The Cheapest LLM APIs in 2026, by Task

The cheapest LLM API depends on what you send and what you expect back. A classifier that returns one word and a coding agent that writes 2,000 tokens per turn have different cheapest models. The price floor sits in two places as of October 2026. One is the vendors' own small models (GPT-6 Luna and Claude Haiku 5.5, both $0.10 per million input tokens and $0.50 per million output). The other is open-weight models on third-party hosts. There, OpenRouter lists gpt-oss-120b from $0.03 input and $0.15 output.

The table gives the option worth running at the lowest price for each task, plus a runner-up. Prices are USD per million tokens at the standard synchronous rate. They were read from the linked pages on 30 October 2026. Batch, cache and off-peak discounts come later, and they change the totals a lot.

On price alone, Ministral 3 3B ($0.10 / $0.10) also undercuts GPT-6 Luna and Haiku 5.5 for short chat. It is a 3B model, so run it on a sample of your own conversations before switching.

Three notes on the table.

Gemini 2.5 Flash-Lite is still on Google's price list at $0.10 / $0.40, which beats the two small models on output. It is a previous-generation model, so the table lists Gemini 3.1 Flash-Lite as Google's current small option instead.

GPT-6 Luna has a 1.05M-token context window at $0.10 / $0.50, which would make it the cheapest long-document model. OpenAI applies a long-context surcharge to GPT-6 Astra and GPT-5.6 Luna (2x input and 1.5x output above 272K tokens). The OpenAI models overview fetched for this article does not say whether GPT-6 Luna carries it too. Check that before you build on it. DeepSeek's flat rate up to 1M tokens is the safer bet today.

The vendors' own models aimed at coding start higher. Gemini 3.8 Flash costs $0.75 / $3.75 ($1.50 / $7.50 from 1 January 2027), GPT-6.1 Terra $1 / $5 and Claude Sonnet 5.5 $2 / $10. Codestral and DeepSeek V4.1-Flash set the floor for completions and small edits. That floor does not extend to long agent runs.

The cheapest output prices

Output tokens cost more than input tokens on almost every model, usually three to six times more. A workload that writes paragraphs pays for output first, so output price is the number to sort by. The chart shows the eleven lowest output prices from the vendor and host pages read for this article. Three models tie at $0.50. The chart includes GPT-6 Luna and Claude Haiku 5.5 and leaves out Z.ai's GLM 5.3 Flash as hosted by Mistral ($0.15 / $0.50).

Ministral 3 charges the same for input and output. No other vendor-owned model in the dataset does that. The same open weights also cost very different amounts depending on the host. Output for gpt-oss-120b costs $0.15 per million tokens through OpenRouter's cheapest provider, and $0.60 on Groq and on Together.

For scale, here is where the floor stood two years earlier.

What moves the bill

The list price is a rate. The bill is that rate times the tokens you send and receive, minus the discounts you claim. Three things move the bill more than the choice between two models at $0.10 input.

Output length

Output costs five times input on Luna and Haiku 5.5. A reply that runs to 600 tokens instead of 300 doubles the output line and nothing else. A max_tokens limit and a request for the answer without the preamble are the cheapest fix there is. Both work on every model.

Caching

Every major vendor now charges a reduced rate for input tokens that repeat a recent prompt prefix. Cache read prices for the small models, per million tokens:

The OpenAI pages fetched for this article do not show a cached-input price for GPT-6 Luna, so it is missing from the table. Gemini's storage charge adds up. A cache kept warm for a whole day costs $24 per million tokens held, which pays off only at high request rates. The prompt caching guide explains how to get hits.

Batch and off-peak

OpenAI, Anthropic and Google all take 50% off both input and output for requests you submit as a batch and collect within 24 hours. Mistral's plan page lists the same 50%. xAI gives 20%, and Amazon Bedrock takes 50% off Amazon's own models. Anthropic's batch discount stacks with caching. Gemini bills cache hits inside a batch at the standard caching rate.

DeepSeek lists no batch discount and prices by the clock instead. Peak rates apply in two weekday windows, 01:00 to 04:00 and 06:00 to 10:00 UTC (Chinese public holidays excluded). Outside those windows every DeepSeek price is 50% of peak, including cache hits. On weekdays the longest off-peak stretch runs from 10:00 to 01:00 UTC, and weekends are off-peak. A job run in the European afternoon or the American working day gets the off-peak rate. A European overnight job at 02:00 UTC does not. Batch APIs compared covers the three batch systems in detail.

The same job priced three ways

The job is one million requests a month. Each request sends 2,000 input tokens and receives 300 output tokens. Of the input, 1,500 tokens are a fixed system prompt and schema that can be cached. The table leaves out cache write costs, because at this volume the cache stays warm and they round to zero. Where batch applies, the cache-read rate stays at its standard price as a conservative assumption. Anthropic says its batch discount stacks with caching, so its real figure may be lower. Totals are rounded to the dollar.

