E-Commerce AI API Guide: Product Descriptions, Chatbots, and Recommendations (with Transparent TRY Cost Math)

In e-commerce, an AI API pays for itself fastest in three jobs: product description generation, customer support chatbots, and classification/recommendations. With the right model class, 10,000 product descriptions cost roughly ₺165 (measured August 5, 2026); on Onysoft AI Gateway one OpenAI-compatible key reaches 708+ models on a Turkish lira balance.

E-commerce is one of the sectors where AI repays itself quickest: content for thousands of SKUs, the support queue in peak season, and margin-setting classification decisions are all text work. Yet the same job can be billed at more than a 25x difference depending on the model you pick — which is why "which model class for which job" must be answered before "which API".

This guide maps each use case to the right model class with prices, converts the cost of 10,000 product descriptions to lira with the math fully shown, shares working code that generates product descriptions, and covers Turkish language quality and KVKK (customer data masking) at a practical level.

Which Jobs Does an AI API Solve in E-Commerce?

The fastest returns come in three areas: catalog content generation, customer support automation, and product data classification/recommendations. All three run on the same API schema, usually through the same key; only the model choice and the prompt change.

• Product descriptions and catalog content: Turning raw supplier attribute lists into SEO-friendly copy in your brand voice; variant texts, category page descriptions, campaign emails. The job best suited to bulk generation — volume is high and the template is clear.
• Customer support chatbot: Order status, returns and exchanges, shipping, and product questions. A chatbot fed by your product catalog and order system shortens the support queue in peak season; only cases that genuinely need a human reach an agent.
• Classification and recommendations: Automatic category assignment for new products, attribute extraction, sentiment analysis on customer reviews, search query intent, and tagging for "similar product" suggestions. Tokens per request are small but request counts are very high — unit price is everything here.

What the three scenarios share: none of them wants a single "best model". Each has a different right model class, and the invoice is determined by that mapping.

Which Model Class Fits Which Use Case?

The rule is simple: high-volume templated work goes to the economy class, real-time customer dialogue to the speed class, and brand-critical copy to the flagship class. The mapping below uses actual sale prices from the live api.onysoft.com catalog:

Use case	Class	Example model	Input ($/1M)	Output ($/1M)
Bulk product descriptions, tagging	Economy	deepseek/deepseek-v4-flash	$0.21	$0.42
Bulk product descriptions, review analysis	Economy	openai/gpt-5.6-luna	$0.15	$0.90
Support chatbot	Speed	anthropic/claude-haiku-4.5	$1.50	$7.50
Support chatbot, multimodal tasks	Speed	google/gemini-3.6-flash	$2.25	$11.25
Brand-critical copy, complex complaints	Flagship	anthropic/claude-sonnet-5	$3.00	$15.00

Measured: August 5, 2026 — api.onysoft.com live catalog.

For chatbots, speed is a real constraint: seconds matter while a customer waits. Here is a measurement from our own production traffic — across 6,959 successful requests through the gateway in the last 14 days, average end-to-end response time was 3.8 seconds, the fastest 0.3 seconds. Long-generation flagship requests pull the average up; short support replies on speed-class models drop below a second. If you would rather not pick models by hand, set the model field to onysoft/auto: OnyRouter analyzes each request and routes it to the right class for free.

Worked Example: What Do 10,000 Product Descriptions Cost in Lira?

Under our assumptions, between ₺110 and ₺2,853 — the model choice is what decides it. The math is fully open: we assume an average of 500 input tokens per product (attribute list + system prompt) and 300 output tokens (a 150-200 word Turkish description). For 10,000 products that totals 5 million input + 3 million output tokens.

Example with openai/gpt-5.6-luna: (5M × $0.15) + (3M × $0.90) = $0.75 + $2.70 = $3.45. In lira: 3.45 × 47.555 = ₺164.06. Applying the same formula to the other models:

Model	Total (USD)	Total (TRY)	Per product
deepseek/deepseek-v4-flash	$2.31	₺109.85	~₺0.011
openai/gpt-5.6-luna	$3.45	₺164.06	~₺0.016
anthropic/claude-haiku-4.5	$30.00	₺1,426.65	~₺0.143
google/gemini-3.6-flash	$45.00	₺2,139.98	~₺0.214
anthropic/claude-sonnet-5	$60.00	₺2,853.30	~₺0.285

Measured: August 5, 2026 — api.onysoft.com live catalog. TRY equivalents calculated at the August 5, 2026 TCMB rate (1 USD = 47.555 TRY). Token assumptions are illustrative; run your own catalog through the cost calculator.

In practice a hybrid strategy usually wins: generate the whole catalog with an economy model, then rewrite the 500 highest-traffic products with a flagship. Under the same assumptions that is 9,500 products on Luna ($3.28) + 500 on Sonnet ($3.00) ≈ $6.28, i.e. about ₺299 — under a tenth of running the whole catalog on a flagship, with flagship quality on your showcase products.

