Claude API Integration: Add Claude to Your App

View this page

How to integrate Anthropic's Claude API into existing software: the Messages API flow, JSON outputs, cost per request from Anthropic's October 2026 prices, rate-limit tiers and data terms.

Organisation
WavX Solutions
Telephone
+919310079927

Description

Integration Anthropic Claude API integration: adding Claude to your app or system

A Claude API integration calls Anthropic's models from your server through the Messages API, asks for output in a JSON schema, and checks it before anything is saved. As of October 2026 Anthropic lists Claude Haiku 4.5 at $1 per million input tokens, Sonnet 5.5 at $2 and Opus 5.5 at $4, with cache reads and batch jobs charged far less, so prompt design sets the running cost as much as model choice.

Discuss an integration Last updated 2 October 2026

What the integration connects

Anthropic's API gives your software access to the Claude models. The integration is the service in between: it takes a record or document from your system, sends it to Claude with fixed instructions, receives a structured answer, checks it, and writes it back where your staff work.

Good first uses are narrow and frequent. Reading supplier invoices into purchase entries, tagging support tickets, summarising long customer threads for the account manager, or answering staff questions from a policy manual. Where the work must take actions in other systems, it becomes an AI workflow automation project; where it only adds a feature to an existing screen, it is covered by add AI to your existing software .

Data flow: reading a supplier invoice into a purchase entry

Accounts team uploads invoice PDF in your app

Your server ── fixed system prompt (company rules, ledger list) ── marked for caching

── the PDF as a document block

── output_config.format: json_schema {supplier_gstin, invoice_no,

date, lines[], taxable_value, cgst, sgst, igst, total}

POST https://api.anthropic.com/v1/messages (model: claude-sonnet-5-5)

Response ── stop_reason: end_turn? max_tokens? refusal?

── parse JSON; re-check totals and GSTIN format in code

── usage: input, cache read, output tokens → cost logged

├── totals match ──▶ draft purchase entry, flagged "AI-read, check"

└── mismatch ──────▶ manual queue with the PDF and the AI's reading

The arithmetic is checked by your code, not trusted from the model. A schema makes sure the fields exist and have the right types; it says nothing about whether the numbers add up.

Field mapping

Your system

Messages API field

Notes

Company rules, ledger names, output instructions

system

Stable text; mark it for prompt caching so repeats are billed at the cache-read rate

The document or record

messages (user turn: text or document block)

Only what the task needs

Output shape

output_config.format with type: json_schema

Objects need additionalProperties: false

Model

model

claude-opus-5-5 , claude-sonnet-5-5 or claude-haiku-4-5 ; kept in configuration

Length cap

max_tokens

If reached, stop_reason is max_tokens and the JSON may be incomplete

Outcome

stop_reason

Handle refusal and max_tokens before parsing

Cost tracking

usage (input, cache creation, cache read, output tokens)

Stored per request

Data location

inference_geo ( global or us )

us costs 1.1 times the standard rate

Cost per request

Prices are Anthropic's per million tokens, read on 2 October 2026, in US dollars.

Model (Anthropic's description)

Input

Cache read

Output

One request: 3,000 in, 400 out

10,000 such requests

Claude Haiku 4.5 ("the fastest model with near-frontier intelligence")

$1

$0.10

$5

$0.005

$50

Claude Sonnet 5.5 ("the best combination of speed and intelligence")

$2

$0.20

$10

$0.010

$100

Claude Opus 5.5 ("long-running agentic coding and knowledge work")

$4

$20

$0.020

$200

The calculation is tokens × price ÷ 1,000,000, input and output added, with no caching. Three levers change it:

Prompt caching. A cache read costs 10% of the input price on most models and 5% on Opus 5.5. Writing to the 5-minute cache costs 1.25 times input and to the 1-hour cache 2 times. A long fixed system prompt repeated on every invoice is where this pays.

Batch processing. The Batch API is 50% off input and output, for work that can wait, such as last month's invoices.

The tokenizer. Anthropic says Claude 4.7 and later models produce about 30% more tokens for the same text than earlier ones. Estimate from a real count, not a word count.

