Cheapest LLM APIs for agents: price list and what a run costs

“Cheapest” depends on what you are paying for. Here it means list price, not the cheapest model that is good enough for your job. Use it to find the floor, then check the benchmark rank and test on your own tasks.

The short version

• On blended list price (3 parts input to 1 part output, an assumption we picked because agent calls re-send a lot of context, not a measured ratio), the cheapest on our list are Claude Haiku 5.5 and GPT-6 Luna (tied) at $0.100 input and $0.500 output per 1M tokens. Check the fine print below: the cheapest prices often come with a prompt-size or time-of-day condition.
• Among the five highest-ranked models on our list, the cheapest is GPT-6.1 Terra at $1.00 / $5.00 per 1M tokens (rank 5).
• For the example workload used across this site (1,000 calls, 2,000 input and 500 output tokens each), the cheapest costs $0.450. The most expensive costs $45.00.

Everything, cheapest first

Rank is our ordering by Artificial Analysis Intelligence Index v5.0 (best published reasoning setting), then LMArena Text score. “AA index” is the Artificial Analysis Intelligence Index score where we have one. “Blended 3:1” is (3 × input + output) ÷ 4. The example workload column is plain arithmetic, not a benchmark: 1,000 calls × (2,000 input + 500 output tokens).

Read the fine print before you pick

A few list prices come with conditions, and they change what “cheapest” means. From the notes in our data:

• GPT-6 Astra: Short-context (<=272K input) price.
• GPT-6.1 Sol: Short-context (<=272K input) price.
• GPT-6.1 Terra: Short-context (<=272K input) price.
• Gemini 4 Argon: Promo price through Nov 30, 2026; $4/$20 from Dec 1, 2026.
• Muse Wonder: API model muse-wonder-1.0; Muse Wonder Full is Muse Code subscription-only.
• Claude Haiku 5.5: Prompts up to 100K tokens; $0.50/$2.50 above.
• Grok 4.7: <200K prompt price; $4/$12 at >=200K.
• DeepSeek V4.1 Flash (1017): API model deepseek-flash. Peak-hour cache-miss price; off-peak is half.
• Gemini 3.8 Flash: Promo price through Dec 31, 2026; $1.50/$7.50 from Jan 1, 2027.
• GPT-6 Luna: Short-context (<=272K input) price.
• DeepSeek V4 Pro (0813): Peak-hour cache-miss price; off-peak is half.

2 models are on a promotional price with an end date. If you are planning a budget past that date, use the post-promo price, not the one in the table.

8 models list a price that only applies below a prompt-size or time-of-day threshold. Long agent loops can cross those thresholds, so check how big your context really gets. The run cost planner shows the peak context for your settings.

Why list price is only half the story for agents

Agent loops re-send the whole conversation on every step. If a run has 12 steps, the first step’s tokens get billed again on each of the following 11. That is why input price usually matters more than output price for agents, and why the cheapest-looking model is not always cheapest per finished task: a weaker model that needs more steps or retries can cost more overall. Prompt caching, batch discounts, reasoning tokens and tool fees are real and are not in this table; the list prices here are standard-tier prices only.

Practical way to use this page: shortlist two or three models near the floor that clear your quality bar, run the same 20 tasks on each, and compare cost per completed task, not cost per token. Then plug your real step counts into the run cost planner.