Most businesses pick an AI model the way they pick a phone: the newest flagship. For automation that runs thousands of times a month, that habit can multiply your bill by 100 with no visible gain.
One job, six prices
Take a common job: reading 10,000 customer emails a month, classifying each one and drafting a reply. Assume about 1,500 tokens in (the email plus instructions) and 300 tokens out per email. That is 15 million input tokens and 3 million output tokens a month.
| Model | List price per 1M tokens (in / out) | Monthly cost for this job |
|---|---|---|
| Claude Haiku 5.5 | $0.10 / $0.50 | $3 |
| Claude Sonnet 5.5 | $2 / $10 | $60 |
| Gemini 4 Argon (introductory price) | $2 / $10 | $60 |
| Claude Opus 5.5 | $4 / $20 | $120 |
| Claude Fable 5.2 | $8 / $40 | $240 |
| GPT-6 Astra | $10 / $50 | $300 |
List prices from each provider, before taxes and before prompt-caching discounts, which can cut repeated input costs sharply. Gemini 4 Argon's price rises to $4 / $20 after its introductory period.
For sorting and drafting routine emails, the $3 model and the $300 model will often produce replies a customer cannot tell apart. Spend the difference where it matters.
Match the model to the task
| Task type | Examples | Start with |
|---|---|---|
| High volume, simple | Tagging enquiries, extracting invoice fields, FAQ replies, call summaries | Cheapest tier (Haiku 5.5, GPT-6 Luna, DeepSeek V4.1-Flash) |
| Everyday professional | Proposals, report drafts, sales follow-ups, data clean-up with judgement | Mid tier (Sonnet 5.5, GPT-6 Sol, Gemini 4 Argon) |
| Hard, high-stakes, low volume | Complex code changes, contract analysis, multi-step research, security review | Top tier (Opus 5.5, Fable 5.2, GPT-6 Astra) |
| Data cannot leave your servers | Patient records, legal files, internal financials | Open weights you host yourself (DeepSeek V4.1-Flash, Muse Spark Open, Sarvam) |
Route, don't choose
The better design is not one model but a router. A cheap model handles every request first and passes only the hard cases up to an expensive one. If 90% of the 10,000 emails above are routine and 10% go to Opus 5.5, the bill is about $15 a month instead of $120, and the hard cases still get the strong model.
Five checks before you commit
- Test on your own data. Run 50 real examples through two or three models and compare. Public benchmarks measure other people's tasks.
- Count tokens, not users. API pricing is per token. Long documents and long chat histories drive cost far more than headcount.
- Check where data goes. India's DPDP Rules were notified in November 2025, with most obligations phasing in over about 18 months. Know which provider processes personal data and where.
- Plan for price changes. This quarter alone, Google announced Argon's price will double after an introductory period, while DeepSeek and OpenAI cut prices. Keep the model swappable.
- Keep a human in the loop for anything that sends money, signs a contract or reaches a customer without review.
Want this done for your processes? Our AI Transformation Assessment maps which of your tasks fit which tier, and our free AI readiness audit is a quick first check. Also see every model launched this quarter.
Sources
- Claude Platform Docs: Pricing
- OpenAI Developer Community: Introducing GPT-6 Astra (3 Sep 2026)
- OpenAI Developer Community: GPT-6 Sol and GPT-6 Luna in the API, Codex and ChatGPT (22 Sep 2026)
- Google: Gemini 4 Argon, our next era of frontier intelligence (30 Sep 2026)
- The Next Web: DeepSeek launches V4.1-Flash and retires V4-Pro (Sep 2026)
- Vals AI: Muse Spark Open model page
- Takshashila Institution: MeitY notifies DPDP Rules 2025 (14 Nov 2025)
Checked on 10 November 2026. Model names, prices and availability change quickly; confirm on the provider's page before you budget.
Which AI model should your business use in 2026? A cost-per-task guide