LLM cost · chatbots · multi-turn · July 2026
What is the cheapest LLM for a chatbot?
For a real multi-turn chatbot, the cheapest LLM as of July 2026 is GPT-4.1 nano at about $8.35 per 1,000 ten-turn conversations (~$4.30 with prompt caching). Among frontier-quality models the value tier is Claude Sonnet 5 and Gemini 3.1 Pro — and caching flips which one wins. Ranking by per-token price alone is misleading, because a chatbot resends its whole history every turn.
Every "cheapest LLM" list ranks models by price per token — the cost of one prompt. But a chatbot is not one prompt. It is a conversation, and because LLM APIs are stateless, the client resends the entire history as input on every turn. So the honest question is not "which model has the lowest token price" but "which model runs a full conversation for the least money" — and the answer changes once you count the whole thing, then changes again once you turn on caching. Below is the ranking, the math, and the two places where "cheapest" isn't what the price page says.
GPT-4.1 nano
prompt (same model)
GPT-4.1 nano
Sonnet 5 ↔ Gemini 3.1 Pro
The ranking
Cheapest LLMs for a chatbot, by real conversation cost
Ranked cheapest → priciest by the cost of one standard 10-turn conversation (defined in Methodology: 63,500 input tokens, 5,000 output tokens, full history resent every turn). The "single prompt" column is what a naïve per-token comparison would suggest — it is roughly 1/16th of the real bill. The "with caching" column is where the ranking quietly reorders. Click any header to sort.
| Model | Provider | Input $/1M |
Output $/1M |
Single prompt / 1,000 (naïve) |
Chatbot / 1,000 conv |
With caching / 1,000 conv |
Caching cut |
|---|---|---|---|---|---|---|---|
| GPT-4.1 nano | OpenAI | $0.10 | $0.40 | $0.52 | $8.35 | $4.30 | 48.5% |
| Gemini 3 Flash | $0.50 | $3.00 | $3.10 | $46.75 | $18.89 | 59.6% | |
| Claude Haiku 4.5 | Anthropic | $1.00 | $5.00 | $5.70 | $88.50 | $42.27 | 52.2% |
| GPT-5.6 Luna | OpenAI | $1.00 | $6.00 | $6.20 | $93.50 | $47.27 | 49.4% |
| Gemini 3.5 Flash | $1.50 | $9.00 | $9.30 | $140.25 | $56.66 | 59.6% | |
| Claude Sonnet 5 | Anthropic | $2.00 | $10.00 | $11.40 | $177.00 | $84.55 | 52.2% |
| Gemini 3.1 Pro | $2.00 | $12.00 | $12.40 | $187.00 | $75.55 | 59.6% | |
| GPT-5.6 Terra | OpenAI | $2.50 | $15.00 | $15.50 | $233.75 | $118.19 | 49.4% |
| Claude Opus 4.8 | Anthropic | $5.00 | $25.00 | $28.50 | $442.50 | $211.37 | 52.2% |
| GPT-5.6 Sol | OpenAI | $5.00 | $30.00 | $31.00 | $467.50 | $236.38 | 49.4% |
| Claude Fable 5 | Anthropic | $10.00 | $50.00 | $57.00 | $885.00 | $422.75 | 52.2% |
All figures USD, per 1,000 conversations. "Single prompt" = first turn only (3,200 input + 500 output tokens) — the number a per-token comparison implies. "With caching" applies each provider's cache-read / cache-write rates to the stable prefix; output is never cached. Snapshot of published list prices, verified 2026-07-13. Data reused from the LLM Conversation Cost Index.
Why "cheapest" isn't obvious
Two places the ranking betrays the price page
1 · A chatbot costs ~16× a single prompt
On GPT-4.1 nano, one first-turn prompt costs about $0.52 per 1,000. The full 10-turn conversation costs $8.35 per 1,000 — 16× more — because the system prompt, tools and every earlier message are resent and re-billed on each turn. Rank by per-token price and you rank the $0.52 world; your users live in the $8.35 world. The multiplier holds across models (~15–16×), so it doesn't reorder the list — but it means every "cheapest" number from a token calculator is off by more than an order of magnitude for a chatbot.
