Open Source vs Proprietary LLMs

"Open source" and "open weights" are not synonymous. Open-source licensing (Apache 2.0, MIT, GPL) applies to source code and permits unrestricted commercial and private use. Open-weights means the trained model weights are downloadable but may be restricted under a specialized license. LLaMA 3.1 is open-weights, not open source — Meta releases the weights under Llama Community License 2.1, which permits commercial use but includes restrictions that prevent calling derivative models "LLaMA" and require attribution.

Proprietary models are neither open weights nor open source. OpenAI (GPT-4o), Anthropic (Claude Opus 4.7), and Google (Gemini 3.1 Pro) do not release model weights. You access them exclusively via API. The weights remain closed; you cannot see, download, modify, or deploy the model yourself.

Understanding this distinction matters for compliance, customization, and data sovereignty. Learn how LLMs work internally to understand why weights matter.

Proprietary LLMs are closed models accessible only via API — the vendor controls the weights, training data, safety alignment, and all updates. OpenAI (GPT-4o), Anthropic (Claude Opus 4.7), Google (Gemini 3.1 Pro), and Mistral API are proprietary. You cannot download weights, view training data, run inference locally, or customize the model weights directly.

Pricing is per-token API billing on a vendor-controlled server. GPT-4o costs $5 per 1M input tokens and $30 per 1M output tokens. Claude Opus 4.7 costs $5/$25. Gemini 3.1 Pro costs $2.00/$12.00 (≤200K context; higher above 200K). You have no infrastructure cost but cannot predict monthly spend precisely — costs scale with usage.

Proprietary vendors maintain control over model updates, behavior, and alignment. When OpenAI updates GPT-4o, you automatically get the new version. Safety alignment, constitutional AI, and instruction-following are vendor responsibilities. For regulated industries, this can simplify compliance — the vendor maintains audit trails and published safety practices.

Open-Weights Model. A large language model whose trained weights (the numerical parameters learned during training) are publicly available and can be downloaded, modified, fine-tuned, and self-hosted. Examples: LLaMA 4 Scout/Maverick (Meta), Mistral Large 2 (Mistral AI), Qwen 2.5 (Alibaba), DeepSeek-R1 (DeepSeek AI). Not to be confused with open-source licensing; "open weights" is about downloadable model files, not necessarily source code or OSI-compliant licensing.

Proprietary LLM. A large language model whose weights are kept private and never released. Access is exclusively through a vendor's API, requiring per-token billing and network connectivity. Examples: GPT-4o (OpenAI), Claude Opus 4.7 (Anthropic), Gemini 3.1 Pro (Google). Users cannot download, inspect, modify, or self-host the model.

Fine-Tuning. The process of retraining a pretrained model on a new, smaller dataset specific to a domain or task. Fine-tuning updates the model's weights to specialize in your use case (e.g., customer service tone, domain vocabulary). Open-weights models support full fine-tuning via LoRA, QLoRA, or full backpropagation; most proprietary models restrict or forbid fine-tuning.

Training Data Cutoff. The date after which a model has no knowledge of events or information. GPT-4o has a cutoff of December 2025; Claude Opus 4.7 has early-to-mid 2025; Gemini 3.1 Pro has January 2025. Models cannot provide accurate information about events after their cutoff date.

Mixture of Experts (MoE). An LLM architecture where the model contains many "expert" sub-networks, but only a fraction are activated per token. LLaMA 4 Scout (16 experts, ~17B active of 109B total) and Mistral use MoE — meaning inference cost scales with active parameters, not total parameters. MoE enables very large total parameter counts with efficient per-token computation, typically 2–4× more throughput than dense models of equivalent total size.

Model Weights. The numerical parameters (billions to trillions of numbers) learned during model training. Weights determine the model's behavior, knowledge, and reasoning patterns. Open-weights models release these files (~15–800 GB depending on model size); proprietary models keep weights secret.

Open-source LLMs (LLaMA 3.1, Mistral, Qwen) make model weights publicly available — organizations can download, inspect, fine-tune, and self-host them. Proprietary LLMs (GPT-4o, Claude, Gemini) are owned by vendors and accessible only through APIs. Users cannot download or modify proprietary weights, but benefit from managed infrastructure and vendor updates.

