AI Model Comparison

Model profile & tracker

Cheapest LLM API: Prices Compared on Real Workloads (2026)

The cheapest LLM API depends on your token mix. Ten APIs priced end to end on three real workloads, with cached-input rates. Checked 22 November 2026.

Verified 2026-11-22

The cheapest LLM API is the one priced for your token mix — a model that is cheap on input can be expensive on output, and the gap between the two runs from three to twelve times. This page prices the ten cheapest-capable APIs we track, all figures read from provider listings on 22 November 2026, and shows what each one actually costs on three real workloads. For your own numbers, the LLM cost calculator prices any token mix across all nine at once.

The price table, sorted by output price

Output price is what dominates most bills, so we sort by it. All prices are USD per million tokens, standard rates.

Model Maker Input $/1M Output $/1M Cached input Context
Inkling 2 Mini Thinking Machines 0.15 0.50 0.015 1.05M
DaisyRose 27B RadixArk 0.15 1.875 — 262K
Mistral Large 4 Mistral 0.70 2.10 0.07 1M
Qwen4-Max Alibaba 1.50 4.50 0.075 1M
Gemini 4 Hydrogen Google 1.50 7.50 0.075 1.05M
Gemini 4 Neon Google 3.00 15.00 — 1.05M
Claude Opus 5.5 Anthropic 4.00 20.00 — 1M
Claude Fable 5.2 Anthropic 8.00 40.00 0.20 1M
GPT-6.2 Astra OpenAI 8.00 40.00 — 1.05M
GPT-6.2 Astra Pro OpenAI 8.00 40.00 — 1.05M

Caveats that matter when you compare: Mistral Large 4's $0.70 / $2.10 is a preview promotion (the list rate is $1.36 / $4.18), GPT-6.2 Astra is listed at $4.00 / $20.00 for batch work, and Gemini 4 Neon at $1.50 / $7.50, both returning results later for roughly half price. A comparison that mixes standard, batch and preview rates is not a comparison. Our methodology page explains exactly which rate we print and why.

What a million tokens gets you

Raw per-million prices hide the workload. Three examples, priced end to end:

A summarisation job: 500K input, 50K output

Model Input cost Output cost Total
Inkling 2 Mini $0.075 $0.025 $0.10
Mistral Large 4 $0.35 $0.105 $0.46
Qwen4-Max $0.75 $0.225 $0.98
Claude Opus 5.5 $2.00 $1.00 $3.00
GPT-6.2 Astra $4.00 $2.00 $6.00

A read-heavy job like this is 60× cheaper on Inkling 2 Mini than on GPT-6.2 Astra. That is the difference between running it on every document and running it on a sample.

A writing job: 10K input, 40K output

Model Input cost Output cost Total
Inkling 2 Mini $0.0015 $0.02 $0.02
Mistral Large 4 $0.007 $0.084 $0.09
Qwen4-Max $0.015 $0.18 $0.20
Claude Opus 5.5 $0.04 $0.80 $0.84
Claude Fable 5.2 $0.08 $1.60 $1.68

Output-heavy jobs move the ranking less than people expect — the cheapest model stays cheapest — but they multiply the absolute bill, which is where cheap models earn their keep.

An agent running all day: 2M input (mostly repeated), 200K output

This is the pattern behind chat assistants and open-source agent frameworks people wire a cheap API into — a large system prompt and tool definitions resent on every call, with fresh output each time. Cached input pricing is what decides this bill:

Model Input cost (2M, ~90% cacheable) Output cost Total
Inkling 2 Mini $0.03 (cached) $0.10 $0.13
Qwen4-Max $0.16 (cached) $0.90 $1.06
Mistral Large 4 $0.15 (cached) $0.42 $0.57
Claude Fable 5.2 $1.68 (90% cached at $0.20) $8.00 $9.68

The cache column turns a $0.27 input bill into a $0.03 one on Inkling, and a $16 one into a $1.68 one on Fable 5.2. If your workload repeats its prompt, filter your shortlist by cached-input price first — it is the most under-used number in every LLM API pricing comparison.

The cheapest LLM API, by what you are building

For volume work — tagging, extraction, classification, first drafts: Inkling 2 Mini at $0.15 / $0.50 with $0.015 cached input. Cheapest on every axis in the table, with open weights if you later want to eliminate the bill entirely.

For documents and code, at the lowest sensible price: Mistral Large 4 at $0.70 / $2.10. The cheapest API that reads images, with a full 1M context window — but it is a preview release at promotional rates, so budget for the $1.36 / $4.18 list price.

The cheapest closed-vendor API: Qwen4-Max at $1.50 / $4.50, released 16 November 2026. The lowest output price of any closed model we track, with $0.075 cached input. The Qwen4-Max page has the benchmarks.

The cheapest with a free tier: Gemini 4 Hydrogen at $1.50 / $7.50, rate-limited but free to start, with five input modalities in one 1.05M window. The Hydrogen page has the full breakdown.

When the data cannot leave your network: DaisyRose 27B, Apache 2.0, $0.15 / $1.875 hosted, 262K context. The self-hostable middle ground.

When you need the cheapest answer to be right: the cheap-first pattern — run Inkling 2 Mini or Mistral Large 4 and escalate only the failures to Claude Opus 5.5 ($4 / $20) or GPT-6.2 Astra ($8 / $40). Two calls on the cheap model still cost less than one on the expensive one, and the use-case guide turns this into a decision table.

How much does an LLM API cost per month

Take the workload tables above and multiply by your volume. At 1,000 summarisation jobs a day (500K + 50K each), the monthly bill is about $3,000 on Inkling 2 Mini and about $180,000 on GPT-6.2 Astra — a 60× gap on identical work. No amount of prompt optimisation closes a gap that size; model choice does. A realistic budget starts from measured token counts on a few hundred real requests, which is an afternoon of logging, not a research project.

Frequently asked questions

Which LLM API is the cheapest?

On output price, Inkling 2 Mini at $0.50 per million tokens, with $0.15 input and $0.015 cached input (prices read 22 November 2026). DaisyRose 27B matches its input price and is open-weight; Mistral Large 4 is the cheapest that also reads images.

Is a cheap LLM API good enough for coding?

For boilerplate, tests and small diffs, usually yes — Mistral Large 4 at $2.10 output is the value pick. For architecture and debugging that has to be right the first time, cheap-first with escalation to a frontier model beats either extreme. Budget retries into any comparison: a cheap model that needs two attempts has doubled its price.

Is there a free LLM API?

Gemini 4 Hydrogen has a rate-limited free tier. Open-weight models (DaisyRose 27B, Inkling 2 Mini, Mistral Large 4) are free to download but not free to run — you pay for GPUs instead of tokens, which wins at high steady volume.

Why do pricing pages disagree with each other?

Because they snapshot at different times and mix rate types — standard, batch, cached, preview. Compare standard rate with standard rate on the same day, and check the date on any figure before you plan around it. This market moves weekly.