First decide what you are buying. If your staff need an AI assistant, buy a seat plan such as ChatGPT Business or Claude Team, at $20 per seat per month billed annually. If you are putting AI inside your own software, start by testing a mid-tier API model (Claude Sonnet 5.5 or GPT-6.1 Sol, both $2 / $10 per million tokens), or a small one for extraction and classification (Claude Haiku 5.5 or GPT-6 Luna, both $0.10 / $0.50).
Move up to a top model only when the cheaper one fails a test on your own data. All prices were checked on the vendors’ official pages on 2026-10-13.
Who this is for: owners and product leads at small and mid-size businesses choosing a model for a task, or choosing between ChatGPT, Claude, and Gemini. Who should look elsewhere: if you need a self-hosted open-weight model for legal or air-gap reasons, this guide only covers the hosted APIs from OpenAI, Anthropic, and Google. If you want raw benchmark scores, see Artificial Analysis, which measures models independently.
Pick your model in under a minute
Answer up to six questions. The picker returns three models to test, each vendor’s own reason for the model, the monthly cost at list price, and the test to run first. It runs in your browser.
The same picker lives on its own page at /tools/ai-model-picker/. Here are its default answers, for 10,000 tasks a month with a 5,000-token prompt:
| Job | Test first | Then | Then | How to test |
|---|---|---|---|---|
| Draft text: emails, product copy, summaries | Claude Sonnet 5.5 ($208) | GPT-6.1 Sol ($160) | Gemini 3.8 Flash (2026 price) ($60.00) | Write 20 real requests. Read the drafts side by side with the model names hidden, and pick the one you would send. |
| Extract or classify: invoices, forms, tickets | Claude Haiku 5.5 ($8.45) | GPT-6 Luna ($6.50) | Gemini 3.5 Flash-Lite ($22.50) | Run 50 real documents where you already know the answer. Count the fields that come back wrong. |
| Answer customers in a chat | Claude Sonnet 5.5 ($169) | GPT-6.1 Sol ($130) | Claude Haiku 5.5 ($8.45) | Replay 30 past tickets. Check every fact against your help docs, and check the tone. |
| Agent that uses tools and takes actions | Claude Opus 5.5 ($1,040) | GPT-6.1 Sol ($400) | Claude Sonnet 5.5 ($520) | Run 10 real tasks end to end. Count wrong tool calls and wrong actions, not just wrong final answers. |
| Read long documents: contracts, reports, manuals | Claude Sonnet 5.5 ($234) | Gemini 3.8 Flash (2026 price) ($67.50) | GPT-6.1 Sol ($180) | Ask 10 questions whose answers sit deep in the document, and check each quote against the source. |
| Hard reasoning, research, or long agent runs | Claude Fable 5.1 ($3,250) | GPT-6 Astra ($2,500) | Claude Opus 5.5 ($1,300) | Run the same tests on the mid-tier model too. Move up only if it fails them. Anthropic: "Tuning effort is often a better lever than switching models." |
Step 1: a chat app for staff, or a model inside your software?
A chat app for staff is a seat plan where the vendor picks the model. A model inside your software is an API you pay per token, and you pick the model by ID. The top-ranking guides for this question compare ChatGPT, Claude, and Gemini as chat apps and stop there, which answers only half the question.
| Chat app for staff | Model inside your software | |
|---|---|---|
| What you buy | Seats on ChatGPT Business, Claude Team, or Gemini in Google Workspace | API access, billed per token |
| Who picks the model | The vendor, with some choice in the app | You, by model ID |
| Price | ChatGPT Business and Claude Team: $20 per seat per month billed annually, $25 monthly. Premium seats $100 / $125 | From $0.45 to $58.50 per 1,000 tasks of 2,000 tokens in and 500 out, depending on the model |
| Sees your own database | Only through a connector or an MCP server | Yes, through your code |
| Right for | Writing, research, and admin work by people | Features your customers or systems use: replies, extraction, agents |
Seat prices come from OpenAI’s ChatGPT Business billing page and Claude’s pricing page. Both list Premium seats with 5x the usage of Standard. Claude Team runs “for teams of 2 to 150” and promises “No model training on your content by default.” We could not read Gemini for Workspace prices on Google’s page as text, so we leave them out rather than guess.
If staff need the assistant to read your own systems, the plan alone won’t do it. Our guide to connecting company data to ChatGPT covers the options.
Step 2: match the model to the job
For a model inside your software, the job decides the tier. These are the current models from each vendor, with the vendor’s own description:
| Tier | Model | Input / output per 1M tokens | What the vendor says |
|---|---|---|---|
| Top | Claude Fable 5.1 | $10 / $50 | ”Anthropic’s most capable model open to all customers” |
| Top | GPT-6 Astra | $10 / $50 | ”Our most capable model for the most demanding work.” |
| Upper mid | Claude Opus 5.5 | $4 / $20 | ”If you’re unsure which model to use, start with Claude Opus 5.5 for most workloads.” |
| Mid | Claude Sonnet 5.5 | $2 / $10 | ”The best combination of speed and intelligence” |
| Mid | GPT-6.1 Sol | $2 / $10 | ”Near-Astra performance for complex work at a lower cost.” |
| Mid | Gemini 4 Argon | $2 / $10, then $4 / $20 from Dec 1, 2026 | ”Our flagship reasoning model” |
| Lower mid | Gemini 3.8 Flash | $0.75 / $3.75, then $1.50 / $7.50 from Jan 1, 2027 | ”Our most intelligent Flash model” |
| Small | Claude Haiku 5.5 | $0.10 / $0.50, more above 100K tokens | ”For high-volume, latency-sensitive tasks such as classification, extraction, and routing” |
| Small | GPT-6 Luna | $0.10 / $0.50 | ”Our most efficient model for focused, high-volume tasks.” |
| Small | Gemini 3.5 Flash-Lite | $0.30 / $2.50 | ”A cost-efficient model, optimized for high-volume agentic tasks, translation, and simple data processing.” |
Sources: OpenAI models, Anthropic models overview and choosing a model, Gemini pricing. Our full AI API pricing table adds cached, batch, and long-prompt rates.
