Cheapest LLM API Finder
Sort and filter the whole public model catalogue by what it actually costs you: input price, output price, a blended price weighted to your own input:output mix, and the context window that decides whether your documents fit at all.
Runs entirely in your browser — nothing you type leaves the page.
Blended = (in × r1 + out × r2) ÷ (r1 + r2)
Loading live prices… · 0 of 0 models shown
Loading live prices…
Prices are the lowest rate among the providers serving each model, in USD per 1,000,000 tokens, and exclude image, audio and search fees. Direct-provider list prices can differ.
Snapshot: ten of the cheapest models right now
Static snapshot dated 2026-10-18, rendered into the HTML so it is readable without JavaScript. The interactive table above fetches the current catalogue.
| Model | Provider | Input $/1M | Output $/1M | Blended $/1M | Context |
|---|---|---|---|---|---|
| Mistral: Mistral Nemomistralai/mistral-nemo | Mistral | $0.029 | $0.030 | $0.029 | 131.1K |
| ScaleDown: Extractscaledown/extract | Scaledown | $0.040 | $0.00 | $0.030 | 1M |
| ScaleDown: Summarizescaledown/summarize | Scaledown | $0.040 | $0.00 | $0.030 | 1M |
| inclusionAI: Ling 3.0 Flashinclusionai/ling-3.0-flash | Inclusionai | $0.021 | $0.063 | $0.032 | 262.1K |
| OpenAI: gpt-oss-20bopenai/gpt-oss-20b | OpenAI | $0.018 | $0.090 | $0.036 | 131.1K |
| IBM: Granite 4.0 Microibm-granite/granite-4.0-h-micro | Ibm Granite | $0.017 | $0.112 | $0.041 | 131K |
| Sao10K: Llama 3 8B Lunarissao10k/l3-lunaris-8b | Sao10k | $0.040 | $0.050 | $0.043 | 8.2K |
| Nex AGI: Nex-N2.5-Mininex-agi/nex-n2.5-mini | Nex Agi | $0.025 | $0.100 | $0.044 | 262.1K |
| DeepSeek: DeepSeek V4 Flash Latest~deepseek/deepseek-v4-flash-latest | DeepSeek | $0.012 | $0.144 | $0.045 | 1.0M |
| OpenAI: gpt-oss-20b (batch)openai/gpt-oss-20b:batch | OpenAI | $0.024 | $0.112 | $0.046 | 131.1K |
How this comparison is built
Every row comes from the public OpenRouter catalogue — the same feed that routes live API traffic — fetched by your browser, cached locally for twelve hours, and never sent anywhere. Prices are converted from the per-token strings the feed publishes into dollars per million tokens, the unit every provider uses on its own pricing page. The conversion is rounded to six decimals, because 0.0000001 × 1,000,000 is 0.09999999999999999 in floating point, and a price column that reads that way discredits every other number on the page.
Ranking models by input price alone is the most common mistake in this comparison. A model at $0.10 per million input tokens looks unbeatable until you notice it charges $2.50 per million for output, and a chat assistant writes far more output than the person reading it types. The blended column makes that trade-off visible: it collapses two prices into one number weighted by how your requests are actually shaped.
- Input price — what the prompt costs, including the system message and any retrieved context you paste in.
- Output price — what the completion costs; usually the dominant term for chat, and negligible for embeddings and classification.
- Blended price — one ranking number at your chosen input:output ratio.
- Context window — the maximum prompt plus completion the model accepts. A model that cannot hold your document is not cheap, it is unusable.
Provider prices change every few weeks and the feed changes with them. The line above the table always states the source and how long ago it was fetched, and a refresh button is there when the cache is older than you are comfortable with. The built-in reference list is only used when the live feed is unreachable — and the page admits it when that happens.
Reading the table: what each column really decides
The cheapest line in a price table is rarely the cheapest model in production. Three things move the total more than the sticker price does: how much text you send, how much text comes back, and whether the beginning of your prompt repeats often enough to be cached.
| Workload shape | What dominates the bill | What to sort by |
|---|---|---|
| Chat assistant, short prompts | Output tokens, by a wide margin | Output price |
| RAG over long documents | Input tokens — every retrieved chunk is billed again | Input price, then context window |
| Classification or extraction | Input tokens; the answer is a few tokens | Input price |
| Batch summarisation | Both, roughly balanced | Blended price at 1:1 |
| Agent loops with tools | Input, replayed on every step | Input and cache-read price |
Prompt caching is the biggest lever and it is not a column here: it is the cache-read price, published for most of the catalogue and typically a tenth of the fresh input price. If your requests start with the same long system prompt, your effective input price is much closer to that figure than to the headline one. This table deliberately shows the uncached rate so the comparison stays like for like; the cost calculator on this site does the caching arithmetic from your own cached share.
