Real Cost of AI Coding in India (2026): ₹ per Month and per Solved Task

Short answer: AI coding in India costs anywhere from ₹0 to over ₹12,000 a month, and the person paying ₹12,000 is not always getting more done. In our model, a student running a coding agent about 200 times a month spends roughly ₹96 on DeepSeek V4.1 Flash, ₹254 with smart routing and ₹886 on Claude Sonnet 5.5. A flat Cursor Pro plan costs about ₹1,980 a month once your bank's forex markup is added, and Claude Pro costs ₹2,000 to ₹2,399.

The more useful finding is that price per token is the wrong number to optimise. What you actually pay for is a solved task, and that depends on how often the model succeeds, how many retries it needs and how long you spend reviewing its code. Measured that way, a model ten times cheaper per token can end up only about five times cheaper per finished task, and your own review time can cost more than the model.

This report combines published 2026 research, current Indian prices and a transparent cost model, so a student, a freelancer or a small team can see what their month will really cost before the bill arrives.

Key findings

Light use is nearly free. About 214 agent steps a month, the average for Cline users in one 2026 gateway study, costs under ₹100 on DeepSeek V4.1 Flash and under ₹900 on Claude Sonnet 5.5.
Heavy use is not. At 1,549 steps a month, the average for Claude Code users in the same study, a frontier model costs ₹6,400 to ₹12,800 a month even with good caching.
Routing beats both extremes. Sending 80% of steps to a cheap model and 20% to a strong one costs about 70% less than running everything on the strong model.
Foreign billing adds up to about 23%. A typical Indian card adds a 3.5% forex markup plus 18% GST on that markup, about 4.13% in total, and some providers add 18% GST on the invoice itself.
Free offers are ending. ChatGPT Go's free year for Indian users began on 4 November 2025, so early sign-ups reach renewal from next week. Gemini's free student offer closed in January 2026, and new GitHub Copilot student sign-ups were paused from 20 April 2026.
Cost per solved task is what matters. In real sessions, only 44% of agent-written code survives into commits, so retries and review time often outweigh the token bill.

Who this report is for

We modelled three Indian personas that match the readers who write to us most often:

Persona	Who they are	Agent steps per month	Basis for the number
Student	B.Tech or BCA student building assignments, a final-year project and a portfolio	214	Average monthly calls per Cline user, Requesty, May 2026
Freelancer or indie dev	Builds client sites, small SaaS products or side projects most weekdays	700	Our midpoint assumption between the two measured averages
Heavy user	Full-time developer or startup founder who works with an agent all day	1,549	Average monthly calls per Claude Code user, Requesty, May 2026

An "agent step" is one request from a coding agent such as Cline, Aider, OpenCode or Claude Code to the model. Each step resends your code and conversation, so it is far larger than a chat message. Our companion guide on what an AI coding agent costs per step explains the four levers behind each step in detail. This report builds on it and asks the bigger question: what does a month, and a finished task, actually cost?

Option 1: pay per token

With pay-per-token access you bring an API key to an open-source agent and pay only for what the model processes. We use the same worked example as our per-step guide: 84,000 input tokens and 1,500 output tokens per step, with a 90% prompt-cache hit rate. Prices are list prices converted at ₹95.04 per US dollar, the rate on 27 October 2026, and exclude GST.

Model	₹ per step (90% cached)	Student (214 steps)	Freelancer (700 steps)	Heavy user (1,549 steps)
DeepSeek V4.1 Flash (off-peak)	₹0.23	₹49	₹161	₹356
DeepSeek V4.1 Flash (peak)	₹0.45	₹96	₹315	₹697
Claude Haiku 5.5	₹0.24	₹51	₹168	₹372
DeepSeek V4 Pro (peak)	₹1.93	₹413	₹1,351	₹2,990
Routed: 80% Flash, 20% Sonnet 5.5	₹1.19	₹254	₹832	₹1,840
Claude Sonnet 5.5	₹4.14	₹886	₹2,898	₹6,413
Claude Opus 5.5	₹8.28	₹1,772	₹5,796	₹12,826
GPT-5.5	₹11.86	₹2,538	₹8,302	₹18,371

Two warnings sit under this table. First, the 90% cache rate is a good month, not a guaranteed one. With no caching, a Sonnet 5.5 step costs ₹17.39 instead of ₹4.14, so the heavy-user month rises to about ₹26,900. Second, reasoning models bill hidden thinking tokens as output, the most expensive kind. Our guide to whether thinking mode is worth it covers how to control that.

DeepSeek's off-peak window is unusually convenient in India. Its peak hours are 6:30 to 9:30 am and 11:30 am to 3:30 pm IST on weekdays, so evening and late-night coding, which is when many students and freelancers actually work, falls in the half-price window. Our DeepSeek price guide for India has the full rupee breakdown.

