Two years ago the answer was easy: if you wanted serious AI in your product, you paid OpenAI, Anthropic or Google per API call, and that was that. In 2026 the answer is genuinely contested. Open-weight models — from OpenAI’s own gpt-oss releases to Qwen, Kimi, GLM and Mistral — are now close to the top closed models on many everyday tasks. So the question lands on our scoping calls weekly: should we switch to open-source and stop paying per call? Here is the honest answer.
What changed
The open ecosystem stopped being a science project. In August 2025 OpenAI released gpt-oss — its first open-weight models since GPT-2, under a permissive license — and the Chinese labs (Qwen, Kimi, GLM) kept shipping models with strong multilingual and tool-use capability. By 2026, for a large share of business tasks — summarising, drafting, extraction, classification, routine support answers — a good open model is simply no longer the weak option.
What has NOT changed: the frontier closed models still lead on hard reasoning, long multi-step agent work, and the messy edge cases. The gap shows up exactly where mistakes are most expensive.
The real comparison is not model vs model — it is bill vs bill
Hosted APIs charge per call and require zero infrastructure. Open models are "free" the way a free puppy is free: you now own hosting, GPUs or inference services, updates, security patches and an engineer who understands the stack. Our rule of thumb from client work:
- Low or spiky volume? Hosted API wins, almost always. If your monthly model bill is under a few hundred dollars, self-hosting cannot beat it — the infrastructure alone costs more.
- High, steady volume on repetitive tasks? Open models start to win. Document extraction, classification, template drafting at tens of thousands of calls a month — this is where teams cut their AI bill dramatically by moving the routine work to an open model.
- Strict data residency? If data legally cannot leave your servers or your country, self-hosted open models may be the only real option — that decision is made for you.
The pattern that actually wins in 2026 is hybrid: an open model for the cheap, repetitive 80% of calls, and a frontier hosted model for the hard 20% — behind an abstraction layer so you can reroute traffic without rewriting the product. This is how we architect every AI build: model-agnostic, because the leaderboard changes every few months and your codebase should not care.
Where switching backfires
- Underestimating ops. A self-hosted model that goes down at 2 AM is your outage, not a vendor status page. If nobody on the team owns infrastructure, the "savings" buy you downtime.
- Benchmark blindness. A model that matches GPT on a benchmark can still fail on YOUR data — your invoices, your Hinglish support messages, your legal templates. We always run a two-week bake-off on real traffic before any migration decision.
- Forgetting the safety layer. Hosted APIs ship with moderation endpoints and abuse controls. With an open model, guardrails, rate limits and content moderation are entirely your job — skip them and the first bad output is on you.
- Quiet quality drift. Without evals, nobody notices the open model doing slightly worse until customers do. Budget for a small evaluation harness — it is the insurance policy of a migration.
Our honest bottom line
For most small and mid-size businesses shipping their first AI feature, start on a hosted API. It is faster, safer and cheaper at low volume — and the running costs are exactly what we quote upfront in every project (we broke the numbers down in our AI development cost guide). Revisit open models when your monthly bill crosses a few hundred dollars on repetitive workloads, or when data residency forces your hand. And whatever you pick, insist on an abstraction layer — the businesses hurting in 2026 are the ones welded to a single provider they chose in 2023.