Cheapest LLM API 2026 — DeepSeek, Llama, Gemini Flash, Haiku

TL;DR

The race-to-the-bottom on LLM pricing is real — and it broke open in early 2026. The genuinely cheap tier is now under $0.30 per 1M tokens for input, with quality that would have been "frontier" eighteen months ago. If you're paying GPT-5 prices (<ModelPrice id="openai/gpt-5" field="input" /> / <ModelPrice id="openai/gpt-5" field="output" />) for a high-volume chatbot, you're probably overpaying by 5–20×.

The four models above cover most production text workloads at a small fraction of frontier pricing. DeepSeek V3.5 is the default pick for general-purpose. Mistral Nemo is the floor for high-volume classification. Llama 4 Scout wins on raw throughput. Gemini 2.5 Flash brings Google's multimodal stack at a price that's hard to refuse.

This article breaks down where each model actually wins, what hidden costs change the picture, and which one to pick for your specific scenario.

What "cheap" means in 2026

Cheap in LLM pricing is a moving target. Two years ago, anything under $5/1M output tokens was a deal. Today, $5/1M output is expensive — that's what you pay for premium reasoning, not for production chatbots.

The actual cheap tier in 2026 has two anchors:

Ultra-cheap — $0.02 to $0.30 per 1M input tokens. Open-weight models hosted by their creators (DeepSeek), or aggressively-priced first-party models (Gemini Flash). Quality is decent, often great for narrow tasks.
Fast and reasonable — $0.20 to $2 per 1M tokens. Distilled or "mini" versions of frontier models (GPT-5 mini, Claude Haiku 4.5). Notably better quality than ultra-cheap, with latency advantages.

A subtlety: input price and output price are wildly different, and what matters depends on your workload. A chatbot that takes 600-token questions and emits 180-token answers is bottlenecked by output cost. A classification job that takes 200 tokens of input and emits a single label is bottlenecked by input cost. Always look at both numbers.

The ultra-cheap tier ($0.02–$0.30 per 1M)

Mistral Nemo

The cheapest model in our entire registry that's still production-grade. Built by Mistral × NVIDIA, optimized for inference economics. The 128K context is generous; quality is "fine for narrow tasks" — don't expect it to handle multi-step reasoning, but it nails classification, extraction, and short responses. If you're doing 10M+ requests/month of a well-defined task, this is the floor.

Llama 4 Scout

Meta's small model, released early 2026. The headline is throughput — 380 tokens/sec sustained, the fastest in our text registry. For real-time UIs (autocomplete, live transcription post-processing, streaming chat) where users feel every millisecond, Scout is unbeatable. Quality lands behind DeepSeek V3.5 but ahead of Mistral Nemo on most benchmarks.

DeepSeek V3.5

The default cheap pick. DeepSeek V3.5 hits a sweet spot: prices that compete with the ultra-cheap tier, quality that competes with $5+/1M models. Our home page Best Value pick for a reason. The 128K context is enough for most RAG patterns. Throughput is solid.

The catch: DeepSeek's terms-of-service have caused enterprise-procurement headaches at some shops (data residency in China). For consumer-facing apps, this rarely matters; for regulated industries, check before you commit.

Gemini 2.5 Flash

Google's small multimodal — text + vision + (in some tiers) audio. Input pricing is competitive with DeepSeek but output is markedly higher. The reason to pick Flash over DeepSeek: native multimodal in one call (no separate vision API), and Google's compliance / SLA story for enterprise. The 1M-token context is the largest of the cheap tier — useful if you're doing long-document RAG.

The "fast and reasonable" tier ($0.30–$2 per 1M)

GPT-5 mini

OpenAI's distilled model. Quality clearly above the ultra-cheap tier — closer to Claude 4.6 Sonnet than to DeepSeek V3.5 on hard reasoning. Latency is great. The price is a tiered jump up from the ultra-cheap models, but you get OpenAI's stack: reliable function calling, clean structured outputs, mature SDKs.

Pick GPT-5 mini when you've outgrown DeepSeek's quality but can't justify GPT-5 frontier pricing.

Claude Haiku 4.5

Anthropic's small model with their characteristic strengths: tool-use reliability, multilingual nuance, low hallucination rate on edge cases. The 200K context matches the larger Sonnet. Output price (<ModelPrice id="anthropic/claude-haiku-4-5" field="output" />) is the highest in this tier, so workloads with lots of generation hurt — but for agentic loops where you need consistent tool calls across 20+ turns, it justifies itself.

DeepSeek R2

Reasoning model at near-ultra-cheap prices. If your workload needs chain-of-thought (math, code planning, complex extraction), R2 gives you GPT-5-tier reasoning quality at a tiny fraction of the cost. Caveat: reasoning tokens count against output budget — actual costs are 3–8× the headline output price for hard problems. Still a bargain.

