In 2023, suggesting an open-source AI model for production work was a career-limiting move. The gap between the best free models and GPT-4 wasn't a gap — it was a canyon visible from orbit. Proprietary AI was the only serious option. Everyone knew it.

Everyone was right. Past tense.

The canyon became a curb

As of March 31, 2026, the Artificial Analysis Intelligence Index — a composite score measuring AI performance across math, science, coding, and reasoning — puts the top proprietary model (Gemini 3.1 Pro) at 57 points. The best open-weights model, GLM-5 by Zhipu AI, scores 50. Claude Opus 4.6 sits at 53.

Seven points. That's the entire distance between "pay us a fortune" and "run it yourself for free." Three years ago, embarrassment was the only unit that captured that distance.

The numbers that should worry closed-model vendors

Let's talk specifics.

Alibaba's Qwen team dropped Qwen3-Coder-Next on February 4, 2026 — a coding-focused model using MoE architecture (Mixture of Experts — a design where the model activates only a small fraction of its "brain" for each task, saving compute while staying smart). It scores 70.6% on SWE-Bench Verified, the benchmark that tests whether a model can actually fix real bugs in real codebases. Not toy problems. Real GitHub issues.

DeepSeek shipped V3.2 on December 1, 2025 — a 685-billion-parameter model (parameters are the learned connections in a neural network — more usually means smarter, but also heavier) with a 128K context window (how much text the model can "see" at once — 128K is roughly a 300-page book). It scores 70–74% on the same benchmark depending on the evaluation setup.

Zhipu AI released GLM-5 on February 11, 2026 — a 744B-parameter beast with only 40B active parameters thanks to its own MoE design. It hits 77.8% on SWE-Bench Verified. Zhipu ships it under the MIT license — meaning anyone can use it for anything, commercially, no strings attached.

Organizations with billions in backing built these. Not hobbyists. Not weekend tinkerers. Companies that treat AI as infrastructure.

The economics that change everything

Here's where it gets uncomfortable for API vendors.

Self-hosting an open model on decent GPU hardware costs roughly $2,000–10,000 per month depending on traffic volume. The equivalent API calls to GPT-5 or Claude Opus for the same workload? $20,000–100,000 per month. At high volumes — 100 million tokens daily and above — self-hosting savings hit 40–90%.

For a startup burning runway, that's not an optimization. That's the difference between survival and a "we regret to inform you" blog post.

And then there's the China factor you can't ignore. Qwen (Alibaba), DeepSeek (High-Flyer), and GLM (Zhipu AI) are all Chinese-backed. When a country with 1.4 billion people decides to subsidize AI development and give the results away under MIT licenses, the competitive landscape doesn't shift — it cracks.

But hold on

Benchmarks lie. Every engineer who's deployed these models knows the gap between "scores well on a test" and "works reliably when your users do something weird" is vast.

OpenAI and Anthropic refine their models through RLHF (reinforcement learning from human feedback — basically, thousands of humans telling the model "good answer" or "terrible answer" until it gets better at the hard stuff). Open models can't easily replicate this scale of human curation.

The 7-point gap on average benchmarks masks a much larger gap on tail-end difficulty. When your AI agent encounters the top 5% hardest queries — novel reasoning, unfamiliar code patterns, ambiguous instructions — Claude and GPT-5 still pull away meaningfully.

Self-hosting isn't free either. Running a 685B model requires multiple H100 GPUs, a team that knows CUDA debugging and tensor parallelism (splitting the model across multiple chips so it actually runs), plus ongoing ops overhead. For many companies, the API cost is genuinely cheaper once you factor in engineering time.

And safety. Anyone can fine-tune open models without restrictions. Great for customization, concerning for everything else. The guardrails Anthropic builds aren't just features — they're engineering investments that open models rarely match.

The framework that actually works

Tier 1 — 70% of workloads: Summarization, simple Q&A, classification, structured data extraction. Open models handle these flawlessly. Using GPT-5 for this is driving a Ferrari to buy milk.

Tier 2 — 25% of workloads: Complex code generation, nuanced writing, multi-step reasoning. Open models are competitive but inconsistent. Proprietary models are more reliable. Your mileage depends on your tolerance for occasional failures.

Tier 3 — 5% of workloads: Frontier reasoning, novel problem-solving, the hardest edge cases. Proprietary wins. The gap is real and worth paying for.

The companies winning in 2026 aren't religious about either side. They run open models for the bulk and route the hard stuff to Claude or GPT-5. This isn't clever architecture — it's basic arithmetic.

The trajectory is the story

The gap went from humiliating to negligible in three years. Every quarter, open models improve faster than proprietary ones can extend their lead. The moat isn't gone — but it's evaporating in real time.

Give it two more years, and "open source is good enough" becomes "open source is the default."

If your business plan assumes proprietary AI will always be dramatically better — update your business plan. The canyon is a curb now. And open source doesn't trip on curbs.