I use Claude Code every day, ship AI features in production, and have paid AI bills for both indie products and enterprise integrations. This post gives you the real per-token pricing math in INR for the models Indian businesses actually use, plus honest guidance on which one wins for which task.
No affiliate links. No sponsorship. Just the numbers as they stand in August 2026, and the reasoning behind them.
If you want to stop reading now:
Now the details.
Prices are per million tokens. USD converted at ₹85 = $1.
| Model | Input price/1M tokens | Output price/1M tokens |
|---|---|---|
| GPT-4o mini | ₹12.75 | ₹51 |
| Claude Haiku 4.5 | ₹68 | ₹340 |
| GPT-4o | ₹212 | ₹850 |
| Claude Sonnet 4.6 | ₹255 | ₹1,275 |
| Claude Opus 4.7 | ₹1,275 | ₹6,375 |
| Llama 3.3 70B (via Together AI) | ₹51 | ₹68 |
A quick refresher: "tokens" are roughly 0.75 of a word. A 1,000-word input is about 1,333 tokens.
Say your AI agent handles 50,000 customer support conversations a month. Each conversation is roughly 500 input tokens (the customer message + a bit of context) and 300 output tokens (the AI's reply).
That works out to:
Here is what the bill looks like on each model:
| Model | Monthly cost | Notes |
|---|---|---|
| GPT-4o mini | ₹1,084 | Cheapest, handles typical support well |
| Claude Haiku 4.5 | ₹6,800 | Slightly better reasoning, still cheap |
| GPT-4o | ₹18,050 | Overkill for most support |
| Claude Sonnet 4.6 | ₹25,500 | Best reasoning, most expensive of the mainstream models |
| Llama 3.3 70B (hosted) | ₹2,295 | Cheaper than GPT-4o mini at higher volume |
At this volume, GPT-4o mini is the obvious choice unless you have a specific reason not to.
The raw per-token price is misleading. Different models use different numbers of tokens for the same task.
A hard example: I ran the same code review task through Claude Sonnet and GPT-4o mini. Sonnet returned a 400-token answer that solved the problem. GPT-4o mini gave a 900-token answer that partially solved it and needed a follow-up prompt.
Total cost:
GPT-4o mini is still cheaper. But for tasks that require careful reasoning, Sonnet can be worth the price premium because it does not need retries.
Rule of thumb: for simple tasks, cheap model wins. For hard tasks, cheap model plus retries can still win, but sometimes the expensive model is cheaper because it needs one call instead of three.
Do not commit to one model. Route different tasks to different models based on difficulty.
An example architecture I use in production:
Real cost breakdown for 50,000 queries/month using this pattern:
vs. running everything through Sonnet: ₹25,500/month.
~9x cost reduction for the same output quality. This is the pattern serious production systems use.
Open source models are "free" to use — you pay for the compute to run them. There are two ways to pay for compute:
Hosted open source is priced per token, similar to closed models but usually cheaper.
At 50,000 messages/month, Llama 3.3 costs about ₹2,295 vs GPT-4o mini at ₹1,084. GPT-4o mini wins.
At 5 million messages/month, Llama 3.3 costs ₹2.3L vs GPT-4o mini at ₹1.08L. GPT-4o mini still wins on price, but you might choose open source anyway for data privacy or vendor independence.
Renting a beefy GPU (like an H100 or a pair of A100s) costs ₹30,000-₹1,00,000/month on cloud providers. You get roughly unlimited inference within that hardware's throughput limits.
At what volume does this break even against paid APIs?
Below that, you are paying for a GPU sitting idle most of the day. Not worth it.
The real cost of self-hosting is not the GPU. It is the ops effort. Someone has to:
If you do not have this expertise in-house, factor in ₹40,000-₹80,000/month for either a devops contractor or the time your engineer will spend on it.
My honest take: unless you have a strong data-privacy requirement, a specific model choice that closed providers do not offer, or genuinely high volume, hosted APIs win in India in 2026.
Indian AI startups are building models tuned for Indian languages and use cases. In August 2026:
Where they win: if a significant fraction of your users interact in Indian languages, these models often outperform Western models per rupee.
Where they lose: for English-only tasks, GPT-4o mini and Claude Haiku remain the value winners.
Try both on your actual traffic before committing. Language performance varies.
Both OpenAI and Anthropic offer prompt caching. If your AI agent has a big system prompt or a repeated context block, prompt caching stores it on their servers and charges you ~90% less for the cached portion on subsequent calls.
Concrete example: an agent with a 3,000-token system prompt handling 20,000 queries/month.
That is a 10x saving for a one-line change in your integration code.
If your vendor is not using prompt caching where it applies, ask why. It is usually laziness, not a technical limitation.
For every AI feature I ship:
If this feels like a lot, this is exactly what a build should include. If your vendor is not doing all seven, they will burn your money.
Want to know what model is right for your specific use case? Message me on WhatsApp with a description of your workload — volume, task type, quality requirement — and I will give you a direct answer.
What is the cheapest AI model for customer support right now?
GPT-4o mini as of August 2026. Handles 80%+ of typical customer support conversations for under ₹0.005 per exchange.
Can I use different models for different tasks in the same product?
Yes, and this is the smart pattern. Route simple tasks to cheap models, hard tasks to expensive ones. Typical cost reduction: 5-10x.
How do I set a hard monthly spending cap?
OpenAI, Anthropic, and Together AI all let you set hard limits in their dashboards. Set them at 2x your expected spend and alert at 80%. If your vendor is not doing this, it is a red flag.
Does using an Indian AI startup save money?
For English-only workloads, usually no. For Indian language workloads, sometimes yes. Test on your actual traffic before deciding.
What is prompt caching and does it actually cut my bill?
Prompt caching stores repeated context on the AI provider's servers. You pay ~90% less for the cached portion. Enable it any time your prompts have a large repeated section (system prompts, long docs). It is nearly always worth it.
Or message me directly for a straight answer on your specific case.