On many tasks, yes. The performance gap has narrowed to 7–8 percentage points on reasoning benchmarks (MMLU). On classification, summarization, and domain-specific tasks, open-weights models like LLaMA 3.1 70B now match proprietary peers. Proprietary models still lead on complex multi-step reasoning, agent orchestration, and multimodal input handling.

Companies should use open-source LLMs when data privacy is mandatory (healthcare, finance, legal), when processing more than 10 million tokens per day, when domain-specific fine-tuning is required, or when EU AI Act compliance demands on-premises data residency. Open-weights models also eliminate vendor lock-in and per-token API billing.

For many use cases, yes. Open-source LLMs are production-ready for classification, summarization, extraction, and domain-specific tasks. Proprietary models maintain advantages on complex reasoning, multimodal input, tool integration, and zero-infrastructure deployment. A hybrid approach — routing tasks based on cost, privacy, and performance requirements — outperforms relying on either model class alone.

Use this framework to decide in 30 seconds. Answer: Does your use case fit one of the categories below? If multiple criteria apply, weight them by importance to your project.

Open-weights models are the leading alternative to proprietary APIs — model weights are downloadable, self-hostable, and fine-tunable without per-token API billing. The landscape shifted significantly in 2025: Meta released the LLaMA 4 family, introducing Mixture-of-Experts (MoE) architecture that dramatically increases parameter counts while keeping inference costs similar to smaller dense models.

Meta's LLaMA 4 family is the new open-weights frontier (released April 2025). LLaMA 4 Scout uses MoE with 16 experts (17B active of 109B total parameters) and a 10M token context window — the largest of any open-weights model. It fits on a single H100 80GB GPU for inference, or can run quantized on an RTX 4090. LLaMA 4 Maverick uses 128 experts (17B active of 400B total) with a 1M context window — requiring multi-GPU deployment but offering stronger performance. LLaMA 4 Behemoth (288B active, ~2T total) has been announced but is not publicly released as of May 2026. All remain under Llama Community License 2.1 (not OSI open-source).

DeepSeek-R1 and DeepSeek v3 are strong open-weights competitors from DeepSeek AI. DeepSeek-R1 is a reasoning-focused model that matches frontier performance on math and science tasks at a fraction of the training cost. Both models are released under MIT license — genuinely open-source. They require significant VRAM for inference (70B+ class) but are available via Hugging Face and Ollama.

Mistral Large 2 and Qwen 2.5 remain strong options. Mistral Large 2 (123B parameters, 128K context) targets the "frontier lite" band with competitive reasoning. Qwen 2.5 72B (128K context) excels on Chinese-language tasks. Muse Spark, released by Meta Superintelligence Labs in April 2026, is an emerging successor model — early details are sparse.

Proprietary LLMs are accessed exclusively through vendor-controlled APIs; model weights are never released and cannot be downloaded, modified, or deployed locally. Users pay per-token API billing and accept vendor control over model updates, safety policies, and performance characteristics.

OpenAI's GPT-4o is the current general reasoning leader, released April 23, 2026. GPT-4o supports 1M token context (922K input + 128K output) with a training cutoff of December 2025. It excels at tool use, agent workflows, multimodal input (images, text), and Terminal-Bench tasks (82.7% on Terminal-Bench 2.0). API pricing: $5 per 1M input tokens, $30 per 1M output tokens. GPT-4o Pro variant: $30/$180 per 1M tokens. GPT-4o is API-only; weights are never released.

Anthropic's Claude Opus 4.7 leads on agentic coding and writing quality. Claude supports 1M token context — enabling full codebase analysis and extended research. Training data cutoff: early-to-mid 2025. API pricing: $5 per 1M input tokens, $25 per 1M output tokens. On SWE-bench Pro (agentic coding), Claude Opus 4.7 scores 64.3% — ahead of GPT-4o at 58.6%. Claude does not offer public fine-tuning. Claude Mythos Preview (invitation-only): $25/$125 per 1M tokens, cybersecurity-focused.

