OpenAI API pricing (May 2026)

Every OpenAI model and its per-token price

All prices in USD per 1 million tokens. Last reviewed 2026-05-15. Provider pricing pages are authoritative, confirm before contracting.

How OpenAI pricing actually works

OpenAI API pricing in 2026 follows the same shape across every model family, a per-1M-token input rate, a per-1M-token output rate (typically 4-6× input), an automatic cached-input rate at 25% of the headline, and an optional batch mode at 50% off. The flagship-to-cheapest spread is roughly 600× ($30 input on GPT-5.5 Pro down to $0.05 on GPT-5 nano), which means model routing is far more cost-relevant than negotiating volume discounts.

The OpenAI API is the most widely deployed LLM API in production. Roughly 65% of developers using AI tools touch it daily according to the 2025 Stack Overflow survey, and the API powers most of the Fortune 500's customer-facing AI products. Microsoft's Azure OpenAI variant serves the same models inside an enterprise procurement and compliance surface, which is the default in regulated industries.

OpenAI offers 50% batch discount, automatic prompt caching, and ZDR on enterprise contracts. Azure OpenAI exposes the same models with VPC residency.

Prompt caching: the 90% discount most teams ignore

OpenAI's prompt caching is automatic: no opt-in required. Any prefix of 1,024 or more tokens that has been seen in the last 5-10 minutes is served from cache at 25% of the normal input rate. The cache hit rate depends on how prompts are structured: stable content (system prompt, tools, RAG context) at the start, dynamic content at the end. Reverse that and you bust the cache on every turn.

On a typical coding agent run that re-sends the same 200K-token codebase across 10 turns, prompt caching reduces effective input cost by 80-90%. The cached-input column in the table above is the right number to plug into a production budget; the headline input rate is the "new conversation" rate, not the steady-state rate.

Batch inference, and half-price overnight

The batch API accepts a JSONL file of up to 50,000 requests and returns completions within 24 hours at 50% of normal pricing for both input and output. There is no minimum size. The right workloads are anything where a few hours of latency is fine; nightly evaluation runs, document enrichment pipelines, synthetic training data generation, large-scale classification.

Batch + cache stack. The combined effective rate for a cache-warm, batched call is often 5-10% of the headline price. For workloads like nightly eval suites, large-scale classification, document enrichment, and synthetic data generation, batching is free money.

Four real production cost scenarios

The routing pattern that cuts OpenAI spend 60-80%

The dominant pattern in 2026 is intent-based routing. Simple queries, and chat, classification, extraction. route to GPT-5 nano or GPT-5.5 mini at sub-cent cost. Workhorse queries route to GPT-5.5 at $5/$30. Hard-reasoning queries (math, multi-step planning, code refactoring) promote to GPT-5.5 Pro. The routing decision is usually made by a tiny classifier model running in front, or by tool-call signals from the previous turn.

A typical production fleet settles into a 70/25/5 split. 70% of requests handled by the smallest competent tier, 25% by the mid-tier workhorse, 5% promoted to the flagship. Done well, this cuts model spend 60-80% versus naive single-model use without any measurable quality drop on the bulk of requests.

With an AI gateway in front, the routing rule is one config block: declare a default model, declare promotion triggers, declare a fallback to a second provider for availability. Applications keep using a single OpenAI-compatible endpoint. See Swfte for a managed runtime that bundles the gateway, observability, eval, and per-team cost ceilings.

Enterprise considerations

Enterprise procurement for OpenAI typically goes through one of three channels. Direct OpenAI Enterprise contracts include zero data retention by default, dedicated capacity options, and a customer success motion. Azure OpenAI exposes the same models with Microsoft enterprise terms, regional residency, and integration with Entra ID / Defender. AWS Bedrock does not offer GPT-5.5, Bedrock is for Anthropic, Meta, Mistral, and AI21 models. For multi-cloud or VPC residency, route through a gateway that can address all three.

• Prompt caching: Available — use it from day one; the headline rate is misleading without it.
• Batch inference: Available, 50% discount, up to 24h SLA.
• Fine-tuning: Supported on most tiers.
• On-prem / VPC: Available via Bedrock / Vertex / Azure or direct VPC contract.
• Zero data retention: Available; default on enterprise contracts.

How OpenAI compares to the rest of the market

Against Claude (Anthropic), OpenAI is broader: multimodal, voice, image generation, fine-tuning, open weights; but Claude is preferred for coding and agent loops. Against Gemini (Google), GPT-5.5 leads reasoning but Gemini wins on price (Gemini 3.0 is 4× cheaper than GPT-5.5 at comparable quality on most general workloads). Against DeepSeek, GPT-5.5 leads quality on the hardest benchmarks but DeepSeek V4 Pro is ~1/8 the cost for similar Arena Elo on bulk workloads.

For a full side-by-side, see the API pricing index and the AI model leaderboard for quality / speed / value rankings.

Frequently asked questions about OpenAI API pricing