2 · Caching flips the value tier: Sonnet 5 ↔ Gemini 3.1 Pro
On raw list price, Claude Sonnet 5 ($177) is cheaper than Gemini 3.1 Pro ($187) per 1,000 conversations. Turn on prompt caching and it inverts: Gemini 3.1 Pro drops to $75.55 while Sonnet 5 lands at $84.55. Google's caching cuts total cost ~60% versus ~52% for Anthropic, and that gap is wider than the 6% list-price gap between the two models — so whichever frontier model is "cheapest" depends entirely on whether your caching is wired. If it isn't, you're leaving that swing on the table.
The practical read: pick the cheapest model that clears your quality bar, then treat caching as part of the price, not a nice-to-have. For a simple FAQ or support bot, GPT-4.1 nano or Gemini 3 Flash are the floor. For a bot that needs frontier reasoning, the value tier is Sonnet 5 / Gemini 3.1 Pro — and if you can cache, Gemini 3.1 Pro is the cheaper of the two. Everything above Opus 4.8 buys quality, not savings.
Methodology
How these numbers are calculated
An LLM API is stateless: to continue a conversation the client resends the entire history as input every turn. So the input tokens on turn t are system + tools + (t−1)·(user + assistant) + user. The standard conversation fixes every variable so all models are compared identically:
- System prompt: 2,000 tokens (resent every turn)
- Tools / context block: 1,000 tokens (resent every turn)
- User message: 200 tokens per turn
- Assistant response: 500 tokens per turn
- Turns: 10
That yields 63,500 input tokens and 5,000 output tokens per conversation. The single-prompt column prices only turn 1 (3,200 input + 500 output = the naïve per-token view). Chatbot / 1,000 prices the full 63,500 + 5,000 at each model's list rate, ×1,000. With caching, turn 1 writes the whole prefix and each later turn reads the prior prefix (at the provider's cache-read rate) and writes only the new user+assistant tokens (at the cache-write rate) — 54,000 cache-read and 9,500 cache-write token-equivalents across the conversation. Output is never cached, on any provider.
Single prompt (turn 1 only) — the naïve per-token view Input = 2,000 + 1,000 + 200 = 3,200 tok -> $0.00032 (@ $0.10 / 1M) Output = 500 tok -> $0.00020 (@ $0.40 / 1M) Per prompt = $0.00052 -> $0.52 / 1,000 Real chatbot (10-turn conversation, no caching) Input, turn 1 = 2,000 + 1,000 + 200 = 3,200 tok Input, turn 10 = 2,000 + 1,000 + 9x700 + 200 = 9,500 tok Total input over 10 turns = 63,500 tok -> $0.00635 (@ $0.10 / 1M) Total output over 10 turns = 5,000 tok -> $0.00200 (@ $0.40 / 1M) Per conversation = $0.00835 -> $8.35 / 1,000 (16x the single prompt) Real chatbot, WITH prompt caching Cache reads = 54,000 tok x 0.25 = 13,500 tok-equiv Cache writes = 9,500 tok x 1.0 = 9,500 tok-equiv Input (cached) = 23,000 x $0.10 / 1M = $0.00230 Output (never cached) = $0.00200 Per conversation = $0.00430 -> $4.30 / 1,000 (saved 48.5%)
Price sources (verified July 13, 2026)
List prices are a dated snapshot from each provider's official pricing page. Providers change prices without notice — always confirm against the source:
- OpenAI (GPT-5.6 Sol / Terra / Luna, GPT-4.1 nano): openai.com/api/pricing
- Anthropic (Claude Fable 5, Opus 4.8, Sonnet 5, Haiku 4.5): anthropic.com/pricing · prompt caching docs
- Google (Gemini 3.1 Pro, 3.5 Flash, 3 Flash): ai.google.dev/gemini-api/docs/pricing
Caveats. This is a planning ranking, not a bill. It assumes fixed average token sizes, a cache hit on every turn (turns within the cache TTL), standard synchronous pricing (no batch discounts, no context-length premiums, no Gemini per-hour cache storage fee). Reasoning/"thinking" tokens are billed as output and would raise the cost of any model run in a high-reasoning mode. Real conversations vary — verify against your provider's usage dashboard, or model your own shape in the Conversation Cost Simulator.
