Blog · AI agents and workflows
Choosing the Right AI Model for a Business Agent (and Keeping Costs Down)
An executive brief on model tiers, routing, caching and batching, with verified price examples and a checklist before you switch models.
By Dijae Dadula · November 2, 2026 · 7 min read
The 60-second brief
- Within one provider, the top model can cost several times more per token than the small one. Most business chats do not need the top model.
- Route by job: small, fast models for routine replies; a larger model only for the hard cases.
- Caching repeated instructions and batching non-urgent work cut the bill further, without changing a word the customer sees.
Decision: default to the smallest model that passes your test set, and pay for more intelligence only where the answer needs it.
This brief is for business owners and teams running an AI chatbot or agent who want to choose a model on purpose instead of by default. The figures come from Google's public Gemini API price list, used here as one verified example. The same levers exist at other providers.
Move 01 · Diagnose
Sort the jobs before you pick a model
List what the agent actually does: answering FAQs, qualifying leads, summarizing chats, drafting replies, reasoning over several tools.
One model for everything means paying top-tier prices for “what are your opening hours?”
Tag each job by difficulty and by volume. The model choice follows from those two tags.
In most business chatbots, the bulk of messages are routine: prices, hours, availability, booking, order status. A smaller share needs real reasoning, such as comparing options, handling a complaint or planning several tool steps. The goal is to price each group separately.
Move 02 · Match
Match each job to a model tier
Providers sell model families in tiers: a small, fast model, a mid-size model and a top reasoning model.
Price rises steeply between tiers, especially for output tokens.
Start every job on the smallest tier and move up only when the test set says so.
Two details matter for budgeting. First, output usually costs much more than input, so long answers are expensive answers. Second, models and prices move fast. Gemini 4 Argon, the newest model on the list, is priced at $2.00 input and $10.00 output per million tokens through November 30, 2026, doubling to $4.00 and $20.00 from December 1, 2026. Gemini 3.8 Flash doubles from January 1, 2027. Chat platforms follow the same tier pattern. Botcake's AI settings, for example, list GPT-5, GPT-5 mini and GPT-5 nano next to older GPT-4.1 and GPT-4o options, plus DeepSeek V4 Flash and V4 Pro. Start on a mini, nano or Flash model and move up only when your test set says so, and recheck the list from time to time, because older generations get retired. For reference, check the pricing page here. OpenAI and Anthropic publish the same kind of tiered price lists; check their current rates here and here.
Move 03 · Route
Route to the small model first, then escalate
A routing rule sends each message to the cheapest model that can handle it.
High-volume, simple messages make up most of the bill, so they set the cost.
Escalate to a larger model only for named cases: complaints, multi-step tool work, low-confidence answers.
Simple escalation rules work well: hand the message to the larger model when the small one says it is unsure, when the customer is upset, or when the job needs more than one tool. Everything else stays on the small model. The same triggers often decide when a person should step in, as covered in the handoff playbook.
Move 04 · Optimize
Cut the bill without cutting quality
Four levers: caching, batching, shorter prompts and capped answers.
None of them change what the customer reads.
Turn on caching for your fixed instructions and knowledge, and batch anything that can wait.
| Lever | How it saves | Example rate (Gemini 3.5 Flash-Lite) |
|---|---|---|
| Context caching | Instructions and knowledge sent with every message are billed at a lower cached rate | $0.03 per 1M cached input tokens vs $0.30 standard, plus $1.00 per 1M tokens per hour of storage |
| Batch API | Non-urgent jobs (summaries, tagging, reports) run asynchronously at a discount | $0.15 input / $1.25 output vs $0.30 / $2.50 standard |
| Shorter prompts | Fewer input tokens on every single call | Remove duplicate rules and old knowledge files |
| Capped answers | Fewer output tokens, the most expensive part | Set a reply length in the instructions and a max output limit |
Illustrative example (the token counts and volume are invented; the prices are the Gemini list prices above): an agent handles 10,000 replies a month, each with 3,000 input tokens of instructions, knowledge and conversation, and a 300-token answer.
