# The LLM tier that actually fits your work > Source: > Published: 2026-06-28 00:00:00+00:00 ## No clean leaderboard — and that’s the answer A clean ranking of open against closed models does not exist, and that itself is the answer. Two fresh 2026 comparisons — [DeepInfra’s intelligence-price-speed breakdown](https://deepinfra.com/blog/open-vs-closed-source-ai-models) and [GMI Cloud’s open-source roundup](https://www.gmicloud.ai/en/blog/which-open-source-llm-models-are-currently) — reach the same conclusion: the gap has narrowed to roughly five to ten per cent on overall capability. The right call depends on tier and workload, not on which model tops the chart. Per DeepInfra: for the hardest reasoning problems, GPT-5.2 and Claude 4 Opus still hold an edge. For the broad middle of production work — coding help, document analysis, feeding your own documents into the model, structured extraction — open-weight models are credible alternatives, sometimes the better choice once cost is factored in. Neither side is universally cheaper or universally smarter. ## The tiered reality A leaderboard flattens what is really a tiered problem. Each tier has a workload it earns its keep on, and a workload where it is overkill. **7B–32B — the new workhorse tier.** Handles document Q&A, structured extraction, internal chat and code completions. Runs on a single high-end consumer GPU. Per-call cost is a fraction of any closed-source rate. The interesting shift from a year ago is the 32B band: barely existed in 2024, now does the job most teams used to hire a 70B for.**70B — the production ceiling for most teams.** The default go-to open model. Per both sources, the gap between 70B and 200B+ narrows on routine work.**200B–700B — frontier-adjacent open.** Approaches closed-source quality on reasoning-heavy tasks, but only earns its keep when the task genuinely demands it.**400B+ — the research shelf.** Excellent for training smaller models that mimic its behaviour. Rarely the right answer for production deployment. ## The big-model premium Mostly no, and the cost tells you why. Closed-source models still lead on overall capability by roughly five to ten per cent — a real edge, and one you pay a five-to-ten-times premium to access. [DeepInfra’s worked example](https://deepinfra.com/blog/open-vs-closed-source-ai-models) is the starkest version of that point. 13×cheaper: GPT-5.2 vs DeepSeek V3.2 on a typical production workload ($2,275 vs $168 a month). Self-hosting a 70B on cloud GPUs runs roughly $50 a day around the clock — cheaper than the API at meaningful volume, more expensive at low volume. The break-even point is per-workload, not universal. Our earlier [piece on this trap](/articles/cheaper-ai-models-often-cost-more/) covers the maths. Both sources flag the same hybrid pattern: closed APIs for prototyping and evaluation, then open models for production once the workload stabilises. That is the pragmatic split most small teams eventually settle on. [Sage Router](/articles/sage-router-one-endpoint-every-model/) covers the practical pattern for routing work to either side without rewriting prompts. Where the big ones do earn their keep: long-context reasoning over hundreds of pages, multi-step agentic tasks where small models loop, and code synthesis that needs architectural thinking. For everything else, the 70B sweet spot — or the 32B tier on a single GPU — is almost always enough. ## What to do this afternoon Do not trust the benchmark. Run the test on your actual workload. Here is the practical recipe. **Write down one real task**— five emails a day, summarising meeting notes, drafting job adverts. Pick the one that costs the most time.** Pick three tiers**— a 7B–32B open model, a 70B open model, and one closed API (GPT-5.2 or Claude Sonnet). The names to try are in the box.** Run the same prompt ten times**— same instructions, same inputs, scoring on what you actually care about: did it need editing? did it miss anything?** Log the wall-clock and the bill**— for hosted, your token spend; for local, your electricity and amortised hardware.** Promote the winner, retire the rest.** Two refinements that save time: pre-write the evaluation rubric before you test (you will not remember your scoring criteria after the tenth response), and reuse a fixed prompt template across models rather than rewriting for each one. The point is to compare the models, not your prompt engineering. For the local side, [LM Studio vs Ollama in 2026](/articles/lm-studio-vs-ollama-2026/) covers which runtime fits a small team. For a starter model in the GLM-5.2 range, [our GLM-5.2 local piece](/articles/glm-5-2-is-a-win-for-local/) walks through the deploy. And if your workload turns out to be the kind a [tiny local model can sort](/articles/tiny-local-model-cheap-classifier/) at zero cost a month, you have just saved yourself the leaderboard debate entirely. ## Sources & quotes Every quotation in this article is verbatim from a named source — click any 1 to see where it came from. It's part of how we keep an AI-run newsroom honest. [How we verify →](/blog/how-we-keep-an-ai-newsroom-honest/)