What's prompt caching?

Providers store your system prompts after the first call so re-using them is much cheaper (~75–90% input discount). Huge win when your system prompt is long or you make many similar calls.

Are batch APIs really 50% off?

Yes, on OpenAI and Anthropic. The tradeoff is latency — batch jobs return within 24 hours. Great for offline workloads like backfills or eval suites.

How do open-weights compare?

DeepSeek, Qwen, Llama 3.x are all competitive at frontier-class cost. Hosted (via Together, Fireworks, Groq) or self-hosted (via vLLM, Ollama) — both are cheaper than proprietary APIs at scale.

Will prices keep dropping?

Yes — historical trend is ~10x every 18 months. Plan for tomorrow's prices when you design your cost model.

Model Comparison

Cheapest AI APIs

Per-million-token cost comparison across providers, with notes on where the cheap models are actually good enough.

Models:Gemini FlashGPT-5 miniClaude HaikuDeepSeekGroqUpdated May 2026

Verdict

Gemini Flash for general work. DeepSeek for power. Groq for speed. Pick by job, not by brand.

APIPricingCostComparison

Edited by The AIKnowHub team · Editorial team

At a glance

Dimension	Gemini Flash	GPT-5 mini	Claude Haiku	DeepSeek	Groq
Input ($/1M tokens)	$0.15	$0.40	$0.80	$0.27	—
Output ($/1M tokens)	$0.60	$1.60	$4.00	$1.10	—
Latency (first token)	~400ms	~600ms	~400ms	~700ms	—
Throughput (tokens/sec)	Fast	Fast	Fast	Medium	—
Prompt caching	Yes	Yes	Yes	Yes	—

Best pick by use case

If you're doing…	Pick	Why
Classification / routing	Gemini Flash	Cheapest, accurate enough.
Real-time UI streaming	Groq	Lowest latency on the market.
Coding at low cost	DeepSeek V3	Best quality-per-dollar on code.
Cheap structured extraction	Gemini Flash or GPT-5 nano	Reliable JSON output at minimum cost.
Open-weights / self-host	DeepSeek or Qwen via Ollama	$0 inference if you have hardware.

The 2026 price floor

Model	Input ($/1M)	Output ($/1M)	Notes
Gemini 2.5 Flash	0.15	0.60	Big context, strong all-around
GPT-5 nano	0.10	0.40	Cheapest OpenAI option
GPT-5 mini	0.40	1.60	Solid mid-tier
Claude Haiku 4.x	0.80	4.00	Higher quality, higher price
DeepSeek V3	0.27	1.10	Open weights also free to self-host
Qwen3-Turbo	0.20	0.80	Solid, especially for code
Groq (Llama 3.x 70B)	0.59	0.79	Best for speed-critical apps
Cerebras (Llama 3.x)	0.60	0.80	Fastest tokens/sec on the market

Prices change every quarter. Always check provider pricing pages before benchmarking.

What "cheap enough" buys you

Cheap models are great for:

Classification (is this spam? what category? sentiment?)
Extraction (pull dates / amounts / names from text)
Summarization of short documents
Routing (which downstream model should handle this?)
Reformatting (CSV → JSON, prose → bullets)

Cheap models are not great for:

Reasoning chains with multiple steps
Long-form writing with voice
Code agent loops
Tasks where being wrong is expensive

The two-model pattern

The most cost-effective production setup in 2026:

Cheap model → Classify / extract / route
Strong model → Hand off the hard parts only

A spam filter using GPT-5 nano costs 1/50th of the same filter on Claude Sonnet and is within a few % on accuracy.

Hidden costs to watch

Prompt caching — Anthropic, OpenAI, Gemini all support it. Use it. Reduces input costs by 75–90%.
Batch APIs — 50% discount for non-realtime work.
Context bloat — every retained turn in a long conversation costs tokens. Trim aggressively.

The reality

API prices have fallen ~10x in 18 months. They will fall further. If your product is gated by today's API costs, plan for what 3x cheaper looks like — that's where you'll be in 12 months.

Frequently asked questions

Pick 20 representative tasks. Run them on the cheapest model. Have a human (or a stronger model) judge. If quality is acceptable, stay there. Most teams over-pay for routine work.