Published 2026-08-12 · reference numbers, verify before budgeting
Run the same content-moderation job — 150 input tokens of context, a 5-token label out the other end — through nine different "cheap" models and the monthly bill lands anywhere from $59.50 to $1,750. Same job, same volume, same output. A 29x spread, and every one of these nine models gets pitched to you as "the budget option" depending on which vendor's docs you're reading.
Pulled from our tracked pricing table (387 models): every model priced under $1/1M input tokens that a provider currently markets for high-volume, low-latency tasks — classification, extraction, moderation, routing. 10 million calls a month, 150 input tokens and 5 output tokens each, so 1.5B input tokens and 50M output tokens total.
| Model | Provider | Input /1M | Output /1M | Monthly (10M calls) |
|---|---|---|---|---|
| Nova Micro | Amazon | $0.035 | $0.14 | $59.50 |
| Ministral 3B | Mistral | $0.040 | $0.04 | $62.00 |
| Command R7B | Cohere | $0.0375 | $0.15 | $63.75 |
| Gemini 3.1 Flash-Lite | $0.050 | $0.20 | $85.00 | |
| GPT-5 nano | OpenAI | $0.100 | $0.40 | $170.00 |
| GPT-4.1 nano | OpenAI | $0.100 | $0.40 | $170.00 |
| GPT-4o mini | OpenAI | $0.150 | $0.60 | $255.00 |
| Claude Haiku 3.5 | Anthropic | $0.800 | $4.00 | $1,400.00 |
| Claude Haiku 4.5 | Anthropic | $1.000 | $5.00 | $1,750.00 |
| Reference pricing from our tracked table, 387 models. Worked example: 10M calls/mo × 150 input + 5 output tokens. Price your own volume on the classification cost calculator. | ||||
Nova Micro, Ministral 3B and Command R7B are within 7% of each other and none of them are the household names in this list. The two OpenAI nano SKUs sit at almost 3x Nova Micro. GPT-4o mini — probably the most-reached-for "cheap" model by brand recognition alone — is already 4.3x Nova Micro. And Claude's Haiku line, marketed as Anthropic's budget tier, is priced closer to some mid-tier chat models than to anything else on this table: 29.4x Nova Micro on input, 35.7x on output.
Every provider ships a cheap model because every provider needs an answer to "what do I run at a billion calls a month." But cheap is relative to each vendor's own flagship, not to the market. Anthropic's flagship-to-Haiku gap and Amazon's flagship-to-Nova-Micro gap are both real discounts within their own catalogs — they just don't start from the same place. Nova Micro was built by a cloud provider that also sells the compute underneath it, so the model can be priced closer to marginal cost. Claude Haiku inherits pricing gravity from being positioned next to Sonnet and Opus, where "10x cheaper than our flagship" still lands at $1/1M because the flagship itself is expensive.
None of this means Haiku is a bad model for a classification job — it's frequently the more accurate one, and accuracy is worth paying for when a wrong label costs more than the token difference. It means "budget tier" is a claim about a vendor's own lineup, not a claim about the market, and the two get conflated constantly in pricing pages and in comparison threads.
For a genuinely high-volume, low-complexity task — spam filtering, intent routing, PII detection, sentiment tagging — start from Nova Micro, Ministral 3B or Command R7B as the price floor, not from whichever brand you already have an API key for. Run a real accuracy check against your labeled data before committing; a 29x price gap is only a good trade if the cheap model's error rate doesn't eat the savings in downstream cost (mis-routed tickets, missed moderation flags, bad extractions that need human review). If accuracy on your specific task doesn't hold up below GPT-4o mini or Haiku, that's the actual price of the task — not $59.50, and not $1,750 by default either. Price your own token mix and volume on the classification cost calculator or the content moderation cost calculator, and compare full lineups on the cheapest LLM API page.
Methodology: prices from our tracked pricing table (387 models) as of August 2026, matching each provider's published per-1M-token rates. The worked example uses a constructed 150-in/5-out token ratio typical of short classification calls, not telemetry from a production system — confirm your own workload's token mix and each provider's current rate before budgeting.