Skip to content
codebrix

Cost guide

How much AI agent development costs, and what each task costs to run

AI agent development cost at Codebrix is an indicative ₹60,000 to ₹2,00,000 ($2,200 to $7,500) when the agent works inside software you already have, and moves into our ₹1,50,000 to ₹6,00,000 web application range when it needs its own app. Running cost is counted per task, because one task can take six or more model calls plus tool fees.

Build, inside your software
₹60,000–2,00,000 indicative
Build, with its own app
₹1,50,000–6,00,000 indicative
Running cost unit
Per completed task
Budget also for
Evals and human review

Updated · By the Codebrix team

Why an agent costs more to run than a chatbot

A chatbot makes one model call per reply. An agent works in a loop: it reads the task, decides which tool to use, calls it, reads the result, then decides again until the job is done or it asks a person. Every pass through the loop is another model call, and each call resends the instructions, the tool definitions and everything the agent has learned so far in that task.

Chatbot and agent, side by side
ChatbotAgent
Unit of workOne replyOne task: triage a ticket, reconcile an invoice, update a record
Model callsOne per replySeveral per task, often 4 to 10
ContextInstructions plus a few passagesInstructions, tool schemas and every tool result so far
Extra feesRareWeb search, file search, third-party APIs
Main riskA wrong answerA wrong action in a live system

That last row is why agents also need evaluation, permission design and human review, which are build costs and running costs a chatbot mostly avoids.

AI agent development cost: what drives the build

Build cost drivers for an AI agent
DriverKeeps cost downPushes cost up
ToolsTwo or three read-only lookupsEight or more tools across several systems
Write accessAgent drafts, a person appliesAgent writes to live systems with rollback
Steps per taskShort, predictable pathsOpen-ended research or multi-system reconciliation
Approval gatesOne approval before any actionRules by amount, customer type and risk level
EvaluationA small set of sample tasksA scored test set run on every prompt or model change
InterfaceRuns inside your existing tool or SlackIts own dashboard, queue and audit screens

Codebrix indicative ranges

An agent added to an existing product or back office is our AI assistant and automation add-on: indicative ₹60,000 to ₹2,00,000 ($2,200 to $7,500), 2 to 4 weeks. The lower end is a single-purpose agent with a few read-only tools and a person approving every action. The upper end has more tools, write access behind approval rules and a scored test set. When the agent needs its own review queue, audit screens and user roles, that is a web application, indicative ₹1,50,000 to ₹6,00,000 ($5,500 to $22,000) over 6 to 16 weeks, with the agent work on top. All ranges are indicative; the fixed price follows a discovery call. The build itself is described on our AI agent development page.

Token cost per task: a worked example

Cost per task = sum over every call in the task of (input tokens × input price + output tokens × output price) ÷ 1,000,000, plus tool fees. Monthly cost = cost per task × tasks per month.

Assumptions for a support-ticket triage agent, stated so you can swap in your own: 6 model calls per task; a fixed 3,000-token prefix on every call (instructions plus tool definitions, including the tool-use system prompt the provider adds, which Anthropic lists as 286 tokens for Claude Sonnet 5.5); the ticket itself at 1,000 tokens; each step adds an 800-token tool result and a 200-token model output to the context; 5,000 tasks a month.

Call 1 reads 4,000 input tokens, call 2 reads 5,000, and so on up to 9,000 on call 6. That is 39,000 input tokens and 1,200 output tokens per task. Note how context growth, not the output, drives the bill.

Example token cost (standard API prices as of November 2026, before tax, no caching)
Model (input / output per 1M tokens)Per task5,000 tasks a month
Anthropic Claude Sonnet 5.5 ($2 / $10)$0.078 + $0.012 = $0.090$450
OpenAI gpt-6.1-terra ($1 / $5)$0.039 + $0.006 = $0.045$225
Anthropic Claude Haiku 5.5 ($0.10 / $0.50, prompts up to 100,000 tokens)$0.0039 + $0.0006 = $0.0045$22.50

The same agent with prompt caching

If steady traffic keeps the 3,000-token prefix in Anthropic's prompt cache, those 18,000 tokens per task are billed as cache hits at $0.10 per million on Sonnet 5.5 instead of $2. Per task: 21,000 uncached input × $2 = $0.042, plus 18,000 cached × $0.10 = $0.0018, plus output $0.012, which is $0.0558. At 5,000 tasks that is about $279 a month instead of $450. We ignored cache-write charges, which Anthropic bills at 1.25 times the input price for a 5-minute cache and which matter more at low traffic.

Tool and API costs on top of tokens

Tool fees from official pricing pages, as of November 2026
ToolPriceNotes
Anthropic web search$10 per 1,000 searchesPlus tokens for the results the model reads
OpenAI web search$10 per 1,000 callsPlus search content tokens at model rates
OpenAI file search$2.50 per 1,000 calls; storage $0.10 per GB a day, first GB freeResponses API only
Anthropic web fetchNo charge beyond tokensA fetched page still adds input tokens
Anthropic Batch API50% off input and outputFor tasks that can wait, such as overnight reconciliation

Your own systems cost too. CRM, enrichment, mapping or payment APIs bill per call or per seat, and an agent can call them far more often than a person would. We count expected calls per task for each paid API during discovery and set hard step limits so a confused agent cannot loop through hundreds of paid lookups.

Evals, monitoring and human review

These are the costs most quotes leave out, and the ones that decide whether the agent is safe to switch on.

