The real cost: beyond the price per token

The price per million tokens is the metric everyone compares and the one that least reflects the real cost.

Self-hosting TCO

Deploying open source models in production isn't free. The Total Cost of Ownership includes:

• GPU infrastructure: From $125,000/year for a minimal setup to $12M/year for enterprise-scale deployments.
• MLOps team: Maintaining the infra, managing versions, monitoring performance. A senior MLOps engineer in Spain: €50-70K/year.
• Opportunity cost: Every hour your team spends on ML infra is an hour it doesn't spend on product.

The breakeven

Self-hosting starts to be cheaper than APIs from 500M to 1,000M tokens/month. Below that, APIs are more economical when you include all costs. And you can use open source model APIs (DeepSeek, Together AI, Fireworks) without self-hosting: open source's low prices without the infra TCO.

Fine-tuning: the hidden difference

With open source models, fine-tuning with LoRA costs $0.48/MTok and the resulting model is deployed with no inference surcharge. With proprietary ones, there's a training cost plus a surcharge on the base model's inference. If you have domain data, fine-tuning an open source model can give you a model that's 90% of the best proprietary model's performance for your specific task, at a fraction of the recurring cost.

The risks nobody wants to discuss

Vendor lock-in: not paranoia, it's arithmetic

67% of companies are already working to reduce dependency on a single AI provider. When you build against a provider's API, you assume that:

• Their prices won't rise (in March 2026 alone, 115 models changed price)
• Their model will keep existing in the version you use (OpenAI has deprecated models with 6 months' notice)
• Their performance won't degrade (it has happened and it will happen)
• Their terms of service won't change in a way that affects your business

None of those assumptions is safe.

AI Act: the deadline many ignore

August 2, 2026: the deadline to comply with the AI Act for high-risk AI systems. Fines: up to €35M or 7% of global revenue. If your system assists in decisions about hiring, credit, insurance, medical diagnosis or access to public services, this applies to you.

The AI Act has transparency, traceability and documentation requirements that are easier to meet when you control the model than when you depend on a proprietary black box. It doesn't mean proprietary models can't comply. It means the burden of proof is different.

Accelerated obsolescence

115 models changed price in March 2026 alone. New models launch every week. The model you chose 3 months ago may not be the best option today. This isn't an argument for paralysis. It's an argument for building your architecture so that switching models doesn't require rewriting your application.

The strategy that works: multi-model with a context playbook

The answer isn't "proprietary" or "open source". It's "the right model for each use case". And the key to making that work without chaos is having two layers: a technical one (routing) and a methodological one (structured context).

AI gateways: the technical layer

An AI gateway is a single entry point to multiple models. Your application talks to the gateway, and the gateway routes each request to the optimal model based on rules you define: cost, latency, task type, privacy requirements. 37% of companies already operate with a hybrid strategy. The gateway is what makes it sustainable.

Context engineering: the layer that makes you model-independent

The technical layer solves routing. But there's a deeper problem: if every interaction with an LLM depends on ad hoc prompts written for a specific model, switching models means rewriting those prompts. And that doesn't scale.

This is where context engineering comes in as a discipline. The idea is simple: instead of relying on loose prompts optimized for a specific model, you work with structured context -- instructions, constraints, output format and success criteria defined in a playbook that's model-agnostic.

That playbook works the same with Claude Opus 4.6 as with DeepSeek V3.2. The level of detail in the result will vary with the model's capability, but the structure of the request, the guardrails and the quality criteria hold. When a new model is better or cheaper tomorrow, you change the gateway configuration. You don't rewrite the playbook, don't retrain your team, don't redeploy.

This applies both to software development (where at onext we implement it as Spec-Driven Development) and to any enterprise use of LLMs: legal document analysis, report generation, customer support, financial processing. The principle is the same: if the context feeding the model is structured, the model becomes an implementation detail, not an architectural dependency.

This is exactly what we see in the teams deploying multi-stage workflows: the advantage isn't the model they use, it's the architecture that lets them switch models when the market shifts. And the market shifts every week.

