The Cheapest LLM APIs in 2026, Compared
Last verified August 2026.
"Cheapest" depends entirely on your input-to-output token ratio, because every provider prices input and output separately — usually output costs 4-6x more than input. A model that looks cheap on paper can end up expensive for output-heavy workloads (long-form writing, code generation) and vice versa.
The budget tier
At the low end, three models stand out for high-volume, latency-tolerant workloads:
- DeepSeek V4 Flash — $0.14 input / $0.28 output per million tokens, with cache hits as low as $0.0028/M input.
- Grok 4.1 Fast — $0.20 / $0.50 per million tokens.
- GPT-5.6 Luna — $0.20 / $1.20 per million tokens, OpenAI's high-volume tier.
For simple classification, extraction, or summarization tasks where quality requirements are modest, these three are within a few percent of each other in practice — the deciding factor is usually latency and API compatibility, not price.
The mid tier
Claude Sonnet 5 ($2/$10, introductory through Aug 31 2026), Gemini 3.1 Pro ($2/$12), Grok 4.3 ($1.25/$2.50), and GPT-5.6 Terra ($2/$12) compete here. Grok 4.3 is notably the cheapest frontier-class model in this bracket on output cost, which matters most for generation-heavy tasks.
The flagship tier
Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30) sit at the top for tasks that need the strongest reasoning — complex agentic planning, hard coding tasks, long-context synthesis. At this tier, price differences matter less than raw capability, since a wrong answer that needs a retry costs more than the price gap between models.
Why the "cheapest" answer changes per workload
A chatbot with short questions and short answers behaves completely differently, cost-wise, than an agent that reads a 5,000-token document and writes a 200-token summary, or a code-generation task that reads a small prompt and writes 2,000 tokens of code. Rather than rely on a headline number, run your actual input/output token counts through a calculator that separates the two — the ranking changes depending on your ratio.
If you're building anything with multiple LLM calls per task — an agent, not a single chatbot reply — see why agent costs run higher than they look.