Limits and gotchas

Issue

What Anthropic documents

What the integration does

Rate limits

Requests, input tokens and output tokens per minute, per model; at the Start tier, 1,000 requests per minute for Opus 5.5, Sonnet 5.5 and Haiku 4.5

Queue; obey the retry-after header on 429

Cached input

For most models, cache reads do not count towards the input-token rate limit

Cache the stable prefix; it raises throughput as well as cutting cost

Monthly spend caps

Start $500, Build $1,000, Scale $200,000; at the cap, requests return 429 until the next month

Alert well before the cap; request a higher tier ahead of launch

New organisations

May begin on an evaluation tier below the standard limits

Do not plan a launch on day-one limits

Schema features

No recursive schemas, no numeric or string-length constraints; enum capitalisation is not guaranteed

Enforce ranges and compare enums case-insensitively in code

Model retirement

Haiku 4.5's retirement is listed as not sooner than 15 October 2026; Opus 5.5 not sooner than 22 September 2027

Keep a test set so a model change is a re-run, not a rewrite

Inference: global or US only; workspace storage: US only

Record this in your privacy notice; strip data the task does not need

What WavX builds

WavX builds the service around Claude: input preparation and redaction, the system prompt and schema under version control with caching in mind, validation in code of every answer, retries and queueing within the rate limits, per-request cost logging, batch jobs for back-office work, and a review screen for staff. The same service can call OpenAI instead; see OpenAI API integration and the OpenAI, Claude and Gemini API comparison.

Effort band

The AI cost calculator prices an AI feature in an existing app from a ₹2,50,000 base: a planning range of ₹2,12,500 to ₹3,12,500 with no custom data, rising when the feature works on your documents (×1.25) or live systems (×1.5). These are planning ranges, not quotes. Usage is billed by Anthropic to your own account.

When not to integrate the API

If the job is occasional drafting or summarising by a few staff, Claude's own apps cost less than building anything. If the task follows fixed rules (GST rate by HSN code, routing by pincode), write ordinary code. If invoices arrive as structured e-invoice JSON already, read the JSON; a model adds cost and risk to data that is already machine-readable.

Frequently asked questions

Should we use Claude or OpenAI?

Test both on a few hundred of your own real cases; the difference for a given task shows up there, not in benchmarks. Both are priced per token and both offer structured JSON output. We build the service so the model is a configuration setting, which keeps the choice open and lets you move if prices or quality change.

Does Anthropic train on what we send through the API?

Anthropic's privacy centre says that by default it does not use inputs or outputs from commercial products, including the API, to train its models. It also says API inputs and outputs are deleted from its back end within 30 days, with exceptions such as the Files API, custom agreements, policy enforcement and legal requirements.

Can inference run in India?

Not as of October 2026. Anthropic's data residency page lists two inference settings, global and US-only, and US as the only workspace storage location. US-only inference costs 1.1 times the standard rate on Claude 4.6 and later models.

Which model should we start with?

Anthropic's own guidance is to start with Claude Opus 5.5 for most workloads. For a high-volume, narrow task such as extraction or tagging, we test Sonnet 5.5 and Haiku 4.5 against the same cases and pick the cheapest one that passes.

Sources

Claude docs: Pricing · read 2 October 2026

Claude docs: Models overview · read 2 October 2026

Claude docs: Rate limits · read 2 October 2026

Claude docs: Structured outputs · read 2 October 2026

Claude docs: Data residency · read 2 October 2026

Anthropic Privacy Center: How long do you store my organisation's data? · read 2 October 2026

Anthropic Privacy Center: Is my data used for model training? · read 2 October 2026

Related

Add AI to your existing software

OpenAI API integration

AI cost calculator

Build your own software — your way, your pricing.

WavX Solutions is here to create your own software in a fully custom way, built exactly how you work — with a pricing model that fits your business. Connect now and let's build it.

Contact Now helpwavx@gmail.com

Reference material from WavX Solutions

Related questions WavX has answered

Aadhaar ekyc and pan verification api integration in loan app?

Aadhaar e-KYC and PAN verification are added to a loan app through a KYC provider's API rather than built from nothing. The app collects consent and the identity number, the server calls the provider, and the response either fills the applicant's verified details or flags a mismatch for manual review. The lender signs the agreement with the provider and holds the credentials; the developer wires the flow, stores only what your policy allows, masks numbers on screen and logs every check. Which checks you are permitted and required to run is your compliance team's call. Source

Data migration from old software to new system what can go wrong?

Migrations go wrong in predictable ways: duplicate customers that merge badly, dates and numbers stored as text, mandatory fields that are empty in old records, codes that changed meaning over the years, and balances that no longer add up after import. Each is caught by profiling the data first, running trial imports into a test system, and reconciling totals against the old system before go-live. Allow time for your own staff to review the mismatches, because only they know which record is right. Source

How do you connect new software to an old system that has no API?