The spread between models at list price is 2.7x. After discounts it is 2.3x, and each model ends up three to nearly four times cheaper than its own list price. The discounts matter more than the model choice at this tier.

Where cheap stops being worth it

A cheap model that is wrong more often costs more than the tokens it saves. The extra cost shows up in three places, none of them on the API invoice.

The first is retries. A reply that fails schema validation goes out again, often to a bigger model. You then pay for the cheap call and the expensive one, and wait out the latency of both.

The second is human review. A cheaper model saves a fraction of a cent per call. Two minutes of a person's time on one bad extraction cost more than thousands of those savings. If a cheap model raises the review rate even slightly, it loses.

The third is wrong answers that go through. They cost the most, and no dashboard shows them until a customer or an auditor finds one.

The test is a two-line comparison. Saving per call equals the price difference times the tokens per call. Cost of failure equals the failure rate times what one failure costs you (minutes of review, a refund, a re-run on a bigger model). When the second line is larger, the cheap model is the expensive option.

Routing gets you out of this trade. Send everything to the cheap model first, and escalate when a signal shows low confidence. Useful signals include a failed schema check and a refusal or hedge in the answer. Two cheap runs at different temperatures that disagree are another signal. Low log-probabilities are a third, where the API exposes them. Each vendor family has a clear ladder to escalate along:

Mistral's own pages disagree on Mistral Large 4. The documentation lists $0.70 / $2.10, and the plans page lists $0.50 / $1.50. The table uses the documentation figure.

If 80% of traffic stays on the cheap model and 20% escalates one rung, the blended price lands far below the big model's. A week of measuring the escalations tells you whether the failure rate also lands close to the big model's. Model routing covers the design.

Hosted open weights vs vendor small models

The cheapest row in the headline table is an open-weight model on a third-party host. That price comes with three differences from a vendor's own small model.

The first difference is that the host sets the price. The dataset shows the same weights at several prices on the same day:

Spreads of 5x and more for identical weights are common in this table, and Qwen3.8-27B reaches 28x on input. Budget for the whole range of prices.

Rate limits are the second difference. Groq's rate-limits page prints base limits for gpt-oss-120b of 30 requests per minute, 1,000 per day, 8,000 tokens per minute and 200,000 per day. The page does not separate free from paid columns. OpenRouter's free model variants allow 20 requests per minute and 50 per day, or 1,000 per day once you hold at least 10 credits. Cloudflare's free allowance is 10,000 Neurons per day. On the vendor side, OpenAI's Build tier (reached with $5 of credit) gives Luna 5,000 requests and 2 million tokens per minute. Anthropic's Start tier gives 1,000 requests and 2 million input tokens per minute. The pages fetched for this article did not list paid-tier limits for the open-model hosts, so ask before committing volume.

Data policy is the third. Google's pricing page marks free-tier content as used to improve its products and paid-tier content as not used. OpenAI's free-token programme is an explicit trade of traffic for tokens. The OpenRouter and Groq pages fetched for this article do not state a training policy. OpenRouter also routes to many underlying providers, so read the policy of the provider that serves your requests as well as the router's. If the data cannot leave your network at all, the comparison shifts to running the weights yourself. See Local LLM vs cloud API.

For most teams, hosted open weights win on price for traffic that can tolerate throttling or rerouting, and for data you would publish anyway. Vendor small models win when you need a rate-limit tier you can plan around. They also give you cache and batch from the same vendor, and a paid-tier data policy in writing.

What to do next

Before choosing a model, measure your real token split. That means input versus output, and how much of the input repeats. The winner on list price can lose once you know those two numbers.

Take the discounts first. Caching and batch cut the worked example's bill by two-thirds to almost three-quarters. Switching between small models changed it by far less.

Run the cheap model behind a confidence check and route failures up one rung. Measure the escalation rate for a week before you decide anything else.

Keep a dated copy of the prices you rely on. Google has already scheduled increases for 1 December 2026 (Gemini 4 Argon) and 1 January 2027 (Gemini 3.8 and 3.6 Flash). The October 2026 pricing snapshot has the full list.

Divero runs this comparison against a client's real traffic logs instead of list prices. The ranking can change once the real token split is in.

Sources

• Claude pricing
• OpenAI models
• OpenAI GPT-5.6 Luna model page
• OpenAI text-embedding-3-small
• Gemini API pricing
• DeepSeek API pricing
• Mistral AI Studio pricing
• OpenRouter: openai/gpt-oss-120b
• Groq models
• Voyage AI pricing