The Cheapest LLM APIs, Compared (And Why the Ranking Flips)
Current per-token pricing across every major provider — and why the cheapest one changes with your workload.
This is part of our guide to controlling your LLM bill.
“Which LLM API is cheapest” has no single answer, and the reason is structural: providers charge different rates for input tokens and output tokens, apply different surcharges for long context, and layer discounts for caching and batch processing on top. A model that wins on a short, input-heavy classification task can lose badly on a long, output-heavy generation task. The cheapest API for one workload is the wrong choice for another.
This is a snapshot of current rates and, more usefully, a guide to reading them. Prices below are per million tokens and current as of August 2026. The LLM pricing market moves fast — providers have cut rates multiple times a year, and models are launched and retired on short notice — so treat every number here as a starting point to confirm against the live pricing page, not a figure to hard-code into a budget.
The current landscape
Rates are USD per million tokens, input / output, on the standard tier.
| Model | Input | Output | Notes |
|---|---|---|---|
| Groq — Llama 3.1 8B | $0.05 | $0.08 | Open-weight, LPU inference, very high throughput |
| OpenAI — GPT-5.6 Luna | $0.20 | $1.20 | Budget tier of the current flagship family |
| DeepSeek — V4 Flash | $0.14 | $0.28 | Frequently cited price floor for a capable model |
| Google — Gemini 3.5 Flash-Lite | $0.30 | $2.50 | 1M context on every tier, including this one |
| Mistral — Small | ~$0.10–0.20 | ~$0.30–0.60 | Budget option with EU data residency |
| OpenAI — GPT-5.4 | $2.50 | $15.00 | Common general-purpose production model |
| Google — Gemini 3.6 Flash | $1.50 | $7.50 | Workhorse mid-tier |
| Google — Gemini 3.1 Pro | $2.00 | $12.00 | Steps up to $4 / $18 above 200K tokens |
| Anthropic — Claude Haiku 4.5 | $1.00 | $5.00 | Low-cost Claude tier |
| Anthropic — Claude Sonnet 5 | $2.00 | $10.00 | Introductory rate through Aug 31, 2026; then $3 / $15 |
| Anthropic — Claude Opus 5 | $5.00 | $25.00 | Flagship; 1M context at standard price, no surcharge |
| OpenAI — GPT-5.6 Sol | $5.00 | $30.00 | Flagship tier |
The absolute floor for a hosted, capable model sits around $0.05–0.30 per million input tokens, on open-weight models served by Groq, DeepSeek, and Mistral. The frontier tiers from OpenAI and Anthropic sit roughly 25–100× higher on output. That spread is the whole story: picking the right tier per task matters far more than squeezing a better rate out of any single provider.
Why the headline rate lies
Four things distort the “cheapest” ranking, and each one can flip it.
Output costs several times more than input. Across nearly every provider, output tokens are priced 3–10× higher than input. A workload that reads a lot and writes a little (classification, extraction, routing) is priced almost entirely off the cheap input rate. A workload that writes a lot (long-form generation, verbose reasoning) is priced off the expensive output rate. The same model can be a bargain for one and costly for the other. This is why a blended “average price” hides the number that actually matters.
Long context carries a surcharge — on some providers. Several models raise their rate once a single prompt crosses a threshold. Gemini Pro roughly doubles above 200K tokens. This is where a structural detail changes the ranking: Claude’s current models include the full 1M-token context window at standard pricing with no long-context surcharge tier at all. On a short prompt, Claude’s headline rate looks higher than a Flash model’s; on prompts that routinely exceed 200K tokens, the absence of a surcharge can make it cheaper in practice. Headline rates can’t see this.
Caching and batch stack on top. Prompt caching discounts a repeated prefix by up to 90 percent, and batch processing takes a flat 50 percent off asynchronous work on every major provider. A model that looks mid-priced can become the cheapest option for a workload that caches heavily or runs in batch — discounts the headline table never reflects.
Tokenizers differ. The same text does not produce the same number of tokens on every model. Newer tokenizers can generate meaningfully more tokens for identical input, which raises the effective cost even when the per-token rate looks lower. Two models at the same headline price can bill differently for the same prompt.
How to actually pick
The reliable method is to match the tier to the task rather than searching for one cheapest model:
High-volume, simple work — classification, routing, extraction, short summaries — belongs on the floor tier: Groq’s open-weight models, DeepSeek, Gemini Flash-Lite, or Mistral Small. These stay cheap even without caching or batching layered on.
Mid-tier production traffic — where accuracy, context, or tool use matters — lands on models like Gemini Flash, GPT-5.4, or Claude Sonnet. The right one depends on your input/output ratio and whether long context is common.
Hard reasoning and complex work — the minority of requests that genuinely need it — escalates to a flagship tier. The mistake that inflates most bills is sending all traffic to a flagship because a fraction of it needs one.
Most cost-efficient deployments route across several tiers rather than standardizing on one. Cheap models absorb the volume; premium models are an escalation path, not a default.
The number that decides it is your own
Every ranking here is built on list prices and assumed token ratios. The cheapest model for a given application depends on that application’s real input/output split, its real prompt sizes, its caching hit rate, and its tokenizer — none of which appear on a pricing page. Two apps calling the same model can see costs that differ by an order of magnitude.
That’s why the only conclusive comparison uses your own token distribution. LLMtrack calculates what current usage would cost on every cheaper alternative using real average token counts per request — not benchmarks — and ranks the alternatives by projected monthly saving. Start tracking free to see which model is actually cheapest for the requests your app is really sending.
Pricing referenced in this article is current as of August 2026 and provided for comparison only. Always confirm current rates on each provider’s official pricing page before making a budgeting or migration decision.