  • Evals: a set of real past tasks with known correct outcomes, run against the agent before launch and after every prompt, tool or model change. Building it takes your team's time to collect cases and our time to score them. Running it costs tokens each time.
  • Monitoring: logs of every step, tool call and cost per task, with alerts for failures and cost spikes. Log storage grows with volume.
  • Permissions: separate credentials for the agent with only the access it needs, so a mistake is contained.
  • Human review: someone checks a share of tasks, every high-risk action and anything the agent flags.

Putting a number on human review

Review time per month = tasks × share reviewed × minutes per review. With 5,000 tasks, 10% reviewed and 3 minutes each, that is 1,500 minutes, or 25 staff hours a month. Multiply by your own staff cost per hour. Early on we suggest reviewing far more, then lowering the share as the eval scores and live error rate justify it.

Hidden costs and how to cut cost safely

  • Retries and loops: a tool that times out can make the agent repeat steps. Cap steps per task and retries per tool.
  • Model retirement: providers retire models, and switching means re-running your evals.
  • Tokenizer differences: Anthropic says its newer models produce roughly 30% more tokens for the same text, so measure on the model you will actually use.
  • Untrusted input: emails, web pages and documents the agent reads can contain instructions meant to mislead it. Hardening against that is build time well spent.
  1. Route steps by difficulty: a small model for classification and extraction, a larger one for the hard decision.
  2. Keep tool results short. Return the five fields the agent needs, not the whole record.
  3. Cache the fixed prefix and keep it identical between calls.
  4. Batch anything that does not need an answer within minutes.
  5. Ask whether each step needs a model at all. A plain rule is cheaper and easier to test.

How to get a fixed quote for an AI agent

Bring 20 to 50 real examples of the task with the outcome a good employee would reach, the systems involved and what access is acceptable, your monthly volume, and who will review the agent's work. On the discovery call we map the steps, estimate tokens and tool calls per task from those examples, and say plainly if a fixed workflow would do the job without an agent. You get a fixed-price proposal with milestones in 2 to 4 days, and the API accounts sit in your name, billed to you with no markup. If the agent mostly answers questions from documents, RAG development may be the better fit; for broader workflow work see AI automation, or ask us for a fixed quote.

Frequently asked questions

How much does it cost to build an AI agent?

With Codebrix, an agent added to software you already use is an indicative ₹60,000 to ₹2,00,000 ($2,200 to $7,500), taking 2 to 4 weeks. An agent that needs its own dashboard, review queue and user roles falls into our web application range, indicative ₹1,50,000 to ₹6,00,000. The fixed price follows a discovery call where we map the task.

What does an AI agent cost to run per month?

Multiply cost per task by tasks per month. In our worked example, a six-call triage agent cost $0.090 per task on Claude Sonnet 5.5, $0.045 on OpenAI gpt-6.1-terra and $0.0045 on Claude Haiku 5.5, at November 2026 prices. At 5,000 tasks that is $450, $225 or $22.50, before tool fees and review time.

Why is an AI agent more expensive than a chatbot?

Because one task means several model calls, and each call resends the instructions, tool definitions and every earlier tool result. Context grows with each step, so input tokens add up fast. Agents also call paid tools and APIs, and they need evals and human review because they take actions in real systems rather than only replying.

Is AI agent development cost lower in India?

Our indicative build ranges are published in rupees and dollars, and you can pay in INR by UPI or Razorpay. Running costs do not change with location: OpenAI and Anthropic list API prices in US dollars on their pricing pages. So the token bill depends on volume, model choice and how the agent is designed, not on where you or we are based.

Can we cap what an AI agent spends?

Yes. We set a maximum number of steps per task, a retry limit per tool, rate limits per user and a monthly spend alert with each provider. Tasks that hit a limit stop and go to a person instead of carrying on. The cost per task is logged, so you can see which kinds of task are expensive.

Do we need human review if the agent is accurate?

Start with it. Accuracy on a test set is not the same as accuracy on next month's unusual cases. We suggest reviewing a large share of tasks and every irreversible action at launch, then lowering the share as evals and live error rates support it. Review time is a real cost, so it belongs in the budget from day one.

Sources

  • AI agents

    AI agent development company for agents that act inside your systems

    AI agent development company building agents that use your CRM, email and APIs, with scoped permissions, human approval on risky steps and full traces.

    Read more
  • AI automation

    An AI automation agency for the repetitive work behind your business

    AI automation agency that moves email triage, lead follow-up, CRM updates and WhatsApp replies to AI workflows, with people approving the risky steps.

    Read more
  • Cost guide

    AI chatbot development cost: build price and monthly running cost

    AI chatbot development cost in two parts: our indicative build range and a worked formula for monthly LLM token and WhatsApp running costs.

    Read more
  • RAG development

    RAG development services for answers your team can check

    RAG development services: ingestion, chunking, hybrid search, reranking and permission-aware retrieval, with retrieval and answers measured separately.

    Read more
  • LLM development

    LLM development services for applications built on language models

    LLM development services: choosing hosted or open-weight models, fine-tuning only when it pays, structured outputs, and latency, cost and privacy limits.

    Read more
  • Hire developers

    Hire AI developers who build, test and ship LLM features

    Hire AI developers who build LLM features with retrieval, agents and evals in Python or TypeScript. Paid first milestone, your keys, rate after a call.

    Read more
Estimate Call WhatsApp