Option 2: a flat subscription

Subscriptions trade a predictable bill for limits you cannot see until you hit them. Here is what the main AI coding plans cost an Indian buyer in October 2026.

Plan	List price	In ₹ at list	Plus 4.13% card forex cost	Payment in India
GitHub Copilot Pro	$10 a month	₹950	about ₹990	International card
Cursor Pro	$20 a month	₹1,901	about ₹1,979	International card
Claude Pro (India pricing)	₹2,000 a month billed yearly, ₹2,399 monthly	GST included	No forex cost	Card or app store, no UPI reported
GitHub Copilot Pro+	$39 a month	₹3,707	about ₹3,860	International card
Cursor Pro+	$60 a month	₹5,702	about ₹5,938	International card
Claude Max (India pricing)	From ₹11,999 a month	GST included	No forex cost	Card or app store

Anthropic began showing rupee prices for Claude plans on 13 July 2026, with GST included. Cursor lists Pro at $20, Pro+ at $60 and Ultra at $200 a month, and each paid plan includes a usage pool for third-party models billed at API rates once you exceed it. GitHub moved Copilot to usage-based AI Credits on 1 June 2026, where one credit is worth one US cent, and the Pro plan includes 1,500 credits.

What does a subscription actually include?

Because Cursor and Copilot now meter usage at roughly API rates, you can translate their allowances into agent steps. At our worked example's Sonnet 5.5 rate:

Cursor Pro's roughly $20 pool covers about 459 well-cached Sonnet steps, or about 109 if caching misses.
Copilot Pro's 1,500 credits ($15 of usage) cover about 344 well-cached Sonnet steps, or about 82 uncached. Code completions stay unlimited on paid plans.

In other words, a subscription comfortably covers the student persona on a frontier model, but a freelancer or heavy user runs past the allowance and either pays overage or switches to a cheaper model inside the tool. That is exactly the trade-off pay-per-token gives you openly.

The hidden costs of paying a foreign provider from India

Three charges sit between a dollar price tag and your bank statement.

Forex markup. Most standard Indian credit and debit cards charge a 3.5% markup on foreign-currency transactions. Zero-markup cards exist but are not the default.
GST on the markup. 18% GST is charged on the markup itself, not the whole bill. A 3.5% markup therefore costs 3.5% × 1.18 = 4.13% in practice.
GST on the service. Foreign digital services sold to Indian consumers attract 18% IGST. Some providers add it to the invoice, and some include it in a rupee price, as Claude now does. If you hold a GSTIN, the treatment differs under reverse charge.

Stack all three and a $20 plan can cost about ₹2,336 instead of the ₹1,901 the sticker suggests, about 23% more. Card mandates for recurring foreign payments are also rejected often enough that many Indian developers simply cannot keep a subscription running without a zero-forex card. This is general information, not tax advice. Check your own case with a chartered accountant.

Option 3: free tiers and student offers, and when they end

India saw an unusual wave of free AI in 2025, and much of it is now expiring.

ChatGPT Go: OpenAI made Go free for a full year for users in India who signed up from 4 November 2025. If you claimed it early, your renewal at the normal ₹399 a month price is due from next week. Check your UPI AutoPay or card mandate now.
GitHub Copilot Student: still free for verified students, but GitHub moved students to a separate Copilot Student plan in March 2026 and removed direct access to premium models such as Claude Sonnet and Opus. One analysis puts its allowance at 200 AI credits, about $2 of compute. GitHub's documentation also says new student sign-ups were paused from 20 April 2026.
Gemini for students: the free-year student promotion closed in January 2026.
Cursor for students: Cursor has offered eligible university students a free year of Pro. Check eligibility on Cursor's student page, since terms change.

Free tiers are excellent for learning. The catch is that they tend to restrict exactly the strongest models you would want for hard debugging, and they can change without much notice. A small paid fallback, even ₹100 a month, keeps you working when a free plan hits its limit the night before a submission.

Why price per token is the wrong number

Every table above assumes each step produces something useful. Real research says otherwise.

Most agent code is not kept. The SWE-chat study of about 6,000 real coding-agent sessions found that only 44% of agent-written code survived into users' commits, and users corrected, interrupted or reported a failure in 44% of turns.
Even the best agents fail often on realistic tasks. On Meta's SWE-Together benchmark, rebuilt from real interactive sessions, Claude Opus 4.8 passed 63% of tasks against about 78% for the original human solutions, and DeepSeek V4 Pro passed 48%.
Weaker agents need more correcting. SWE-Together found a strong negative correlation, about −0.92, between an agent's pass rate and how often users had to correct it.