Models whose id ends in :free really do cost $0.00 per token today, which is why the free toggle exists — but they are rate limited, they can be withdrawn without notice, and their capacity is not a service level you can build a product on. Treat them as a way to prototype, never as a line in a forecast.
Worked example: picking a model for a support bot
Suppose the bot answers 20,000 messages a month, and each one is a 1,500-token prompt (system instructions, three retrieved help-centre articles, the conversation so far) with a 300-token reply. That is 30 million input tokens and 6 million output tokens per month. Set the blended ratio to 5:1 to match that shape, then read the blended column as the ranking for this workload rather than for a generic one.
Run the same arithmetic for a frontier model at $3 and $15 per million and the identical traffic costs $99 a month — twelve times more; swap in a small model at $0.10 and $0.40 and it costs $5.40. The table does not answer which of those you should ship. It prices the experiment, so you can run the expensive model on a sample, measure quality against the cheap one, and know exactly what the difference costs per month before you commit.
What these numbers do not include
A price table is a lower bound, and it is honest to say so. Four things sit outside it:
- 1Routing differences. The catalogue lists the lowest price among the providers serving a model; billing the same request to one specific upstream can cost more.
- 2Non-token charges. Some models bill per image, per audio second, per web search or per request. Those fees live in the same pricing object, are usually zero, and never appear in a per-token comparison.
- 3Tiers and minimums. Long-context tiers, priority tiers and batch discounts change the effective rate; the table shows the standard tier only.
- 4The cost of being wrong. A model that fails the task costs retries and engineering time. Cheapest per token is a starting filter, not a decision.
Frequently asked questions
- Are the prices on this page live?
- Yes. Your browser fetches the public OpenRouter model catalogue — no key, no account — and caches it locally for twelve hours. The line above the table states the source and how long ago the data was fetched, and you can force a refresh. If the feed cannot be reached, the page falls back to a hand-checked reference list and says so instead of implying the numbers are current.
- Which LLM API is actually the cheapest?
- It depends on the shape of your traffic, which is exactly why the blended column exists. At the default 3:1 ratio the cheapest paid models are small, fast ones, usually between $0.02 and $0.10 per million input tokens. For output-heavy work the ranking changes by a factor of ten or more, so set the ratio to match your own input:output split before trusting the top row.
- What does "blended price" mean?
- It is the weighted average of the input and output price: (ratioIn × input + ratioOut × output) ÷ (ratioIn + ratioOut). With the default 3:1 weighting and a model at $1 per million in and $2 per million out, the blended price is $1.25 per million tokens. It gives one number that ranks models fairly across different mixes, and it is the same arithmetic the cost calculator performs once you give it real token counts.
- Why are the free models hidden by default?
- Because "$0.00" wins every sort and would bury the comparison you came for. Free variants are the same weights served for evaluation, with strict rate limits, no uptime promise and no notice before they are withdrawn. They are useful for prototyping, so the toggle is one click away — but a table whose first ten rows all cost nothing tells you nothing about the market.
- Why do these prices differ from the provider's own pricing page?
- The catalogue lists the lowest price among the providers currently serving each model, while an AI lab's own page quotes the price of going direct with your own account, including any committed-spend discount. Both figures are true. Use this table to compare models, then confirm the exact rate on the route you will actually be billed against.
- How often do LLM prices change?
- Frequently. New generations usually arrive cheaper per token than the ones they replace, and existing models are regularly discounted or split into tiers. That is why every tool here refreshes at most once per twelve hours and prints the age of what it is showing. If a forecast must not move under you, snapshot the CSV export and record the date with it.
- Does the table include prompt caching, batch discounts or priority tiers?
- No — it deliberately shows the standard, uncached rate for every model so the comparison is like for like. Cache-read prices are published for most models and are typically a tenth of fresh input, and batch or priority tiers appear as variants of the same model. The cost calculator has a cached-share field precisely so you can model those savings on top of this comparison.
- Can I export the comparison?
- Yes. The copy button puts the currently filtered and sorted table on your clipboard as CSV — model name, API id, provider, input, output and blended price per million tokens, context window and maximum output. It respects your filters, so narrow to a provider and a context floor first if that is the shortlist you want to share.
Learn this in the course
This tool is the hands-on companion to these stages of the free 11-stage track.
Related tools
- LLM API Cost Calculator →
Turn a token count and a traffic guess into a real monthly bill. Set the prompt size, the answer size and how many requests you expect, and every priced model is compared side by side.
- AI Token Counter →
Paste any text to see exactly how a language model will split it into tokens — the units that decide what you pay and how much of the context window is left. The counter runs entirely in your browser.