2026-07-09 · Levi

ChatGPT API Pricing 2026: Comparison with Claude, Gemini and HKD Conversion

Two completely separate pricing systems — confusing them is the most common source of budget errors.

ChatGPT API Claude API Gemini API LLM cost comparison

Businesses searching for "ChatGPT pricing" usually face two completely separate pricing systems at once: subscription monthly fees and API usage-based billing. Confusing the two is the most common source of budget errors. This article separates the two logics first, then lists July 2026 API prices verified directly against official pricing pages for the three major providers, with HKD conversion for a real business scenario.

Subscription vs API: Two Separate Pricing Logics

Subscription (ChatGPT Plus, Business, Enterprise, etc.) bills per user per month, suitable for staff using the web or app interface directly. ChatGPT Plus is US$20/month; Business and Enterprise bill per user count — check OpenAI's official pricing page for details.

API bills by token volume, with input and output priced separately, suitable for integrating AI into company systems — quotation workflows, document processing, customer service replies, data extraction. API usage requires a separate balance in an API account, settled independently from subscriptions. The two accounts are completely separate.

One-line test: staff typing and using it themselves → subscription; system calling it automatically → API. Enterprise automation projects are almost universally the latter.

2026 Official API List Prices from Three Providers

The following prices were verified against all three providers' official pricing pages on 5 July 2026 (USD per million tokens):

GPT-5.4 nano (Economy)Input $0.20 · Output $1.25

Gemini 3.1 Flash-Lite (Economy)Input $0.25 · Output $1.50

Claude Haiku 4.5 (Economy)Input $1.00 · Output $5.00

Gemini 3.5 Flash (Standard)Input $1.50 · Output $9.00

GPT-5.4 (Standard)Input $2.50 · Output $15.00

Claude Sonnet 4.6 (Standard)Input $3.00 · Output $15.00

Gemini 3.1 Pro (Premium)Input $2.00 · Output $12.00

Claude Opus 4.8 (Premium)Input $5.00 · Output $25.00

GPT-5.5 (Premium)Input $5.00 · Output $30.00

Output costs 5–6× more than input. Output length is the largest budget variable; limiting response length is the most direct cost control lever.

Tier selection affects cost by multiples. Same workload, input unit price ranges from US$0.20 (GPT-5.4 nano) to US$5.00 (GPT-5.5/Opus 4.8) — a 25× spread. Placing classification and extraction tasks on economy models while reserving complex reasoning for premium models is the largest cost lever.

Discount structures are consistent. All three providers' Batch APIs (non-real-time bulk processing) offer approximately 50% discounts. Cached repeated content reads at approximately 10% of standard input price.

A Real Scenario in HKD

Assume a document processing workflow: 8 pages per document (input ≈ 7,000 tokens), 600 tokens structured output, 300 documents per month, ×1.3 system overhead, exchange rate 7.8:

GPT-5.4 nano≈ HK$0.02 per item
≈ HK$7/month (300 items)

Gemini 3.5 Flash≈ HK$0.16 per item
≈ HK$48/month (300 items)

GPT-5.4≈ HK$0.27 per item
≈ HK$81/month (300 items)

Claude Sonnet 4.6≈ HK$0.30 per item
≈ HK$91/month (300 items)

Tens to a hundred HKD per month at this scale makes one point: at typical SME document volumes, model fees are a minor item in overall costs. The real cost is in the other two layers — human review and one-time system integration. The complete calculation method is in How to Calculate AI Automation Costs.

Three Things to Watch Beyond the Bill

Reasoning tokens. Multiple newer generation models bill their internal "thinking" process at output prices — Google's official pricing page explicitly states that output prices include thinking tokens. A short answer can carry a longer bill. Budget accordingly.

Long document pricing tiers. OpenAI and Google set higher rates for very long context requests (e.g., Gemini 3.1 Pro charges $4/$18 for requests exceeding 200K tokens). Anthropic's current main models use flat pricing for 1M context. For workflows frequently handling multi-page documents, this difference affects model selection and costs.

Tokeniser differences. The same text produces different token counts across models — identical list prices may produce different actual costs. The reliable method for precise budgeting is to test with a small batch of real documents.

Price Validity

All prices in this article were verified against official pricing pages on 5 July 2026 (OpenAI: developers.openai.com, Google: ai.google.dev, Anthropic: platform.claude.com). LLM prices change frequently — re-verify against official pages before finalising budgets.