For three years, the dominant AI narrative has been simple: closed frontier labs build the best models, open-source labs ship a few months behind, and the gap is permanent. That story stopped being true sometime in 2025, and most of the industry hasn’t caught up yet. In specific, measurable categories, open-source AI now matches or beats GPT-5, Claude Opus 4, and Gemini Ultra. In others — and this part also matters — it’s still genuinely behind.
📋 Contents
This is the honest accounting. Where open source has won, where it’s drawing level, and where it still loses badly enough that frontier labs can keep charging premium prices.
Where Open Source Has Quietly Won
Three categories where the open ecosystem now leads:
Specialist coding models. DeepSeek-Coder V3, Qwen-Coder 32B, and the latest CodeLlama variants now match or beat closed models on most coding-specific evaluations — including the ones that haven’t been gamed. More importantly, they run on a single GPU you can buy for under £2,000. A developer with a workstation can have a coding assistant that performs in the same league as Claude Sonnet, with zero per-token costs and zero data leaving their machine.
Multilingual performance. Open models trained with strong non-English data — particularly from China, the Middle East, and India — outperform GPT-5 and Claude on languages with smaller training-data footprints. If your business operates in Vietnamese, Thai, Arabic, or any of two dozen other languages, the open ecosystem now genuinely beats the frontier labs.
Embeddings and retrieval. The best open embedding models — bge-large, jina-embeddings-v3, nomic-embed — match OpenAI’s text-embedding-3-large on retrieval benchmarks while costing nothing to run. For RAG systems, search applications, and semantic similarity at scale, there’s no good reason to pay closed API fees in 2026.
Where the Gap Has Closed to Practical Equivalence
Several categories where open and closed models trade blows depending on the specific task:
General reasoning on standard benchmarks. The latest Llama, Qwen, and DeepSeek variants score within 2–5 percentage points of GPT-5 on MMLU, GPQA, and similar tests. That’s noise. On specific reasoning tasks an open model will sometimes win, on others a closed model will. Neither is consistently better.
Long-form writing. Closed models still hold an edge in nuanced creative writing — but barely. For business writing, summaries, email drafting, and technical documentation, open models are indistinguishable in blind tests. Most readers cannot tell which model wrote a piece of corporate copy.
Function calling and structured output. Once a major closed-model advantage, this is now broadly solved. Open models with the right prompting templates produce reliable JSON, call tools accurately, and work in agentic loops without breaking.
Where Frontier Labs Still Win Decisively
Three categories where the gap remains real and probably durable for now:
Agentic workflows over long horizons. When a task requires twenty or thirty sequential tool calls, planning, backtracking, and error recovery, frontier closed models are still meaningfully better. The reasoning over long contexts holds up. Open models tend to lose track of the task or fail to recover from a misstep. This is the gap that keeps Claude Code and similar agentic IDEs running on closed models.
Multimodal reasoning at frontier quality. GPT-5, Claude Opus, and Gemini handle screenshots, charts, diagrams, and video frames with a fidelity the open ecosystem hasn’t matched. Llava, Qwen-VL, and others are good — but for tasks like “read this 30-page PDF with embedded charts and tell me what the data shows”, closed models are noticeably more reliable.
Safety and refusal calibration. Whether you like the result or not, frontier labs have spent enormous effort on what their models will and won’t do. Open models tend to be either too cautious (refusing benign requests) or too permissive (helping with things they shouldn’t). For consumer-facing products where edge cases matter, this still drives many companies to closed APIs.
What’s Driven the Convergence
Three factors have closed the gap faster than most analysts predicted:
Knowledge distillation has worked. Teams have got very good at using frontier model outputs to train smaller open models. The frontier labs effectively do the expensive thinking; the open community packages it. This is legally murky and increasingly contested, but it’s been the technical engine behind much of the open ecosystem’s gains.
Architectures matter less than data. The dirty secret of 2025–26 is that most of the performance gains have come from better data curation, not novel architectures. Open labs can do data curation with grants from universities, foundations, and Chinese tech firms. They don’t need a £100m frontier compute budget to compete on this axis.
Inference compute has become the real cost. Training a frontier model still costs hundreds of millions. But the daily cost of running it for users has become dominant in the frontier labs’ P&L. Open models, run on local or cheap rented GPUs, sidestep this entirely. The economics have flipped against closed labs for high-volume use cases.
What This Means for Choosing a Model
The right framing in 2026 is not “open vs closed” but “what specifically am I doing”:
- Coding assistant for your own work? Open-source coder models on local hardware match the frontier and cost nothing to run.
- Powering a customer-facing chatbot? Closed APIs are still safer due to refusal calibration, even though the underlying capability is similar.
- Search, retrieval, embeddings? Open every time. The closed APIs are paying-for-the-brand.
- Agent that handles 30-step workflows? Closed models still genuinely better. Don’t fight it yet.
- Writing marketing copy at scale? Open models are now good enough that you should be running your own.
- Multimodal document understanding? Closed for now. Wait six months and reassess.
The Strategic Implication
If you’re a business spending more than a few thousand pounds a month on AI APIs, you should be running an open-source pilot in parallel. Not because open will be better in every dimension — it won’t — but because the categories where it’s already better are categories you almost certainly use. The question isn’t whether to add open-source AI to your stack. It’s which tasks to move first.
And for individuals: a £2,000 workstation with the right open coder model genuinely replaces a £20-per-month subscription for the way most people actually use AI. That math gets harder for the closed labs every month.
Key Takeaways
- Open-source AI now beats closed frontier models on specialist coding, multilingual tasks, and embeddings — three categories that cover most enterprise use
- The general-reasoning gap has closed to noise on public benchmarks; closed models still win on long-horizon agentic workflows and frontier multimodal tasks
- Knowledge distillation, better data curation, and the rising cost of inference have driven the convergence faster than most predicted
- The right framing isn’t open vs closed — it’s matching the model to the specific task; most workloads now have a credible open option
- Anyone spending serious money on closed APIs should be piloting open alternatives in parallel right now
AI Maestro covers AI without the vendor narratives. We test what we recommend.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
[newsletter_form]
