Cheapest LLM APIs in 2026: Per-Token Pricing From $0.01 to $75 per Million

Every Major LLM API, Sorted by Input Cost (April 2026)

Prices per million tokens. Input is what you send. Output is what the model returns. Output typically costs 2-8x more than input. For coding agents and long-context workloads, input cost dominates because you send far more tokens than you receive.

Budget Tier: Under $1 per Million Input Tokens

This is where most production workloads should start. Models in this range handle classification, extraction, simple code generation, summarization, and formatting. The quality gap between a $0.20/MTok model and a $3.00/MTok model is smaller than the 15x price difference suggests.

The surprise in this tier: OpenAI's reasoning model o4-mini at $0.55/$2.20 is cheaper than many general-purpose mid-tier models. It replaced o3-mini and outperforms it on most benchmarks. For tasks that need structured reasoning on a budget, o4-mini undercuts everything except DeepSeek's reasoner mode.

Mid Tier: $1-5 per Million Input Tokens

The workhorses. These models handle complex code generation, multi-step reasoning, long-form writing, and most agentic workflows. The price difference between $1.25/MTok (Gemini 2.5 Pro) and $3.00/MTok (Claude Sonnet 4.6) matters at scale but both are viable for production.

Two models stand out. Gemini 2.5 Pro at $1.25/$10.00 offers frontier quality at nearly half the price of Claude Sonnet 4.6. Its 1M context window matches Claude's and costs less. The catch: Gemini is newer for coding agent use cases and the ecosystem (tool use, function calling) is less mature than Anthropic's or OpenAI's.

Mistral Large 3 at $2.00/$6.00 has the cheapest output in this tier. If your workload is output-heavy (long-form generation, detailed code), Mistral Large saves 60% on output versus Claude Sonnet ($6 vs $15) and 25% versus o3 ($6 vs $8).

Premium Tier: $5+ per Million Input Tokens

Reserve these for tasks where quality directly translates to value. Architecture decisions, complex debugging, security audits, production incident analysis. Running a premium model on routine code formatting is burning money.

GPT-5.4 at $2.50/$15.00 is technically in this tier but borderline mid-tier on input pricing. It is the cheapest way to access OpenAI's flagship-class quality. Claude Opus 4.6 at $5.00/$25.00 costs 2x the input and 1.67x the output, but includes a 1M token context window at standard pricing (GPT-5.4's context is 272K, with prices doubling past that).

The truly expensive models (GPT-5.4 Pro at $30/$180, o1-pro at $150/$600) exist for specialized research and evaluation. No production coding agent should run on these.

Inference Providers: Same Model, Different Price

Open-source models (Llama, Qwen, Mixtral, DeepSeek) run on multiple providers. The same model at the same quality can cost 2-5x more depending on where you host it.

For most teams: start with Groq or Together AI for open-source models. If you need dedicated capacity or fine-tuned models, look at Fireworks on-demand or self-hosting.

Batch Discounts and Prompt Caching

Raw per-token pricing is the sticker price. Actual cost depends on whether you use the discount stack every major provider offers.

The math on stacking. Take Claude Sonnet 4.6 at $3.00/MTok input:

OpenAI's GPT-5.4 follows the same pattern. Standard at $2.50, cached at $0.25, batch at $1.25, batch + cached at $0.13. The effective price of a frontier model with full discount stack is cheaper than the sticker price of most budget models.

Quality-Adjusted Cost: Cheap but Bad = Expensive

A $0.05/MTok model that fails 40% of tasks and needs 3 retries per success costs more than a $3.00/MTok model that succeeds on the first attempt. The metric that matters is cost per completed task, not cost per token.

DeepSeek V3.2 at $0.47 per completed task is the value sweet spot. It is 12x cheaper per completed task than Claude Sonnet, with a 75% single-attempt success rate that is good enough for most automated workflows with retry logic. Claude Sonnet's 92% success rate matters when retries are expensive (user-facing, time-sensitive, or when errors compound in multi-step agents).

The cheapest models (Llama 3.1 8B, GPT-4.1 Nano) look cheap per token but burn through tokens on retries. They are cost-effective only for tasks where they reliably succeed on the first attempt: classification, extraction, formatting, simple completions. Using them for complex code generation is a false economy.

Local Inference via Ollama: $0 Per Token

Local inference eliminates per-token costs entirely. You pay for hardware upfront and electricity ongoing. Ollama, vLLM, and llama.cpp make self-hosting straightforward.

Break-even is roughly 50 million tokens per month, or about 1.7M tokens per day. Below that, APIs are cheaper. Above that, local inference wins on raw cost, though you lose access to frontier-quality models. A Mac Studio running Llama 3.1 70B under full GPU load draws about 60W, costing under $15/month in most US electricity markets.