There are three workable routes when an old system has no API. Read its database directly, ideally through a read-only copy, and write back only through channels the old system already accepts. Exchange files: most old systems can export and import CSV or XML on a schedule. As a last resort, automate the screens with a script that types and clicks like a user, which works but breaks when the screen changes. A small connector service usually sits between the two systems to translate and log. Source

What does WavX Solutions charge for adding a public API to a software build?

A public API is a ₹70k add-on to the base build in WavX's published software cost model, and third-party integrations are a separate ₹60k. Those figures apply inside the estimator alongside a build type, such as an internal tool from ₹2,00,000, and scale. An API added to software written by another developer needs a code review first, so its price depends on what is found. Exposing only the endpoints partners need keeps the work small. The figure is open to discussion; the quote follows a free consultation, requested at helpwavx@gmail.com. Source

What does WavX's cost model show for adding an AI feature to an existing app?

Adding an AI feature to an existing app starts at ₹2.5 lakh in WavX's published cost model, with a range of ₹2.13L–₹3.13L before add-ons, roughly $2,600–3,800 as an approximate conversion; the quote itself is built in INR. A data multiplier applies: 1.25 when the feature reads documents and 1.5 when it connects to live systems. Because it plugs into the app you already have, there is no new platform to pay for. Builds under ₹3 lakh fall in the 3–6 week band. WavX is open to negotiating the figure after a free consultation. Source

How to move item and customer data from marg or busy billing software to a new system?

Export, clean, import, then verify with a parallel run. Take the item master, customer and supplier ledgers and opening stock out of the old software as Excel or CSV files, or from a database backup. Clean before loading: remove duplicate items, fix units and HSN codes, standardise phone numbers and GSTINs. Import into a test copy of the new system and compare stock value and outstanding balances with the old reports. Do a physical stock count at cut-over. Old bills can stay archived in the old software instead of being migrated. Source

My business in India already runs on a custom web system; what does adding a mobile app for it cost?

A companion mobile app is a ₹2,00,000 add-on in WavX's software estimator when the backend and APIs already exist, which is why it costs less than a standalone build. If the web system was built elsewhere or has no APIs, the app is priced through the app model instead, where a simple app starts at ₹1,50,000 (₹1.28L–₹1.88L) and a medium one at ₹4,00,000. The condition of the existing APIs decides which applies. Reusing your backend is the main saving. WavX can negotiate the figure once it reviews the system. Source

Our company runs on an older ERP with no API, just a SQL database. Management wants a mobile app so sales staff can check stock and customer balances and place orders. How do you put an app on a system like that safely?

Put a small API layer between the app and the ERP instead of letting phones touch the database. The layer reads stock and balances through read-only queries or views, caches them, and exposes only what the app needs. Orders from the app go into a holding table or the ERP's own import route, so its business rules still apply, and nothing is written directly into core tables. Add authentication, per-user permissions and an audit log. WavX builds custom APIs and legacy integrations of this kind. Source

WavX starting prices (base scope, before add-on features)

From the cost model behind the WavX calculators. A quote follows a scoping call; features, integrations and scale move these figures. Annual maintenance is typically 15–25% of the build cost.

BuildStarting priceTypical timeline
Landing pagefrom ₹40,0003–6 weeks
Business websitefrom ₹50,0003–6 weeks
E-commerce storefrom ₹1,50,0003–6 weeks
Internal tool / dashboardfrom ₹2,00,0003–6 weeks
Custom web applicationfrom ₹4,50,0006–10 weeks
SaaS productfrom ₹7,00,00010–16 weeks
ERP / CRM systemfrom ₹9,00,00010–16 weeks
AI chatbotfrom ₹1,50,0003–6 weeks
RAG assistant (your data)from ₹3,50,0006–10 weeks
AI agent / automationfrom ₹5,00,0006–10 weeks
AI feature in an appfrom ₹2,50,0003–6 weeks

Calculators: website, software, AI, mobile app. Pricing explained: https://www.wavxsolutions.in/pricing.

About WavX Solutions

WavX Solutions (WAVX — Web Application & Venture Exchange) is a remote-first, 100% custom software development company founded in 2022 in Gurgaon, Delhi NCR, India. WAVX designs, builds, scales and maintains custom websites, web applications, iOS and Android apps, Shopify stores, ERP/CRM software, loyalty systems, AI integrations and performance marketing for startups, businesses, brands and institutions across India.