Cheapest AI API 2026:
Lowest Cost LLMs Ranked

Cheapest AI APIs Ranked by Input Price (November 2026)

Prices are per million tokens (1M tokens ≈ 750,000 words). Sorted by input token cost, lowest first. Current models only — retired models (Claude 3.x, Gemini 1.5/2.0) have been removed.

Top 3 Cheapest AI APIs — Detailed Breakdown

🥇 #1 (tie): GPT-6 Luna — $0.10/M input

OpenAI's cheapest current-generation model at $0.10/$0.50 per million tokens — frontier-family quality at a budget price, with a ~1M token context window. A strong default for high-volume classification, extraction, and everyday chat. Batch/Flex pricing halves it again to $0.05/$0.25.

• Best for: High-volume chat, classification, extraction, everyday agentic tasks
• Avoid for: Tasks requiring frontier-level reasoning (use GPT-6.1 Sol or Claude Sonnet 5.5 instead)
• 1M token cost: $0.10 input + $0.50 output = $0.30 total for a 500K in / 500K out split

🥇 #1 (tie): Claude Haiku 5.5 — $0.10/M input

Anthropic's new budget model matches GPT-6 Luna at $0.10/$0.50 per million tokens for prompts up to 100K tokens (prompts above that are billed at $0.50/$2.50), with a 1M context window. Note it uses Anthropic's newer tokenizer, so the same text counts as roughly 30% more tokens than on older Claude models.

• Best for: Classification, extraction, routing, subagent tasks, structured output at scale
• Watch out for: Very long prompts — the over-100K rate is 5× higher

🥉 #3: Qwen3.8-Flash — $0.15/M input

Alibaba's budget Qwen model costs $0.15 per million input tokens and $0.47 per million output tokens (international rate, prompts up to 1M tokens). Output is actually a little cheaper than GPT-6 Luna and Claude Haiku 5.5, so for long, output-heavy replies it can come out cheapest overall. It's billed through Alibaba Cloud Model Studio, which takes a few more setup steps than the US providers.

• Best for: Output-heavy, high-volume work like drafting, summaries and translation
• Watch out for: Data-residency and compliance requirements. Check where your requests will be processed before sending sensitive data.

Also worth a look: DeepSeek Flash (V4.1) — $0.30/M input

DeepSeek charges $0.30 input / $1.20 output per million tokens at peak times, and half that ($0.15 / $0.60) off-peak. Peak hours are weekdays 01:00–04:00 and 06:00–10:00 UTC (9am–noon and 2–6pm Beijing time), so most US working hours fall in the cheaper off-peak window. GPT-4o mini (legacy) also sits at $0.15/M input, but it's scheduled for API shutdown on January 14, 2027, so start new projects on GPT-6 Luna instead.

How Much Can You Save by Switching Models?

Assume a customer support chatbot processing 1 million requests/month with 500 input + 200 output tokens each:

Should You Use Open-Source Models (Llama, Mistral)?

Meta's Llama and Mistral's open-weight models can be significantly cheaper when self-hosted, but running your own infrastructure adds complexity and fixed costs — and hosted pricing varies by provider. AIModelCalc currently tracks five first-party providers (OpenAI, Anthropic, Google, DeepSeek and Alibaba Qwen) where published per-token pricing is stable; for open-weight models, check your hosting provider's current rate, since it's often competitive with the cheapest closed-source options above.

Self-hosting makes sense if you process more than 50M tokens/month and have engineering capacity. Below that threshold, hosted APIs are simpler and often cheaper when you factor in GPU costs and maintenance.

Tips for Minimizing AI API Costs

• Use prompt caching — Anthropic and OpenAI both offer 90%+ discounts on repeated input context
• Trim system prompts — Every token in your system prompt is charged on every request
• Match model to task — Use small models for simple tasks, save premium models for complex work
• Batch non-urgent requests — OpenAI's Batch API offers 50% off for async workloads
• Monitor output length — Output tokens cost 5x or more than input; set max_tokens limits