Hidden costs that change the picture

Headline $/1M numbers are the start, not the end. Five things that quietly inflate (or deflate) your real bill:

Output >> input cost ratio. Most cheap models charge 2–5× more for output than input. A chatbot that emits long answers feels expensive even on a "cheap" model. Calculate (input_tokens × input_price + output_tokens × output_price) / 1,000,000 for your real ratio — don't average.
Reasoning token multiplier. Reasoning models (DeepSeek R2, GPT-5 with reasoning.effort: high) emit hidden chain-of-thought tokens that count against your output budget. A "1000-token answer" can cost as much as 8000 tokens of generation. The model's headline price isn't lying — your prompt is just spending more output than you realize.
Cache discount. OpenAI, Anthropic, Google all offer 50–90% off on cached input. If your prompts share a long system message or document, caching can drop effective input cost by 5–10×. Cheap models often skip caching support — DeepSeek V3.5 doesn't have it; GPT-5 mini does. For a RAG workload with stable docs, this can flip the cost ranking.
Batch API discount. OpenAI and Anthropic offer 50% off on async batch inference (24-hour SLA). If your workload is non-realtime (overnight summarization, batch classification), batch tier turns "fast and reasonable" models into ultra-cheap territory. Most cheap models don't have batch API — they're already that cheap.
Per-image vision fee. Multimodal models (Gemini Flash, GPT-5 mini, Claude Haiku) charge separately per image input — typically $0.50 per 1024×1024 image at high detail. A chatbot processing 1M images/month adds $500K to your bill on top of token cost. Always factor in.

Per-scenario picks

The right cheap model depends entirely on your usage shape. Five common production scenarios with real monthly costs:

High-volume customer support chatbot

Pick: DeepSeek V3.5 if data residency isn't a concern. Gemini Flash if you need vision (image attachments) or enterprise compliance. Avoid Claude Haiku at this volume — its output price compounds badly.

RAG / chat-with-docs (long input, short output)

Pick: Gemini 2.5 Flash. Its 1M context handles long documents in one call without chunking, and the input price (where this workload lives) is competitive. DeepSeek V3.5 is close behind but caps at 128K context — if you're doing whole-PDF queries, that hurts.

Code completion / IDE agent

Pick: Qwen 3 Coder for raw code-task quality at low price. DeepSeek V3.5 is a reasonable generalist alternative if your IDE agent does mixed tasks (code + docs + chat). Claude Haiku wins for refactoring agents that need long tool-use loops.

Translation at scale (low input + output, very high volume)

Pick: Mistral Nemo. At this volume, even small per-token differences dominate. Mistral Nemo's price floor wins. If you need certain language pairs (CJK, Indic) and Nemo's quality is borderline, jump to Gemini Flash — Google's multilingual is the strongest in the cheap tier.

Classification at massive scale (10M+ requests/month)

Pick: Mistral Nemo. Classification is the workload that punishes any unnecessary per-token premium. Llama 4 Scout is competitive on raw price but Nemo edges it on input cost — and at this volume, that's all input cost. Don't use a "fast and reasonable" tier model here; you're throwing money away.

What to avoid

Three patterns that look smart on a pricing page and lose money in production:

Picking by raw input price without checking output ratio. A model at $0.05 input / $5.00 output looks cheap until your workload's output dominates. Always do (in × in_price + out × out_price) for your real ratio.
Free-tier models with rate limits. Free tiers exist for a reason — they're throttled, deprecated quickly, and have no SLA. They're fine for prototyping; production needs a paid tier with attribution. Models priced at $0/$0 in our registry are filtered out of CHEAPEST picks for this reason.
Ultra-cheap models without function calling for agent loops. Some cheap models lack reliable function calling (or have it but it hallucinates JSON). For an agent doing 10+ tool calls per session, the savings vanish in retries and error handling. Mistral Nemo, for instance, is great for single-shot tasks but flaky for tool use — pick Claude Haiku 4.5 or GPT-5 mini for agents.

Decision framework

A 2×2 grid that captures most cases:

	Basic quality OK	Decent quality required
Massive volume (10M+/mo)	Mistral Nemo · Llama 4 Scout	DeepSeek V3.5 · Gemini 2.5 Flash
Moderate volume (< 1M/mo)	DeepSeek V3.5 · Gemini Flash	Claude Haiku 4.5 · GPT-5 mini

Two extra dimensions that move you across cells:

Need function calling reliability? → push right (DeepSeek → GPT-5 mini)
Need vision? → push to Gemini Flash, GPT-5 mini, or Claude Haiku
Need reasoning? → DeepSeek R2 (cheap), then GPT-5 (frontier)
Long context (>128K)? → Gemini Flash (1M) or Claude Haiku (200K)

Our default pick for 2026

If you have to pick one cheap model for a generic production workload and stop optimizing: DeepSeek V3.5. It's the best ratio of price to quality in our entire text registry. It's why our home page tags it BEST VALUE.

Pick something else when:

You need vision in the same call → Gemini 2.5 Flash
You're at 10M+ requests/month and quality is "fine" → Mistral Nemo
You need OpenAI-grade tool use → GPT-5 mini
You need long-loop agent reliability → Claude Haiku 4.5
Data residency requires non-China hosting → Gemini Flash or Claude Haiku

The Cheapest LLM API for Production in 2026