Verdict: to build AI application prototypes and early production backends, start on the Gemini API with gemini-3.8-flash. Google lists it at $0.75 per 1M input tokens and $3.75 per 1M output tokens through December 31, 2026, which is 2.67x cheaper than Claude Sonnet 5 at $2 and $10. Choose Claude only when a task needs Opus-class depth and the budget can absorb it.
TL;DR
- Gemini 3.8 Flash costs $0.75 input and $3.75 output per 1M tokens until December 31, 2026, then rises to $1.50 and $7.50.
- Even after the rise, it stays 25% below Claude Sonnet 5 at $2 and $10.
- A sample workload of 1,000 requests a day costs $45 a month on Gemini and $120 a month on Sonnet 5.
- Setup takes one API key from Google AI Studio and one
pip install. - Indian businesses registered for GST generally self-assess 18% IGST on these imported services under reverse charge.
- Budget for the January 2027 price change now, not after launch.
Which API is cheaper when you build an AI application?
Gemini, on both input and output. Prices below are per 1M tokens, taken from Google's pricing page and Anthropic's pricing page. Opus 5.5 comes from Anthropic's rate card effective 22 September 2026.
| Model | Input (USD) | Output (USD) | Notes |
|---|---|---|---|
| Gemini 3.8 Flash (intro, to 31 Dec 2026) | $0.75 | $3.75 | Output includes thinking tokens |
| Gemini 3.8 Flash (from 1 Jan 2027) | $1.50 | $7.50 | Standard pricing |
| Gemini 3.1 Flash-Lite | $0.25 | $1.50 | Ultra-budget option |
| Claude Sonnet 5 | $2 | $10 | Standard rate |
| Claude Opus 5.5 | $4 | $20 | Batch is half price |
Google also states that the free tier is free of charge, and lists context caching at $0.075 per 1M tokens during the introductory period. For a prototype with a few hundred test calls a day, the free tier may be all you need.
One caveat: token price is not the same as task cost. A cheaper model that needs two attempts to get a correct answer erases part of its advantage. Run your own prompts on both before committing.
What does a realistic monthly bill look like?
In our own worked analysis, an app handles 1,000 requests a day, each with 500 input tokens and 300 output tokens. Over a 30-day month that is 15M input tokens and 9M output tokens. Using the prices in the table above:
- Gemini 3.8 Flash, introductory: 15 x $0.75 plus 9 x $3.75 gives $45 a month, using Google's listed prices.
- Gemini 3.8 Flash from 1 January 2027: 15 x $1.50 plus 9 x $7.50 gives $90 a month, using Google's listed prices.
- Claude Sonnet 5: 15 x $2 plus 9 x $10 gives $120 a month, using Anthropic's listed prices.
- Claude Opus 5.5: 15 x $4 plus 9 x $20 gives $240 a month, using Anthropic's rate card.
The doubling in January is real, so plan for $90 rather than $45 if the app will still be running. Even then Gemini sits below Sonnet 5. Thinking tokens count as output on Gemini, so a request set to a high thinking level will cost more than the arithmetic above suggests. The model supports low, medium and high thinking levels, with medium as the default, so start at low for simple classification and extraction.
How do I build an AI application on the Gemini API step by step?
The process is short. Google's quickstart covers it, and the steps below add the Windows details.
- Open Google AI Studio and create an API key. It creates a project and key for you at no charge.
- Install the SDK with
pip install -U google-genai. - Store the key as the environment variable
GEMINI_API_KEY. - Make a first call.
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-3.8-flash",
input="Summarise this support ticket in one sentence: ..."
)
print(interaction.output_text)
The model has a 1M-token context window and 64k maximum output, which is enough to pass in a whole codebase or a long document set. Moving to the paid tier needs Cloud Billing and, per the quickstart, a minimum prepay of $10.
How do I set this up on Windows or WSL?
Pick one of the two routes. Do not mix them, because Windows environment variables are not automatically visible inside WSL.
PowerShell (native Windows):
- Run
setx GEMINI_API_KEY "your-key-here"in PowerShell. - Close the terminal and open a new one, since
setxonly affects new sessions. - Create a virtual environment with
python -m venv .venvand activate it using.venv\Scripts\Activate.ps1. - Run
pip install -U google-genai, then your script.
WSL (Ubuntu or similar):
- Run
python3 -m venv .venv && source .venv/bin/activate. - Add
export GEMINI_API_KEY="your-key-here"to~/.bashrcand runsource ~/.bashrc. - Run
pip install -U google-genai. - Confirm the key is visible with
python3 -c "import os; print(bool(os.environ.get('GEMINI_API_KEY')))", which should printTrue.
Keep the key out of version control. A .env file listed in .gitignore is the usual approach.
How does GST apply to Gemini and Claude API bills in India?
A supplier based outside India selling to an Indian business is treated as an import of services. The GST Council's sectoral FAQ for IT and ITeS puts IT services at 18% and says a GST-registered Indian recipient pays IGST under reverse charge. That reverse-charge tax is paid through the cash ledger, and eligible input tax credit can be claimed afterwards. Corroborating guidance points to the same 18% rate for foreign software imports.
At about Rs 95 to the dollar (the market rate on 3 November 2026 was roughly 95.3), the worked example above becomes:
| Option | Monthly cost before GST | With 18% GST |
|---|---|---|
| Gemini 3.8 Flash, introductory | about Rs 4,275 | about Rs 5,045 |
| Gemini 3.8 Flash, from 1 Jan 2027 | about Rs 8,550 | about Rs 10,089 |
| Claude Sonnet 5 | about Rs 11,400 | about Rs 13,452 |
If you are a registered business, the GST is largely recoverable as credit, so the pre-tax figure is closer to your true cost. If you are an individual developer without registration, the treatment can differ, and OIDAR rules apply to some providers directly. Confirm with a chartered accountant, because this article is not tax advice.
When should I pick Claude instead?
Choose Claude when output quality on hard reasoning or long, multi-step engineering work outweighs price. Opus 5.5 at $4 and $20 per 1M tokens costs more, but a single correct answer can be cheaper than several rejected ones. A sensible pattern is to route routine calls to Gemini and escalate the difficult ones to Claude.
Note that API pricing differs from subscription pricing. For chat plans, see our guide to Claude Opus price in 2026. For tools that wrap these models, read about the best AI agent for coding and agentic AI coding tools. If you only need a design front end, the Google Stitch and Gemini 3.8 Flash guide is closer to what you want, and anyone avoiding code entirely should compare AI website builders.
FAQ
Q: What is the cheapest way to build an AI application in 2026?
A: Start on the free tier of the Gemini API, then move to gemini-3.8-flash at $0.75 input and $3.75 output per 1M tokens while the introductory price lasts.
Q: Will Gemini 3.8 Flash get more expensive?
A: Yes. Google's pricing page shows $1.50 input and $7.50 output per 1M tokens from January 1, 2027.
Q: Is Claude Sonnet 5 worth the higher price?
A: For many routine tasks it is not, since it is priced at $2 and $10 per 1M tokens. It earns its place when your own tests show fewer failed attempts than the cheaper model.
Q: Do I need a credit card to start with the Gemini API?
A: No, creating a key in Google AI Studio is free. Paid usage needs Cloud Billing and a minimum prepay of $10.
Q: Do I pay GST on API usage from India?
A: Generally yes, at 18% for IT services, usually self-assessed as IGST under reverse charge by registered businesses. Check your own case with an accountant.
Prices and tax rules were last verified on 3 November 2026 and can change. This article was drafted with AI assistance and reviewed by the editorial team before publication. See how we work for our sourcing and correction policy.
Top comments (0)