Google's Gemini 3.1 Pro offers competitive multimodal capabilities and Google Workspace integration. Context window: 1M tokens (note: the 2M context was Gemini 2.5 Pro; 3.1 Pro is capped at 1M). ARC-AGI-2 score: 77.1%. Training data cutoff: January 2025. API pricing: $2.00 per 1M input tokens, $12.00 per 1M output tokens (≤200K context); $4.00/$18.00 above 200K. Fine-tuning is available for Gemini models.

Benchmark performance depends heavily on the task type — proprietary models no longer lead across all categories. MMLU (broad academic reasoning) still shows a 7–8 point gap, but agentic coding (SWE-bench Pro) shows Claude Opus 4.7 ahead of GPT-4o, and ARC-AGI-2 (abstract reasoning) is led by Gemini 3.1 Pro.

MMLU (broad reasoning) — 2024/2025:

GPT-4o: 88.7% | Claude Opus 4.7: ~88% | LLaMA 3.1 70B: 80.5% | Mistral Large 2: 81.2% | Qwen 2.5 72B: 82.1%

SWE-bench Pro (agentic coding, 2026):

Claude Opus 4.7: 64.3% (SWE-bench Pro) / 87.6% (SWE-bench Verified) | GPT-4o: 58.6% | LLaMA 4 Maverick: data pending

Terminal-Bench 2.0 (CLI & system tasks, 2026):

GPT-4o: 82.7% | Claude Opus 4.7: 69.4%

ARC-AGI-2 (abstract reasoning, 2026):

Gemini 3.1 Pro: 77.1% | GPT-4o: data pending | Claude Opus 4.7: data pending

The benchmark picture is more fragmented in 2026 than in prior years. No single model leads all benchmarks. On agentic coding, Claude Opus 4.7 leads. On CLI reasoning, GPT-4o leads. On abstract reasoning, Gemini leads. On MMLU (academic knowledge), GPT-4o and Claude are near-tied. Open-weights (LLaMA 4 Maverick, DeepSeek-R1) are closing the gap on reasoning tasks.

For task-specific deployment, run your own benchmarks. See how to pick the right model for your use case.

Direct cost comparison: proprietary APIs dominate at low volume; open-weights self-hosting wins at scale. The crossover point is typically 5–10M tokens per day. Below this threshold, API simplicity and no infrastructure cost favor proprietary. Above this, open-weights self-hosting becomes cost-effective.

API pricing as of May 2026:

Self-hosting infrastructure cost: NVIDIA A100 80GB rents for ~$2/hour on cloud; RTX 4090 consumer hardware costs ~$1.50/hour in electricity + amortization (3-year lifespan). For Mistral 7B, inference throughput is ~50–100 tokens/second per GPU, or ~180–360M tokens/day per GPU. Mistral Large 2 or LLaMA 70B: ~20–30 tokens/second per A100, or ~1.7–2.6M tokens/day. At these throughputs:

At 5M tokens/day: A100 self-hosting costs ~$2.50/day. API costs for Claude Opus 4.7: 2.5M × $5/1M + 2.5M × $25/1M = $12.50 + $62.50 = $75/day (assumes 50% input, 50% output). APIs still cheaper. Using GPT-4o at $5/$30: $12.50 + $75 = $87.50/day.

At 50M tokens/day: Need ~20 A100s self-hosting = ~$50/day (or fewer H100s with LLaMA 4 Scout MoE efficiency). API costs: $750/day (Claude Opus 4.7). Open-weights wins decisively.

At 100M tokens/day: Self-hosting ~$100/day (A100). API costs: $1,500/day (Claude Opus 4.7). Open-weights is 15× cheaper.

Verify pricing: OpenAI Pricing · Anthropic Pricing · Google Pricing — rates change quarterly. See tokens, costs, and limits explained for detailed token cost breakdown.

Open-weights models deployed locally = zero data leaves your infrastructure. When you run LLaMA 3.1 via Ollama on your private GPU, no inference data, metadata, or query logs leave your network. This is data sovereignty: you maintain complete control. Proprietary APIs (OpenAI, Anthropic, Google) require you to send requests over the network to external servers. Even with contractual data deletion, the data briefly transits vendor infrastructure and is logged for compliance.