A rough chatbot scenario with the same transparency: 5,000 support conversations per month, assuming ~10,000 input + 1,500 output tokens per conversation (history is re-sent with every message). That totals 50M input + 7.5M output; on gpt-5.6-luna $14.25 ≈ ₺677.66/month (~₺0.14 per conversation), on gemini-3.6-flash $196.88 ≈ ₺9,362.39/month (~₺1.87 per conversation). At your traffic level, the gap between the two classes can single-handedly change the feasibility of a chatbot project.

How It Works on Onysoft: Description-Generating Code and Cost Safeguards

Only base_url changes in your existing OpenAI SDK code; the script below works as-is with any model in the catalog. Generate an sk-ony- prefixed key from the dashboard and top up a lira balance:

from openai import OpenAI

client = OpenAI(
    base_url="https://api.onysoft.com/v1",  # the only line that changes
    api_key="sk-ony-YOUR_KEY",
)

product = {
    "name": "Stainless Steel French Press 600 ml",
    "features": "double-wall body, dishwasher safe",
    "category": "Kitchen > Coffee Gear",
}

response = client.chat.completions.create(
    model="openai/gpt-5.6-luna",  # any model in the catalog
    messages=[
        {"role": "system", "content": "Write an e-commerce product description: "
            "150-200 words, turn features into benefits, avoid hype adjectives, "
            "use the product name in the first sentence."},
        {"role": "user", "content": str(product)},
    ],
)

print(response.choices[0].message.content)

The same request with curl:

curl https://api.onysoft.com/v1/chat/completions \
  -H "Authorization: Bearer sk-ony-YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-5.6-luna",
    "messages": [
      {"role": "system", "content": "Write an e-commerce product description: 150-200 words."},
      {"role": "user", "content": "Stainless Steel French Press 600 ml; double-wall body"}
    ]
  }'

In bulk generation the real question is not "does the code run" but "can the budget run away". The safeguards at the gateway layer work like this:

• Pre-request balance check: On every request the estimated cost, with a 1.2x buffer, is compared against your balance; if it does not cover the request, a 402 is returned before anything reaches a provider. When the balance runs out the system stops automatically — no negative balance, no surprise invoice. You can run the 10,000-product script overnight and sleep well.
• Per-key model token limits: Give the content script's key its own cap, managed independently of the chatbot key.
• Real cost in every response: The cost field returns the request's actual cost in USD; the dashboard shows the lira equivalent. Usage reports filter by date/model/status, download as PDF, or go out as a logo-branded email report — handing accounting a monthly breakdown is one click.

To try it without writing code, run the same product prompt across several models side by side in the in-panel Playground; schema details are in the API documentation.

How Do You Guarantee Language Quality in Turkish Product Copy?

The biggest quality lever is not the model but a good system prompt and a 50-product pilot test. Economy-class models generally produce adequate Turkish for templated product descriptions; the risk is publishing to 10,000 products without measuring first.

Typical issues to watch in Turkish: translation-flavored phrasing, suffix and inflection errors, inflation of the same adjectives ("perfect", "unique"), and misused category jargon. Most of this is solved at the prompt level:

• Add a brand voice definition (casual/corporate, target audience) and a banned word list to the system prompt.
• Spell out length, heading format, and the "turn features into benefits" rule; include two example descriptions (few-shot).
• Pilot: Generate the same 50 products with two or three models, read them side by side in the Playground, have the category owner score them. Roll out to the rest of the catalog with the winning model + prompt pair.

For premium segments or brand-sensitive categories (e.g. your showcase products), the hybrid strategy from the previous section kicks in: on those products the output difference of a flagship such as anthropic/claude-sonnet-5 is easy to justify at roughly ₺0.29 per product.

KVKK and Invoicing: Customer Data Masking, Lira E-Invoices

Product description generation involves no personal data; where KVKK attention is genuinely needed is in scenarios where customer data flows — chatbots and review analysis. The practical rule: never send the model any personal data it does not need to do the job.

• Mask or pseudonymize names, phone numbers, addresses, and emails before adding them to a request (e.g. customer_7421); resolve order numbers in your own system, not in the model.
• Instruct the chatbot in its system prompt not to request personal data from users.
• State log and conversation retention periods in your privacy notice.

The counterparty question should also be settled: with Onysoft, the party to your data-processing agreement is Onysoft Veri Merkezi A.Ş. (İzmir), established in Türkiye — your privacy notice names a concrete counterparty under Turkish law. The full framework is in the KVKK-compliant AI guide.

Invoicing matters for e-commerce companies: after registering, enter your tax number (VKN) in the billing details and click "Query at GİB" — the registry is queried live; if you are a registered taxpayer your company title auto-fills and an e-Fatura is issued, otherwise an e-Arşiv invoice. Invoices are in lira at the TCMB rate; the KDV-2/withholding obligations that arise with foreign AI subscriptions do not occur with a domestic invoice. This section is not legal or tax advice; consult your own KVKK and tax advisors.

To get started, open a free account and generate your first product description in the Playground; for the broader picture see the AI API guide.

Last updated: August 5, 2026 · Data: api.onysoft.com live catalog