The operational playbook

AI Centers of Excellence exist precisely for this: to implement this multi-model architecture with the methodologies and governance that make it sustainable. It's not something you improvise team by team.

Frequently asked questions

Which is better for companies, a proprietary LLM or an open source one?

It depends on the case, there's no absolute winner in 2026. Proprietary models (Claude Sonnet 4, GPT-4o) win on out-of-the-box accuracy, consistent latency and enterprise support. Self-hosted open source models (Llama 3.x, DeepSeek-V3, Qwen 2.5) win on recurring cost —between 5× and 10× cheaper per million tokens— and on privacy of sensitive data that can't leave the corporate environment. The right decision evaluates four axes: effective cost at your real volume, tolerable latency, data privacy and performance measured on your own internal evaluation, not on public benchmarks.

How much does each LLM cost per million tokens in 2026?

At official API prices (May 2026): Claude Sonnet 4 costs $3 input / $15 output per million tokens; GPT-4o, $2.50 / $10; DeepSeek-V3, $0.27 / $1.10; Llama 3.x hosted on Together AI, $0.60 / $0.80. Self-hosted open source models on your own GPUs drop to a range of $0.10 – $0.30 per million tokens, but require your own operational engineering (scaling, observability, patches). "Small" models like GPT-4o mini ($0.15 / $0.60) or Claude Haiku 4.5 ($0.80 / $4) are the most cost-effective option for classification, summarization or routing tasks where you don't need the flagship model.

When is a self-hosted open source LLM worth it vs a proprietary one?

Self-hosting pays off from around 10 million tokens a month —that's where the absolute saving justifies the complexity of maintaining GPUs, models and observability— and whenever there's data that can't leave the enterprise environment for compliance or contractual reasons. Below a million tokens a month, the absolute saving is a few hundred euros: it doesn't compensate against the development speed and quality of a hosted proprietary model. The decision also changes if you already have an MLOps team with your own GPU pool: the marginal cost of adding an extra model is low and self-hosting wins sooner.

Which is the cheapest LLM for companies in 2026?

By hosted API, the most competitive on cost/performance are DeepSeek-V3 ($0.27 / $1.10 per million tokens) and GPT-4o mini ($0.15 / $0.60). If the priority is absolute cost and you have a team to self-host, Llama 3.x on your own GPUs reaches $0.10 – $0.30 per million —between 10× and 30× cheaper than the top proprietary models— but requires serious setup: GPU provisioning, scaling, monitoring and patching. "Cheapest" without context is misleading: the real cost depends on volume, the required latency, the minimum acceptable accuracy and how operational costs are charged.

Is it safe to use proprietary LLMs for confidential enterprise data?

Yes, if you use an enterprise variant with zero data retention and sign the corresponding DPA: Anthropic Claude via AWS Bedrock, Azure OpenAI Service and Google Vertex AI offer that contract with European residency and an explicit exclusion of using data for training. For data that by contract or regulation can't leave the enterprise environment (private banking, healthcare, defense, parts of public administration), the only robust option remains self-hosted open source with prompt governance, input/output logging and full observability of the chain. The key difference in 2026 between a "secure" proprietary setup and a truly sovereign one isn't the model: it's the end-to-end operational control.

The smart decision isn't picking a side

The LLM market in March 2026 looks less and less like "proprietary vs open source" and more and more like an ecosystem where both coexist in the same architecture. 37% of companies already operate this way. In 12 months, it'll be the norm.

• Performance: Gap narrowed to ~10%, except in complex reasoning and multi-step agents where proprietary models keep an edge.
• Cost: From $0.014 to $30 per million tokens. The multi-model strategy saves 40-50% without degrading quality.
• Privacy: The AI Act isn't the future, it's August 2026. €35M in fines isn't a theoretical risk.
• Dependency: 115 models changed price in March alone. Betting on one is a decision that expires fast.

The smart decision is to build the architecture that lets you use the right model for each case, switch when the market shifts, and comply with regulation without rewriting your system. That requires structured context, not loose prompts. Governance, not improvisation.