OpenAI API Pricing Calculator

How OpenAI's pricing actually works

OpenAI charges per token — not per request, not per message, not per «conversation». A token is roughly three-quarters of an English word (the exact ratio varies by language — Russian and Chinese consume substantially more tokens per word, see the language multiplier in the calculator's conversation mode). When you call the API, you pay for two things:

• Input tokens— everything you send to the model: your system prompt, the user's message, any prior conversation context you replay for memory, any retrieved documents from RAG.
• Output tokens — the model's response. Output is typically 3-4× more expensive per token than input across most OpenAI models. This asymmetry matters for cost optimization: long input contexts are cheaper than long outputs.

For a typical SMB customer support chatbot answering 10,000 conversations per month with 8 messages per conversation and 40 words per message, the calculator estimates roughly 480,000 input tokens and 480,000 output tokens per day — or about 14.4M input + 14.4M output tokens per month. At GPT-4o mini pricing, that's $2.16 input + $8.64 output = $10.80 per month. At GPT-4o pricing, the same workload costs $180 per month — a 17× cost increase for the upgrade.

When GPT-4o vs GPT-4o mini vs GPT-3.5 Turbo

The biggest cost-optimization lever in OpenAI pricing is model selection. Use the calculator above to compare exact numbers for your workload, but here's the rule-of-thumb decision framework:

• GPT-4o mini — default for production SMB chatbots. Handles intent classification, FAQ retrieval, lightweight reasoning, and most customer support scenarios at 6% of the cost of GPT-4o. Use this unless you have evidence GPT-4o is measurably better for your specific use case.
• GPT-4o— reach for it when conversation quality is the unique differentiator (premium B2B sales bots, healthcare triage, legal Q&A). The quality improvement over GPT-4o mini is real but workload-dependent; benchmark before committing.
• GPT-5 / GPT-5 mini— currently OpenAI's flagship for complex reasoning, agentic workflows with multiple tool calls, and high-quality content generation. GPT-5 mini is the new «balanced tier» sweet spot.
• GPT-3.5 Turbo— legacy. Useful only for cheap classification tasks where you specifically don't want any reasoning capability. Consider migrating to GPT-4o mini, which is comparable in cost and substantially better in quality.
• o1-preview / o1-mini— reasoning models that «think» before responding. Significantly more expensive per output token because they generate internal reasoning chains. Use only when the reasoning matters (math, complex code review, multi-step planning) — not for chat.

Top 5 cost optimization tactics

1. Enable prompt caching

If your prompts include a long static system instruction (e.g., your bot's persona, brand voice rules, escalation criteria), enable prompt caching. Cached input tokens cost about 10% of full price. For a typical bot with 300-token system prompt and 10,000 calls per month, this saves 20-30% on input cost. Implementation difficulty: easy — just add cache_control markers to your prompt structure.

2. Use Batch API for non-realtime workloads

OpenAI's Batch API gives you 50% off both input and output tokens in exchange for a 24-hour SLA. Perfect for: bulk classification, weekly embedding refreshes, large RAG knowledge base indexing, content summarization at scale. Not suitable for live chatbot conversations.

3. Cap output token length

Set max_tokensaggressively. Most chatbot responses don't need to be more than 200 tokens. Setting a hard cap prevents runaway responses that drive up cost unpredictably. Industry pattern: 200-token cap for support bots, 500-token cap for informational bots, 1500-token cap for long-form content generation.

4. Use smaller models for classification, larger for generation

Architect your bot as a two-stage pipeline: GPT-4o mini classifies the user's intent and routes to a handler; only complex generation handlers escalate to GPT-4o. This routing pattern can cut your overall LLM bill 60-80% while preserving quality for the generation-critical paths.

5. Tier-down embeddings unless quality matters

Use text-embedding-3-small ($0.02 per 1M tokens) instead of text-embedding-3-large ($0.13 per 1M tokens) for typical RAG knowledge bases. The quality difference is real but small — and most production RAG systems are bottlenecked by chunking strategy and retrieval logic, not embedding quality.

OpenAI vs Anthropic vs Gemini — cost landscape

OpenAI is not always the cheapest choice. As of 2026, Anthropic's Claude family (Haiku, Sonnet, Opus) competes directly on the cost-vs-quality frontier. Google's Gemini Flash family is aggressively priced for high-volume workloads. For multi-provider comparison, see our LLM API Cost Calculator which extends this OpenAI-specific calculator to compare across providers.

Most production chatbots today use a multi-model architecture: cheap fast models for classification + intent recognition, premium models for nuanced generation, fallback models for cost-overage protection. The cost calculator above models OpenAI-only deployments; the cross-provider calculator models the multi-vendor architecture.

Related Chatbotscape resources

• Large language model — definition + concepts
• System prompt — what it is + why size matters
• Chatbase review — OpenAI-powered chatbot platform
• Botpress review — multi-LLM chatbot platform
• Voiceflow review — visual designer + OpenAI
• WhatsApp Business API cost calculator
• How we source and refresh pricing data

FAQ

Which OpenAI model is cheapest for chatbots?

GPT-4o mini at $0.15 input + $0.60 output per 1M tokens. Run the calculator above with your specific volume to see exact monthly numbers; the cheapest model can vary by use case when you weight quality.

How accurate is conversation-mode token estimation?

The calculator uses OpenAI's documented heuristic of ~0.75 tokens per English word, with language multipliers for non-English. For production accuracy, switch to token mode with direct telemetry, or integrate js-tiktoken in your application for exact counting.

What is prompt caching?

OpenAI's feature that lets you cache repeated context (typically system prompts) for 5 minutes between calls. Cached portion costs ~10% of full input price. Toggle it in the calculator above to see the effect on your monthly cost.

When should I use Batch API?

For workloads where 24-hour latency is acceptable: bulk classification, embedding generation, content summarization at scale. Never for realtime chatbots. The 50% discount on both input and output makes it the single largest cost lever for batch-friendly workloads.

How often is pricing refreshed?

Weekly, from OpenAI's official pricing page. Current data last refreshed 2026-05-26.

Can I embed this calculator on my site?

Yes — free. Copy the iframe snippet from the embed section above. Embed strips Chatbotscape navigation, keeps the calculator + attribution badge.

Does it include GPT-5 and o1 models?

Yes — full GPT-5 family, o1-preview, o1-mini, plus all GPT-4 family variants and both text-embedding-3 models.

How do I convert monthly conversations to API tokens?

The calculator does this for you in conversation mode using the formula: conversations × messages per conversation × words per message × (0.75 tokens-per-word × language multiplier). Half is treated as input, half as output.