Notice that both OpenAI and Anthropic tell you to start high. OpenAI’s page says “If you’re not sure where to start, use GPT-6 Astra.” That is safe advice for the vendor and expensive advice for you. Anthropic’s same guide offers the other path, “Start efficiency-first,” which begins with Claude Haiku 5.5 and upgrades “only if necessary for specific capability gaps.” For most small business tasks, we start one tier lower than the vendor suggests and move up on evidence.
Step 3: let price and length narrow it
Three numbers change the answer more than any benchmark:
- Volume. At 100,000 tasks a month, a support reply on Claude Sonnet 5.5 costs about $1,690 and on Claude Haiku 5.5 about $85, with a 5,000-token prompt. If the small model is good enough, that is the decision made. Our AI cost per user guide shows the same math per user.
- Prompt length. OpenAI’s GPT-6 models charge “2x input and cache rates and 1.5x output for the full request” when the prompt passes 272,000 tokens. Claude Haiku 5.5 costs five times more above 100,000. Claude Sonnet 5.5 and Opus 5.5 stay flat to 1 million. For contract packs and long reports, that flat price can beat a cheaper list price.
- Dated price changes. Gemini 4 Argon doubles on December 1, 2026, and Gemini 3.8 Flash on January 1, 2027. Our Gemini price increase guide has the full list. If you build on a model now, price your plan on the rate it will have in three months.
Two more filters:
- Customer data. Use the paid API, not a free tier. OpenAI’s data guide says “data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us).” Anthropic’s privacy center says it won’t use inputs or outputs from “our commercial products (e.g. Claude for Work, Anthropic API…)” to train by default. Google’s Gemini API terms say paid-tier prompts aren’t used to improve its products, but free-tier prompts are.
- Model lifespan. Older models shut down on fixed dates, and your code breaks if it still calls them. Check the AI model retirement calendar before you pin a model ID.
Step 4: test before you choose
No table can tell you which model handles your invoices or your customers’ questions. A short test can:
- Collect 20 to 50 real inputs with known good answers. Past tickets, real invoices, real documents.
- Run them through your top two or three picks with the same prompt.
- Score the outputs blind. Hide the model names. For extraction, count wrong fields. For writing, pick the one you’d send.
- Read the usage numbers. Every API response reports tokens used. Multiply by the price for the real cost per task. Claude’s newer tokenizer “produces approximately 30% more tokens for the same text,” per Anthropic’s pricing page, so equal list prices don’t mean equal bills.
- Try the effort setting before a bigger model. Anthropic says “Tuning effort is often a better lever than switching models.” OpenAI’s GPT-6 models offer the same kind of reasoning setting.
Before any of this, ask whether the job needs a model at all. Our checklist, should this feature be AI, catches the cases where a rule or a regular expression is cheaper and never wrong. And once you have picked, the AI SDKs we reach for explains how to keep the model swappable, so the next price change is a one-line edit.
FAQ
Is ChatGPT, Claude, or Gemini best for a small business?
For staff, ChatGPT Business and Claude Team cost the same, $20 per seat per month billed annually. Pick the one your team prefers after a two-week trial, and check the admin controls you need. Gemini makes sense if you already pay for Google Workspace. For a feature inside your software, test models from at least two vendors, because the best one depends on your task.
What is the cheapest good AI model for business tasks?
GPT-6 Luna and Claude Haiku 5.5, both $0.10 per million input tokens and $0.50 per million output tokens, checked 2026-10-13. They suit extraction, classification, routing, and short replies. Test them on your own inputs before relying on them for customer-facing answers.
Do I need the most powerful AI model?
Rarely. Top models such as Claude Fable 5.1 and GPT-6 Astra cost $10 / $50, five times the mid tier. Anthropic’s guide suggests moving to Fable 5.1 only “if your evals at xhigh or max effort still fall short.” Most business tasks never get there.
Can I switch AI models later?
Yes, if you plan for it. Keep prompts in your code, route all model calls through one place, and keep your test set. Then a switch is a re-run of the test and a one-line change. Vendors retire models on fixed dates, so you will switch at some point anyway.
Does the AI model train on my business data?
On the paid API from OpenAI, Anthropic, and Google, not by default, per each vendor’s own data pages. Free tiers and consumer plans differ: Google uses free-tier Gemini API prompts to improve its products, and ChatGPT’s pricing page says content on its Free, Go, Plus, and Pro plans is used to train its models, with an opt-out.
We help businesses pick, test, and wire in the right model as part of our AI integration work, with the test results in writing before anything ships.