So the number that matters is:

cost per solved task = (cost per attempt ÷ success rate) + your review time

A worked example

Assume a typical task takes about 25 agent steps. At the per-step costs above, one attempt costs about ₹11.25 on DeepSeek V4.1 Flash (peak) and about ₹103.50 on Claude Sonnet 5.5. Now assume, for illustration, that Flash solves 40% of your tasks and Sonnet solves 65%. These rates are our assumptions, chosen to sit inside the range the benchmarks above report, so measure your own.

Strategy	Model cost per solved task	Notes
Flash only	about ₹28	Cheapest per task, but you hit 60% failures and must still finish those yourself
Sonnet 5.5 only	about ₹159	Fewer failures, about 5.7x the model cost of Flash
Cascade: try Flash, escalate to Sonnet on failure	about ₹107	About 33% cheaper than Sonnet only, with Sonnet's final success rate

A ten-fold gap per token became a roughly 5.7-fold gap per solved task, because the cheaper model fails more. The cascade wins by paying the strong model only when the cheap one cannot finish. Production research backs the idea: PayPal's RLM-Cascade, which sends simple turns to DeepSeek and hard ones to Claude Opus, cut API cost by 45.8% on a Claude Code workload with no loss of quality on its test set.

Your time is the biggest line item

Now add review. If a freelancer values their time at ₹800 an hour and spends 10 to 20 minutes reading and fixing each agent-written change, review costs ₹133 to ₹267 per task. That is more than the model bill on every strategy in the table above. Team data agrees: Faros AI's telemetry across 22,000 developers found that heavy AI adoption came with pull requests 51% larger, review times 91% longer and 54% more bugs per developer.

The practical lesson for Indian freelancers is blunt. Saving ₹50 on tokens by using a weak model is a bad trade if it costs you an extra half hour of debugging. Saving ₹5,000 a month by not running a frontier model on routine edits is a very good one. Our guide on measuring cost per solved task on your own code shows how to test this in an afternoon.

Recommended setup for each persona

Student

Budget: ₹100 to ₹300 a month.
Setup: use any free student plan you still qualify for, plus a pay-per-token key as a fallback. Run Cline or Aider on DeepSeek V4.1 Flash for everyday work and switch to a stronger model only for the bug you are stuck on.
Why: at 214 steps a month, even full routing costs about ₹254, less than one month of ChatGPT Go. You also learn how agents behave across models, which is a hiring signal in its own right. See our report on AI coding agents and Indian fresher jobs.

Freelancer or indie developer

Budget: ₹800 to ₹1,500 a month.
Setup: route by difficulty. Plan and review with Claude Sonnet 5.5 or similar, and do routine edits on DeepSeek V4.1 Flash or Claude Haiku 5.5. Give each client project its own API key with a monthly cap.
Why: routing at 700 steps costs about ₹832, less than half of Cursor Pro after forex, and per-key caps make it easy to bill AI costs to the right client.

Heavy user or small team

Budget: ₹2,000 to ₹7,000 per developer a month.
Setup: either a rupee-priced plan such as Claude Pro or Max if you live in one tool, or routed pay-per-token if you use several agents. Track cache hit rate weekly, because it moves the bill more than anything else.
Why: at 1,549 steps, routing costs about ₹1,840 against ₹6,413 for Sonnet only. If your work genuinely needs a frontier model on most steps, a flat plan with GST included can be the simpler choice.

Ten ways to cut your AI coding bill

Measure a normal week first. Log steps, cached tokens and output tokens before changing anything.
Route by difficulty. Cheap model by default, strong model for planning, review and hard bugs.
Protect the cache. Keep timestamps and session IDs out of the start of prompts and avoid changing tools mid-session.
Start a fresh session per task. Long sessions resend ever-growing history.
Keep rules files short. A bloated AGENTS.md or CLAUDE.md is sent on every step. See our context engineering guide.
Cap reasoning effort. Default to low or medium and raise it per task.
Code off-peak when using DeepSeek. Indian evenings and nights are half price.
Use a zero-forex card for foreign subscriptions, or pay in rupees, to avoid the 4.13% markup cost.
Set a budget on every key. A looping agent can burn a month's budget overnight.
Track cost per solved task, not per token. It is the only number that tells you whether a cheaper model is actually cheaper.

How to try this with one rupee wallet

Velona is an INR-native AI API gateway. You top up with UPI from ₹10, pay in rupees with no international card, and use one OpenAI-compatible endpoint, https://velona.in/v1, for DeepSeek, Claude, GPT, Qwen and hundreds of other models. That makes the routed setup above a configuration change rather than three separate accounts. You can create one key per agent or client, each with its own budget, and see the rupee cost of every request on your usage page.

Setup steps for Cline, Aider, OpenCode and OpenClaw are in our comparison of open-source coding agents. To build your own router, start with our Python model router tutorial. Create a free Velona account to test your own month.

Frequently asked questions

How much does AI coding cost per month in India?

