The 10 Cheapest LLM APIs in 2026 — Real Prices From 97 Models

We pulled the live per-token price of every model listed on our gateway — all 97 of them, June 2026 snapshot. The spread is wild: the cheapest usable model costs $0.075 per million blended tokens, about 150× less than GPT-5.5. Here is the full ranked list, the best budget pick for each job, and the honest part — when cheap quietly gets expensive.

TL;DR — the cheapest picks by job

• Cheapest overall: gpt-oss-120b — $0.04 in / $0.18 out per million ($0.075 blended).
• Cheapest reasoning model: Qwen3 235B Thinking — a flat $0.10 in and out.
• Cheapest 1M-token context: Gemini 2.5 Flash Lite and GPT-4.1 Nano, tied at ~$0.175 blended.
• Cheapest current-gen small model: GPT-5.4 Nano — $0.20 / $1.25, with a 400K window.
• Flagships for contrast: Claude Opus 4.7 ($10.00 blended) and GPT-5.5 ($11.25 blended).

How we built this list (real data, reproducible)

Nothing here is copied from vendor marketing pages. On June 12, 2026 we exported the live price list from our own gateway — the same table you can see on the public pricing page — covering 97 models across OpenAI, Anthropic, Google, Qwen, DeepSeek, MoonshotAI, xAI and more. We then ranked every model by blended price = 75% input rate + 25% output rate, the 3:1 input-to-output ratio we see in typical production traffic. Change the ratio and the order shifts slightly, but the tiers below are stable.

The 10 cheapest LLM APIs, ranked

For perspective, the flagships on the same list: Gemini 3.1 Pro blends to $4.50, Claude Sonnet 4.6 to $6.00, Claude Opus 4.7 to $10.00 and GPT-5.5 to $11.25. Everything in the top-10 table is 17–150× cheaper than that band. Cross-reference quality before you commit — LMArena's community leaderboard and Artificial Analysis both track these models — and note that several top-10 entries are previous-generation models: that's largely why they're cheap, not a defect.

What $1 actually buys you

Price-per-million numbers are abstract; here's the same data flipped into blended tokens per dollar:

A worked example: the support-bot bill

Take a customer-support assistant doing 1M requests/month at ~2K input + 0.5K output tokens each (2B input + 500M output tokens monthly):

• On GPT-5.5: 2,000 × $5 + 500 × $30 = $25,000 / month
• On gpt-oss-120b: 2,000 × $0.04 + 500 × $0.18 = $170 / month

Same traffic, $24,830/month difference. Even if the budget model needs a flagship fallback for the hardest 10% of tickets, you're still cutting the bill by roughly 85-90%. That fallback pattern is exactly what routing is for — more below.

The right cheap model for each job

When cheap quietly gets expensive

The honest section. Three ways a $0.08 model costs you more than a $5 one:

• Retry tax. If a budget model fails your task 30% of the time and you retry with a flagship, you pay for both calls — and added latency. Measure the end-to-end cost per solved task, not per call.
• Long-generation drift. Small models degrade faster on multi-thousand-token outputs (reports, multi-file code). Our GPT-5.5 vs Claude Opus 4.7 comparison covers what the flagship tier buys you there.
• Output-heavy workloads flip the math. Blended price assumes 3:1 input:output. A chatty agent at 1:1 makes output rates dominate — recheck the ranking with your own ratio on the pricing page.

Rule of thumb we see across gateway traffic: route the easy 80% to a top-10 model from this list, keep a flagship behind it for the hard 20%, and you capture most of the savings with none of the quality cliff. Official rate cards if you want to verify upstream: OpenAI pricing, Anthropic pricing, Google's Gemini API rates.

One endpoint, cheap by default

Every model in this article is callable through the same OpenAI-compatible endpoint on DataLLM Lab — switching from a $11 flagship to a $0.08 workhorse is a one-string change:

# Same request — only the model string changes
curl https://api.datallmlab.com/v1/chat/completions \
  -H "Authorization: Bearer $DATALLMLAB_API_KEY" \
  -d '{
    "model": "openai/gpt-oss-120b",  # was: "openai/gpt-5.5" — bill drops ~150×
    "messages": [{"role": "user", "content": "Classify this ticket..."}]
  }'

Model IDs for all 97 priced models are in the model directory; live rates on the pricing page.

FAQ