| Setup | Cost per reply | Per month |
|---|---|---|
| Gemini 4 Argon for everything (price to November 30, 2026) | $0.0090 | $90.00 |
| Gemini 4 Argon for everything (price from December 1, 2026) | $0.0180 | $180.00 |
| Gemini 3.8 Flash for everything | $0.0034 | $33.75 |
| Gemini 3.5 Flash-Lite for everything | $0.0017 | $16.50 |
| 3.5 Flash-Lite, 2,500 of the input tokens cached | $0.0010 | $11.55 incl. storage |
The absolute amounts are small at this volume. The ratio is the point: the same traffic costs about eight times more on the top model than on the cached small model (about sixteen times once the December price applies), and the gap grows with every new channel and campaign. For the full cost picture beyond tokens, see how much an AI chatbot really costs.
Pay for intelligence only where the answer needs it. Every other message goes to the smallest model that passes the test.
Move 05 · Prove
Prove the choice with numbers
Run the same test set on each candidate model before you switch.
A cheaper model that answers wrongly costs more in lost sales and corrections.
Compare pass rate and cost per resolved conversation, not price per token.
Success metric: cost per resolved conversation, meaning total model cost divided by conversations that ended with an answer or a booked next step, with the pass rate on your test set held at or above today's level. The test method is in how to test an AI chatbot, and the weekly review that catches drift after a switch is in AI agent monitoring after launch.
| Decision checklist before you switch models | |
|---|---|
| Jobs are tagged by difficulty and volume | |
| The small model passes the test set for routine jobs | |
| Escalation rules are written and tested | |
| Caching is on for fixed instructions and knowledge | |
| Non-urgent jobs run as batch | |
| Cost per resolved conversation is tracked weekly |
An honest limitation: model names, prices and quality change every few months, and a public benchmark says little about your customers' questions. Treat any model choice as a decision you re-test each quarter, and re-check the price list before you budget.
Frequently asked questions
What is the best AI model for a business chatbot?
For most routine customer questions, a small, fast model from a major provider is enough. Use a larger reasoning model only for complex cases, and choose based on your own test set rather than public rankings.
How can I reduce AI agent costs?
Route routine messages to a small model, cache the instructions and knowledge sent with every message, batch non-urgent jobs, shorten prompts and cap answer length. Then track cost per resolved conversation.
Why do output tokens cost more than input tokens?
Providers price generated text higher than text you send. On the Gemini API price list, output tokens cost several times more than input tokens on every tier, so long answers raise the bill quickly.
How often should I review my AI model choice?
Review it at least every quarter and whenever a provider changes prices or releases a new model. Re-run your test set before switching.
Get a model and cost plan for your agent
Want your jobs mapped to model tiers, with routing, caching and a cost dashboard set up? Book a free AI Automation Blueprint call.
Prefer to learn it step by step? Join the free weekly newsletter.
References
- Google AI for Developers: Gemini API pricing (paid tier per 1M tokens, context caching, Batch API). Accessed November 2026.
- OpenAI API pricing. Check current rates before budgeting.
- Anthropic Claude API pricing. Check current rates before budgeting.
Keep reading
Get one practical idea like this every week
Join the free newsletter on AI agents, automation and turning your skills into income. Prefer video? Watch the walkthroughs on YouTube.
Want a system like this built for your business? Book a free AI Automation Blueprint call, or see the done-for-you services at my consultancy, IRD Digital Marketing Services.
DIJAE DADULA
Dijae Dadula · AI automation and systems specialist · Cavite, Philippines · Founder of IRD Digital Marketing Services
Home · Blog · Book a call · Newsletter · YouTube · Threads · Transparency
Some links on this site are affiliate links. If you buy through them, I may earn a commission at no extra cost to you.
© 2026 Dijae Dadula. All rights reserved.