It depends mostly on how many agent steps you run and which model you use. In our model with good caching, light use of about 214 steps costs about ₹96 on DeepSeek V4.1 Flash, about ₹254 with 80/20 routing and about ₹886 on Claude Sonnet 5.5. Heavy use of about 1,549 steps costs about ₹697, ₹1,840 and ₹6,413 respectively, before GST.

What is the cheapest way to use an AI coding agent in India?

Run a free open-source agent such as Cline, Aider or OpenCode with a pay-per-token key on a low-cost model such as DeepSeek V4.1 Flash, and escalate to a stronger model only for hard tasks. Paying in rupees avoids the roughly 4.13% forex markup cost of a foreign card.

How much is Cursor Pro in India?

Cursor Pro lists at $20 a month, about ₹1,901 at ₹95.04 per dollar. With a typical 3.5% card forex markup plus GST on the markup, it costs about ₹1,979, and about ₹2,336 if 18% GST is also charged on the invoice.

How much is Claude Pro in India?

Anthropic's India pricing, introduced on 13 July 2026, lists Claude Pro at ₹2,000 a month billed annually and ₹2,399 billed monthly, with GST included. Claude Max starts at ₹11,999 a month.

Is GitHub Copilot still free for students in India?

Copilot remains free for verified students through the Copilot Student plan, but premium models such as Claude Sonnet and Opus were removed from direct selection in March 2026, and GitHub paused new student sign-ups from 20 April 2026. Check GitHub's education page for the current status.

When does the free ChatGPT Go year end in India?

The free year began for users who signed up from 4 November 2025, so the earliest sign-ups reach renewal from 4 November 2026. ChatGPT Go normally costs ₹399 a month in India.

What is cost per solved task?

It is the total cost of getting one task done, not one request. It equals the model cost per attempt divided by the success rate, plus the time you spend reviewing and fixing the result. A cheap model that fails often can cost more per solved task than a stronger one.

Is a subscription or pay-per-token cheaper for coding?

Pay-per-token with routing is cheaper for most students and freelancers. A subscription can be better for heavy users who need a frontier model on most steps and want one predictable bill, especially a rupee-priced plan with GST included.

How we calculated this

Per-step cost: 84,000 input and 1,500 output tokens per step, 90% of input served from cache, the same worked example as our per-step cost guide. Claude cache writes are priced at 1.25 times the input rate.
Prices: provider list prices converted at ₹95.04 per US dollar (27 October 2026), excluding GST unless stated.
Steps per month: 214 and 1,549 from Requesty's May 2026 study of coding-agent traffic. 700 is our own midpoint assumption.
Routing: 80% of steps on DeepSeek V4.1 Flash at peak rates and 20% on Claude Sonnet 5.5.
Cost per solved task: 25 steps per attempt, with illustrative success rates of 40% and 65%. The cascade assumes the strong model's success rate is unchanged on tasks the cheap model failed.
Forex: 3.5% markup plus 18% GST on the markup, the common default on Indian cards.

Limits of this report

This is a cost model built from published prices and published usage averages, not a controlled benchmark on one codebase. Success rates, cache hit rates and steps per task vary widely by repository, language and habit. Subscription allowances change often, and the Copilot Student allowance comes from a third-party analysis rather than GitHub's own pricing page. Treat every number here as a starting estimate and replace it with a week of your own usage data.

Sources

The Coding Agent Economy (Requesty, May 2026)
SWE-chat: Coding Agent Interactions From Real Users in the Wild (arXiv 2604.20779, April 2026)
SWE-Together: Evaluating Coding Agents in Interactive User Sessions (arXiv 2606.29957, June 2026)
RLM-Cascade: Response-Level Speculative Decoding for Cost-Efficient LLM API Serving (arXiv 2606.22840, June 2026)
Are AI coding assistants really saving time? (Faros AI)
Cursor Pricing 2026: Pro, Pro+, Ultra and the Usage Pools (Krater, read 3 October 2026)
Cursor vs Copilot pricing, June 2026 (Morph)
Anthropic Claude INR pricing in India, July 2026 (explainx.ai)
Claude India pricing 2026 compared with US list (ecorpit)
Access GitHub Copilot for free as a student (GitHub Docs)
GitHub Copilot Student plan removes direct access to premium models (Roboin, March 2026)
Student developer perks in 2026 (BuildMVPFast)
ChatGPT, Gemini and Perplexity Pro free in India (Digit)
Forex markup and GST on Indian credit cards (CardTrail)
The results of the 2026 Developer Survey (Stack Overflow, October 2026)
What does an AI coding agent really cost in India? (Velona)
Are you overpaying for AI? The real monthly cost of AI in India (Velona)

The Real Cost of AI Coding in India (2026): Students, Indie Devs and Cost per Solved Task in ₹