Cheapest LLM API 2026: $0.06 to $0.15 per Million Tokens Ranked
The eight cheapest production-grade LLM APIs at public list price in July 2026, ranked by input cost per million tokens — with the output price, blended cost and the use case each is genuinely best at.
By the LLMCost data team · Published 2026-07-26 · Last reviewed 2026-07-26
The headline
The cheapest production LLM API in July 2026 is Amazon Nova Lite at $0.06 per million input tokens, with Gemini 2.0 Flash and Mistral Small tied at $0.10 for teams not on Bedrock. For workloads that still need frontier-class reasoning at the lowest possible price, DeepSeek V4 Flash at $0.14 is the practical pick.
| # | Model | Input / 1M | Output / 1M | Blended* | Best for |
|---|---|---|---|---|---|
| 1 | Nova Lite (Bedrock) Amazon · 300K ctx | $0.06 | $0.24 | $0.10 | High-volume classification, routing and extraction where you are already on AWS. |
| 2 | Gemini 2.0 Flash Google · 1M ctx | $0.10 | $0.40 | $0.18 | Long-context chat, summarisation and tool use with a 1M-token window. |
| 3 | Mistral Small Mistral · 32K ctx | $0.10 | $0.30 | $0.15 | EU-hosted workloads, function calling and structured output at the lowest non-hyperscaler price. |
| 4 | DeepSeek V4 Flash DeepSeek · 128K ctx | $0.14 | $0.28 | $0.18 | The cheapest model that still feels frontier on code and reasoning for the price. |
| 5 | GPT-5.4 Mini OpenAI · 128K ctx | $0.15 | $0.60 | $0.26 | Default drop-in for teams already on the OpenAI SDK — best tool-calling reliability at this price. |
| 6 | Gemini 2.5 Flash Google · 1M ctx | $0.15 | $0.60 | $0.26 | Newer Flash tier with better reasoning than 2.0 at the same price band. |
| 7 | Llama 4 Maverick (via Groq) Meta · 1M ctx | $0.20 | $0.60 | $0.30 | Latency-critical features (real-time UI, autocomplete) thanks to Groq’s inference hardware. |
| 8 | DeepSeek-V3 DeepSeek · 128K ctx | $0.27 | $1.10 | $0.48 | Full-strength DeepSeek reasoning when V4 Flash is not quite enough. |
*Blended assumes a 3:1 input-to-output token ratio (a typical chat workload). Source: each provider’s official pricing page, July 2026.
Model-by-model notes
#1 Amazon Nova Lite (Bedrock)$0.06 in / $0.24 out
Best for: High-volume classification, routing and extraction where you are already on AWS.
Watch out: Weakest reasoning on this list. Best as a Tier-1 router in front of a stronger model.
#2 Google Gemini 2.0 Flash$0.10 in / $0.40 out
Best for: Long-context chat, summarisation and tool use with a 1M-token window.
Watch out: Flash is fast and cheap but not the tier you want for hard reasoning — step up to Gemini 2.5 Pro for that.
#3 Mistral Mistral Small$0.10 in / $0.30 out
Best for: EU-hosted workloads, function calling and structured output at the lowest non-hyperscaler price.
Watch out: Smaller context window than Gemini Flash. Pick it for tool use, not for long-document ingestion.
#4 DeepSeek DeepSeek V4 Flash$0.14 in / $0.28 out
Best for: The cheapest model that still feels frontier on code and reasoning for the price.
Watch out: Throughput and latency vary by region; provision in the region closest to your users.
#5 OpenAI GPT-5.4 Mini$0.15 in / $0.60 out
Best for: Default drop-in for teams already on the OpenAI SDK — best tool-calling reliability at this price.
Watch out: Prompt caching and the Batch API cut this further; effective cost is often well under list.
#6 Google Gemini 2.5 Flash$0.15 in / $0.60 out
Best for: Newer Flash tier with better reasoning than 2.0 at the same price band.
Watch out: Identical list price to GPT-5.4 Mini — choose on capability and ecosystem fit, not price.
#7 Meta Llama 4 Maverick (via Groq)$0.20 in / $0.60 out
Best for: Latency-critical features (real-time UI, autocomplete) thanks to Groq’s inference hardware.
Watch out: You are paying for speed, not for the cheapest possible token. Llama on CPU/GPU elsewhere is cheaper.
#8 DeepSeek DeepSeek-V3$0.27 in / $1.10 out
Best for: Full-strength DeepSeek reasoning when V4 Flash is not quite enough.
Watch out: Output is 4x the input price — watch your generation length on long completions.
The right architecture: cheap router + strong executor
The cheapest API is rarely the right single model for a whole product. The pattern that wins on cost without sacrificing quality is a two-tier router: a cheap model (Nova Lite, Gemini 2.0 Flash or Mistral Small) classifies the request and handles easy cases, while a stronger model (DeepSeek V4 Flash or GPT-5.4 Mini) handles anything the router is not confident about. In practice 70-90% of requests stay on the cheap tier, and the blended cost lands far closer to the bottom of this table than to any single model’s price.
Try the calculator → to see what your specific token mix would cost across all of these models.
FAQs
What is the cheapest LLM API in 2026?
At public list prices per million tokens, Amazon Nova Lite (via Bedrock) is the cheapest at $0.06 input / $0.24 output. Among non-hyperscaler options, Gemini 2.0 Flash and Mistral Small are tied at $0.10 input. If you want the cheapest model that still performs at a frontier level on code and reasoning, DeepSeek V4 Flash at $0.14 input is the practical winner.
How is "cheapest LLM API" actually measured?
Two numbers matter: price per million input tokens and price per million output tokens, both in USD at the provider's public list rate. For a realistic comparison we also show a blended figure assuming a 3:1 input-to-output ratio, which matches a typical chat workload. Always check both numbers — a model cheap on input but expensive on output (DeepSeek-V3 at $0.27/$1.10) costs more than it looks for generation-heavy workloads.
Are there discounts below the list prices on this page?
Yes, often materially. OpenAI prompt caching can cut input cost by 50% on repeated prefixes; the Batch API gives 50% off for non-real-time jobs. Google, Anthropic and Bedrock all offer volume and committed-use discounts. The numbers on this page are the floor — your effective price is usually lower once caching, batching and volume tiers are applied.
Which cheap LLM API is best for high-volume classification or routing?
Amazon Nova Lite at $0.06/$0.24 is the cheapest option and is well-suited to high-volume classification, routing and extraction. Pair it with a stronger model behind a router (DeepSeek V4 Flash or GPT-5.4 Mini) for the cases where confidence is low — this two-tier pattern is how most production teams keep costs down without losing quality.
Is the cheapest LLM API good enough for production?
For most non-reasoning workloads — classification, summarisation, extraction, routing, formatting — yes. The cheapest tier (Nova Lite, Gemini 2.0 Flash, Mistral Small) handles these comfortably. For multi-step reasoning, agentic tool use and complex code generation, step up one tier to DeepSeek V4 Flash, GPT-5.4 Mini or Gemini 2.5 Flash. The right architecture is usually a cheap router model in front of a stronger executor, not one model for everything.
How often do LLM API prices change?
Frequently. Since 2023, frontier-class input pricing has fallen roughly 60-90% per generation. Treat any ranking as a snapshot and re-check the provider's pricing page before locking in a build. Our LLM API price history tracker keeps a longer view of how these numbers have moved.
Related
- LLM API Price History 2024–2026 — how these prices have fallen over time.
- Per-model pricing pages — cost-per-1K/1M/10M and a monthly calculator for each model.