How to lower the bill
Three levers, in order of impact
Turn on prompt caching
Cache the stable prefix — system prompt, tools, early history. It cuts total cost 48–60% on every model here, and is the single biggest lever. Without it, half your published-price "cheap" is gone.
Cap or summarize history
Cumulative input grows with roughly the square of the turn count. Don't resend an ever-growing transcript — truncate, window, or summarize old turns so turn 30 doesn't cost 3× turn 10.
Tier your models
Send simple turns to GPT-4.1 nano or Gemini 3 Flash and reserve a frontier model for the hard ones. Most chatbot turns are easy; paying Opus rates for "what are your hours?" is pure waste.
Embed this on your site (free)
Writing about LLM pricing? Drop in this live badge — it always shows the current cheapest option and links back to the full ranking. No script, no tracking.
<iframe src="https://waynetools.rdkworld.com/embed/cheapest-llm-badge.html" width="360" height="210" style="border:0" title="Cheapest LLM for chatbots — waynetools" loading="lazy"></iframe>FAQ
Common questions
What is the cheapest LLM for a chatbot?+
For a real multi-turn chatbot, the cheapest LLM as of July 2026 is GPT-4.1 nano, at about $8.35 per 1,000 ten-turn conversations, or about $4.30 with prompt caching. Next cheapest are Gemini 3 Flash (~$46.75) and Claude Haiku 4.5 (~$88.50). These rank by total cost per conversation, not just price per token.
Is GPT or Claude cheaper for a chatbot?+
It depends on the tier. In the budget tier, OpenAI's GPT-4.1 nano (~$8.35/1,000) is far cheaper than Claude Haiku 4.5 (~$88.50). In the frontier tier, Claude Sonnet 5 (~$177) is cheaper than the comparable GPT-5.6 Terra (~$233.75). So GPT wins the low end and Claude Sonnet is the cheaper frontier workhorse — but with caching, Google's Gemini 3.1 Pro undercuts both frontier options.
Does prompt caching change which model is cheapest?+
Yes — in the value tier it flips the ranking. Without caching, Sonnet 5 ($177) beats Gemini 3.1 Pro ($187). With caching, Gemini 3.1 Pro drops to $75.55 and Sonnet 5 to $84.55, so Gemini becomes the cheaper frontier model. Google's caching cuts total cost ~60% versus ~52% for Anthropic and 48–49% for OpenAI, enough to reorder the tier.
Why is price per token misleading for a chatbot?+
Because LLM APIs are stateless: to continue a chat the client resends the whole history as input on every turn. A single 3,200-token first prompt looks cheap, but a 10-turn conversation bills 63,500 input tokens — the system prompt and every earlier message paid for again and again. The full chat costs about 16× a single prompt, so per-token or single-prompt rankings understate a chatbot's real bill by more than an order of magnitude.
What is the cheapest frontier-quality LLM for a chatbot?+
The value tier is Claude Sonnet 5 and Gemini 3.1 Pro, both near $180 per 1,000 conversations without caching. Sonnet 5 (~$177) edges out Gemini 3.1 Pro (~$187) on list price, but with caching enabled Gemini 3.1 Pro (~$75.55) becomes cheaper than Sonnet 5 (~$84.55). Premium models — Opus 4.8, GPT-5.6 Sol, Fable 5 — cost 2–5× more for the same conversation shape.
How do I actually lower my chatbot's LLM bill?+
Three levers, by impact: (1) turn on prompt caching for the stable prefix — it cuts total cost 48–60% on every model here; (2) cap or summarize history so you aren't resending an ever-growing transcript, since cumulative input grows with roughly the square of the turn count; (3) route by tier — cheap model for simple turns, frontier model only for hard ones.
Model it yourself