The EU AI Act (2024) designates certain LLM applications as "high-risk," requiring risk documentation, bias testing, and audit trails. Categories include systems that make significant decisions (hiring, credit, legal discovery, benefits determination). High-risk systems must maintain records of how decisions are made, prove non-discrimination, and support human oversight. Open-weights models deployed on-premises make this easier — you control the audit trail and data storage. Proprietary APIs make this harder — you depend on vendor compliance reports, which may be inadequate for regulated industries.

For regulated industries (healthcare, finance, legal services), open-weights is often mandatory. HIPAA (healthcare), SOX (finance), and attorney-client privilege require data residency — meaning data cannot leave your jurisdiction. Proprietary APIs based in the US or other countries violate these requirements. Teams in these sectors typically deploy open-weights models (LLaMA, Mistral, or commercial distributions) on on-premises infrastructure.

Open-weights models permit full fine-tuning; proprietary models restrict it or forbid it. Fine-tuning means retraining the model weights on your own data to specialize it for your domain. You can use LoRA (Low-Rank Adaptation) for efficient fine-tuning, QLoRA for quantized training, or full backpropagation training if you have the compute. After fine-tuning, the model becomes yours — you own the resulting weights, can deploy them anywhere, and can update them offline.

Proprietary fine-tuning availability: OpenAI fine-tuning API works only for GPT-4o mini, GPT-4 (older models). Not available for GPT-4o flagship. Anthropic does not offer fine-tuning for Claude via API. Google offers limited fine-tuning for Gemini. None of these permit ownership of the fine-tuned weights — you rent a fine-tuned copy of the proprietary model.

Security consideration: When fine-tuning on proprietary APIs, your training data is uploaded to vendor servers. For sensitive domains, this violates compliance rules. Open-weights fine-tuning stays on-premises. See prompt injection and security for additional attack surface considerations when using external APIs.

Open-weights models cost less at scale and enable full customization; proprietary models deliver faster time-to-value and managed infrastructure at higher per-token cost. Below 5M tokens/day, proprietary APIs are usually cheaper. Above 10M tokens/day, self-hosted open-weights wins on cost. Choose based on your volume, privacy requirements, and infrastructure readiness.

Open-weights models enable deeper prompt experimentation at lower cost. You can run the same prompt 100 times against a local LLaMA 3.1 instance and iterate on wording, temperature, and structure without per-token billing. Fine-tune the model on prompt-response pairs from your domain. Experiment with jailbreaks and edge cases in your private infrastructure. This sandbox environment is ideal for research, prototyping, and understanding model behavior.

Proprietary APIs are faster to test and easier to scale. You write a prompt, call the GPT-4o or Claude API, and get results in milliseconds with zero infrastructure setup. No need to manage VRAM, quantization, or model downloads. For quick A/B testing, production deployment, and handling variable traffic, proprietary models reduce operational complexity.

Hybrid approach: prototype on open-weights, validate on proprietary. Develop and refine prompts locally with LLaMA 3.1 8B (fast iteration, no cost). Once the prompt strategy is locked, test on GPT-4o or Claude 4.6 to confirm frontier performance. Deploy the better performer to production. This combines open-weights flexibility with proprietary reliability.

Choose open-weights when data privacy, cost at scale, or deep customization requirements dominate your constraints. Open-weights excel in:

Choose proprietary when absolute performance, managed infrastructure, or safety alignment matters most. Proprietary excels in:

Organizations can use hybrid AI architectures that route requests to open-weights models for sensitive data and cost-sensitive tasks, while dispatching complex reasoning and multimodal work to proprietary models. This approach combines the cost efficiency, privacy, and customization of open-weights with the performance and managed infrastructure of proprietary LLMs.

Open-weights is not always cheaper than proprietary APIs. At <5M tokens/day, proprietary APIs (GPT-4o mini, Claude Haiku, Gemini Flash) are often cheaper because infrastructure cost (GPU amortization, electricity, DevOps labor) exceeds API billing. Only above 10M tokens/day do open-weights self-hosting become cost-optimal.

Vague questions don't elicit useful comparisons. Specific context (volume, constraints, requirements) drives better decision-making.