Decision guide
Best AI Model by Use Case: Price, Context and Open Weights
Which AI model to pick for cost at volume, long context, image input, open weights or quality — prices checked 22 November 2026.
Picking a model is three questions in a fixed order: what does a million output tokens cost, how much context do I actually need, and do I have to own the weights? Answer those and most of the field drops out. What follows is the short list for each answer, using the prices in our model index, all checked 22 November 2026. If the question is purely price, the cheapest LLM API breakdown prices the cheap tier end to end on three real workloads; if it is code specifically, the best LLM for coding ranks the tiers by what a real refactor costs.
Cheapest output tokens: Inkling 2 Mini
$0.50 per million output tokens. The next cheapest model here, Mistral Large 4, is $2.10 — 4.2 times more — and the cheapest closed model, Qwen4-Max, is nine times more at $4.50. With open weights, a 1.05M context window and 3.4B active parameters, it is the model to put behind a high-volume feature: summarising support tickets, tagging documents, drafting first passes, where the output is checked by a human or a second model anyway.
The catch is not in the price. It is a small active-parameter model, so on a task that needs deep reasoning you will spend the savings on retries. The pattern that works: run Mini first, escalate to a frontier model only on the requests that fail.
Cheapest closed model: Qwen4-Max
$1.50 input / $4.50 output with a 1M context window and image input, released 16 November 2026. Its output price is 30% of Gemini 4 Neon's $15, the lowest of any closed model in the index, and Alibaba publishes average cost per task falling alongside it. The weights are closed, so if open weights matter, see below.
Images and documents: Mistral Large 4
$0.70 input / $2.10 output (list rate $1.36 / $4.18) with a 1M window and native text and image input, in public preview since 6 October 2026 with the weights already released. For document, chart and screenshot pipelines this is the obvious starting point — it reads images without a separate OCR step, at under half Qwen4-Max's output price. One caveat repeated deliberately: a preview price is not a promise, so budget against the list rate before you commit a workflow to it.
Longest context: the 1.05M group
GPT-6.2 Astra, Claude Fable 5.2, Claude Opus 5.5, Gemini 4 Neon, Gemini 4 Hydrogen, Mistral Large 4 and Inkling 2 Mini all list a context window of about 1,048,576 tokens. That is roughly 750,000 words — whole codebases, whole contract sets, a year of email. Every price tier is represented in that group, from Inkling's $0.50 output to Astra's $40, so a long window alone no longer buys any pricing excuse.
Two warnings. A long window is not the same as good long-context recall; models degrade in the middle of very long inputs, and the only way to know is to test on your own documents. And a long prompt is billed as input on every single call unless the provider supports prompt caching — check the cached-input rate before you build a workflow around a 900K-token system prompt.
Open weights: DaisyRose 27B, Inkling 2 Mini and Mistral Large 4
If data cannot leave your network, or you want to fine-tune, these are the three in the table. Mistral Large 4 (1.05T total / 52B active, 1M context, text and image input) is the strongest of the three and the only one with multimodal input; DaisyRose 27B (Apache 2.0, 262K context) is the better fit for multi-step coding and reasoning on modest hardware; Inkling 2 Mini is the better fit for volume and very long inputs. Self-hosting turns a per-token bill into a GPU bill — cheaper at high steady volume and more expensive at low or spiky volume.
Best all-round balance: Gemini 4 Neon
$3 input / $15 output with a 1.05M window and image input. It is the cheapest of the three frontier-priced models and the one with the widest window relative to its price. If you want a single default model and do not want to build a routing layer, this is the least bad place to start.
When quality is the only thing that matters: GPT-6.2 Astra or Claude Fable 5.2
$8 / $40, the top of the table by output price. It only makes sense on tasks where a wrong answer costs more than the tokens — hard engineering, scientific reasoning, visual analysis. Astra Pro is listed at the same price; treat the two as one tier with different routing rather than as a cheaper and a dearer option. Fable 5.2 is Anthropic's own entry at this tier — same price, 1M context, 128K max output, always-on thinking at high effort by default — so the choice between them is which lab's failure modes your task punishes less, not which one is cheaper.
The decision table
| If your constraint is… | Start with | Why |
|---|---|---|
| Output cost at volume | Inkling 2 Mini | $0.50/1M output, open weights |
| Images and documents | Mistral Large 4 | $2.10/1M output with image input, open weights |
| Cheapest closed model | Qwen4-Max | $4.50/1M output, 1M context, image input |
| Very long inputs | Gemini 4 Neon | 1.05M context at $3/$15 |
| Data cannot leave your network | DaisyRose 27B | Apache 2.0, self-hostable |
| One model, no routing layer | Gemini 4 Neon | Cheapest frontier-tier with full window |
| Correctness over cost | GPT-6.2 Astra / Claude Fable 5.2 | Hardest-task tier |
What we are not claiming
None of the above is a quality ranking. The prices are measured; "best for coding" is a judgement built on what each provider positions the model for, and on the price and context that comes with it. Run your own evaluation on 50 real tasks before you commit — it takes an afternoon and it beats any leaderboard.