The practical limitation: local models top out at 70B parameters on consumer hardware (120B with quantization tricks). Tasks that require frontier reasoning (complex architecture, nuanced debugging) still need API calls to Claude Opus, GPT-5.4, or Gemini 2.5 Pro. The hybrid approach works best: route simple tasks to local models, escalate hard problems to API.

How Morph Compact Makes Every LLM Cheaper

Every optimization above reduces the price per token. Morph Compact reduces the number of tokens. These are complementary strategies, and they multiply.

Compact compresses LLM context by 50-70% at 33,000 tokens per second. The compression is extractive: every surviving line is character-for-character from the original input. No paraphrasing, no synthesis, no hallucination risk. The model decides which lines are relevant and discards the rest.

The savings per request look modest at the cheap end. But coding agents send 20-100 requests per session, each carrying accumulated context. A 30-turn agent session on Claude Sonnet 4.6 with 200K tokens of context per turn:

At 10 sessions per day, that is $108/day saved, or $3,240/month. The savings scale linearly with token volume and model cost. Compaction is most valuable on expensive models (where each saved token is worth more) and on long-running agent sessions (where context accumulates the most).

Claude Sonnet 4.6, a frontier coding model, at $0.06 effective per MTok of original context. That is cheaper than the sticker price of Llama 3.1 8B on Groq ($0.05/$0.08), while delivering 92% first-attempt success rate versus 35%. The discount stack inverts the pricing hierarchy.

Frequently Asked Questions

What is the cheapest LLM API in 2026?

The absolute cheapest is Llama 3.1 8B on Groq at $0.05/$0.08 per million tokens. Amazon Nova Lite at $0.06/$0.24 is close. Among first-party providers, DeepSeek V3.2 at $0.28/$0.42 offers the best quality-to-cost ratio. Gemini 2.5 Flash-Lite at $0.10/$0.40 is cheapest from a major US provider. GPT-4.1 Nano at $0.10/$0.40 is cheapest from OpenAI.

How much does GPT-5.4 cost?

GPT-5.4 costs $2.50 input / $15.00 output per million tokens. Cached input at $0.25. Batch at $1.25/$7.50. GPT-5.4 Mini is $0.75/$4.50. GPT-5.4 Nano is $0.20/$1.25. GPT-5.4 Pro (the expensive one) is $30.00/$180.00.

Is DeepSeek cheaper than ChatGPT?

Yes. DeepSeek V3.2 at $0.28/$0.42 is 9x cheaper on input and 36x cheaper on output compared to GPT-5.4. DeepSeek V4 at $0.30/$0.50 is the latest flagship and still 8x cheaper on input. Cache hits at $0.028-$0.03/MTok are the cheapest cached input from any provider. Quality is competitive for coding and general tasks but below frontier models on the hardest reasoning benchmarks. DeepSeek R1 at $0.55/$2.19 offers chain-of-thought reasoning at a fraction of o3's cost ($2.00/$8.00).

What is the cheapest LLM for coding?

Dedicated coding models: Qwen3 Coder 480B A35B at $0.22/$0.90. Codestral at $0.30/$0.90. DeepSeek V3.2 in reasoner mode at $0.28/$0.42. General models that code well at low cost: GPT-4.1 Mini at $0.20/$0.80, o4-mini at $0.55/$2.20. For the best coding quality, Claude Sonnet 4.6 at $3.00/$15.00 leads most benchmarks but costs 10x more.

How do batch API and prompt caching reduce costs?

Batch APIs (OpenAI, Anthropic, Google) process requests asynchronously within 24 hours at 50% off. Prompt caching stores repeated context so subsequent requests pay 10% of standard input price (Anthropic, OpenAI, Google). These stack: batch + cache hit gives 95% savings on input. DeepSeek's cache hit price of $0.028/MTok makes it the cheapest cached input from any provider.

Is running a local LLM with Ollama cheaper than an API?

Above 50 million tokens per month, yes. A Mac Studio M4 running Llama 3.1 70B costs ~$15/month in electricity. Below 50M tokens, APIs are cheaper because you avoid the $4,000-8,000 hardware investment. Quality ceiling is 70-85% of frontier models. Best used in a hybrid setup: local for simple tasks, API for complex reasoning.

How does Morph Compact make LLMs cheaper?

Compact compresses context by 50-70% before sending it to any LLM. A 100K prompt compacted to 40K costs 60% less. The compression is extractive (verbatim lines, no paraphrasing) with zero hallucination risk. Combined with prompt caching and batch processing, it brings Claude Sonnet 4.6's effective input cost to $0.06/MTok, cheaper than the sticker price of most budget models. See how Compact works.