Last month I did something slightly embarrassing: I audited my own AI bill and realized I had been running a $20-per-million-token model to reformat CSV files. A task that a $0.18 model does just as well.
That is a 114x price gap for the exact same output. And it is the single most useful thing I have learned about building with LLMs in 2026: the model you reach for by default is almost never the one the task needs.
I pulled the standard list API prices for 17 major models across OpenAI, Anthropic, Google, xAI, and DeepSeek (June 2026, from each provider's own pricing page). Here is the spread that nobody puts on a slide:
| Model | Provider | Blended $/1M tokens |
|---|---|---|
| DeepSeek V4 Flash | DeepSeek | $0.18 |
| Grok (cheap tier) | xAI | $0.28 |
| Gemini Flash | $0.56 | |
| GPT (mini tier) | OpenAI | $0.69 |
| Claude (mid) | Anthropic | $8.00 |
| Claude Fable 5 | Anthropic | $20.00 |
(Blended = a 3:1 input-to-output ratio, so everything ranks on one number. Batch and cached-input discounts can knock 50 to 90 percent off, but the relative spread holds.)
Median blended cost across all 17 is about $2 per million tokens. The frontier models are 10x that. So the question is never "which is the best AI." It is "which is the cheapest model that clears the bar for this task."
The 4 questions I actually run
I stopped trying to memorize model names (they change monthly anyway) and started running the same four questions on every task. This is the whole decision tree.
1. Does this task fail loudly or silently?
If a wrong answer is obvious and cheap to catch (reformatting, extraction, boilerplate, first drafts you will edit anyway), use the cheapest capable model. DeepSeek Flash and Gemini Flash are stupidly good at these now. If a wrong answer is subtle and expensive (legal reasoning, security-sensitive code, medical or financial content, anything a human will trust without checking), pay for the frontier model. The cost of one bad silent answer dwarfs the API savings.
2. How long is the context?
Long documents are where the "cheap" models quietly get expensive, because you pay per token and a 200-page PDF is a lot of tokens. Here the math flips: a model with strong long-context handling and a lower input price often beats a "smarter" model that needs multiple passes. Check the input price specifically, since long-context tasks are input-heavy and output prices (a median 6x higher than input) matter less.
3. Is this interactive or batch?
If a human is waiting on the response, latency is a feature and you pay for speed. If it runs overnight or in a queue, use the batch API (roughly half price everywhere) and a cheaper model. I moved every non-interactive job to batch and my bill dropped more than any model swap ever did.
4. Will you send the same prompt 10,000 times?
If the system prompt or context is stable across calls, cached input is often 90 percent off. That changes the entire calculation. A model that looks expensive per call can be the cheapest at scale once the cache kicks in. Almost nobody factors this in.
The honest per-provider read
- DeepSeek is the price floor. Median blended around $0.36. For high-volume, fail-loudly work it is hard to argue with.
- xAI (Grok) and Google (Gemini Flash) sit in the cheap-but-capable band, roughly $0.5 to $1 blended. My default for "good enough, high volume."
- OpenAI spans the widest range (mini tiers under $0.70, frontier over $11). You can run the whole ladder on one API, which is underrated for switching models per task.
- Anthropic is where I go when the task fails silently. Median blended $8, up to $20 for Fable 5. Expensive, and worth it exactly when a wrong answer would cost more than the tokens.
The trap is treating any one of these as "my model." The people spending smart in 2026 route per task: Flash for the grunt work, frontier for the 5 percent that actually matters.
A shortcut if you don't want to memorize this
I turned this exact logic into a free tool. You answer three questions (task type, budget, and what you care about most: quality, cost, speed, or privacy) and it ranks the top 3 models for that specific job, with an honest one-line "why" and the real API cost. No signup, no login, no email wall:
It is built on the same pricing data above (all sourced from provider pages, rebuilt monthly). If you catch a stale number, tell me and I will fix it.
The one-line takeaway
Do not ask "what is the best AI model." Ask "what is the cheapest model that clears the bar for this task, at this context length, at this volume." Run those four questions and you will cut your bill without noticing a quality drop, because the drop happens on the tasks where quality was never the constraint.
I write up more of these honest, tested breakdowns (which AI tools are actually worth it, which are hype) in a small daily AI channel on Telegram if that is useful: t.me/aitoolsinsiderhq. No spam, just what I find that works.
What is your routing rule? I am curious whether people default to one model or actually switch per task. Tell me in the comments.
Top comments (0)