OpenAI Pricing Explained 2026: Every Plan, Every Model, Real Math

Executive summary

OpenAI's pricing in 2026 splits into two worlds. ChatGPT plans (Free, Plus, Pro, Team, Enterprise, Edu) are seat-based and predictable — useful for end-user productivity but not where most enterprise AI cost sits. The API side is consumption-based, billed per token, and is where production AI applications actually spend money. Understanding the difference, and the four discount levers on the API side, is the single highest-leverage thing a business leader can do to bring AI spend under control.

This guide covers both sides, with real-workload math and budgeting heuristics. Model your specific usage in the LLM cost calculator.

ChatGPT plans (seat-based, end-user)

| Plan | Price | Best for | | ---------- | ------------------ | -------------------------------- | | Free | $0 | Limited GPT-5 and 4o access | | Plus | $20/user/month | Individual power users | | Pro | $200/user/month | Researchers, heavy reasoning use | | Team | $25–$30/user/month | Small teams, SSO basic | | Enterprise | Custom | SSO, audit logs, data controls | | Edu | Custom | Universities |

Use seat plans for human-in-the-loop productivity. Move to the API the moment a workflow becomes programmatic — the unit economics flip hard.

API pricing (consumption-based, where production cost sits)

| Model | Input ($/1M) | Output ($/1M) | Context | Use case | | ------------ | ------------ | ------------- | ------- | ---------------------------- | | GPT-5 nano | $0.05 | $0.40 | 400k | High-volume routine | | GPT-5 mini | $0.25 | $2.00 | 400k | Default for most production | | GPT-5 | $1.25 | $10.00 | 400k | Balanced flagship | | GPT-5.4 | $2.50 | $15.00 | 1.05M | Long context, hard reasoning | | GPT-5.4 pro | $30.00 | $180.00 | 1.05M | Hardest reasoning only | | GPT-4.1 nano | $0.10 | $0.40 | 1M | Legacy long-context | | o3 | $10.00 | $40.00 | 200k | Reasoning model (legacy) | | o4-mini | $1.10 | $4.40 | 200k | Cheaper reasoning |

Output is 4–10× input on every tier. Plan for that ratio.

The four discounts that change the bill

Prompt caching (50% off cached input). OpenAI caches automatically for prompts over 1,024 tokens that share a common prefix. No code change required — just structure prompts with stable content first, variable content last. For workloads with consistent system prompts, this drops effective input cost by 30–60%.

Batch API (50% off both input and output). Submit a JSON file of requests; results come back within 24 hours. Use for evals, nightly summarization, embedding jobs, content moderation backfills, and any workload tolerant of async completion.

Realtime API (separate pricing, audio-included). For voice applications, the Realtime API has its own pricing schedule with input/output and audio rates. Budget separately from text workloads.

Volume tiers and enterprise commits. Significant spend ($50k+/month) opens up rate-card discussions. OpenAI rarely publishes these but they exist.

Real-workload math

B2B SaaS customer-support chatbot, 500k conversations/month, 2,000 input / 400 output tokens.

GPT-5 standard: $1.25 × 1,000 + $10 × 200 = $3,250/month. GPT-5 mini standard: $0.25 × 1,000 + $2 × 200 = $650/month (5× cheaper). GPT-5 mini with 70% input cached: effective input = $0.25 × 300 + $0.125 × 700 = $162.50; output $400; $562/month.

For this workload GPT-5 mini with caching wins by 6× over GPT-5 default — same vendor, same code path, different tier and config.

Document analysis, 100k reports/month, 20,000 input / 3,000 output tokens.

GPT-5: $1.25 × 2,000 + $10 × 300 = $5,500/month. GPT-5.4 (long context): $2.50 × 2,000 + $15 × 300 = $9,500/month. GPT-5 with batch API: $2,750/month.

Pick GPT-5 standard if latency matters, batch if it doesn't.

Agentic coding workflow, 5k sessions/month, 80,000 input / 15,000 output tokens.

GPT-5.4: $2.50 × 400 + $15 × 75 = $2,125/month. o4-mini reasoning: $1.10 × 400 + $4.40 × 75 = $770/month. GPT-5.4 pro: $30 × 400 + $180 × 75 = $25,500/month.

The flagship-pro tier is rarely justified unless the work is provably 10× more valuable than the cheaper alternative.

Costs most teams miss

Function-calling and tool overhead. Tool schemas count as input tokens on every call. With 10 tools defined, expect 1,500–3,000 input tokens of overhead before the user prompt.

Embeddings. text-embedding-3-small at $0.02 per million tokens looks free but at billion-token corpora becomes a real line item.

Fine-tuning. Training is per-1M tokens (often $25–$50/million for GPT family). Hosting fine-tuned models adds a per-token premium versus the base model — typically 1.5–3× the base rate. Rarely justified in 2026 given the flexibility of base models.

Failed completions. OpenAI does not charge for HTTP errors but does charge for outputs that pass validation but fail downstream — 5–15% of bills in poorly-instrumented systems.

A practical budgeting heuristic

For most B2B AI features, expect $0.001–$0.005 per end-user interaction with disciplined tier selection and caching. If your math is much higher than that, the leverage is in the four discounts above, not in switching vendors. See our GPT vs Claude comparison for cross-vendor benchmarks and the Anthropic pricing guide for the equivalent breakdown on the other side.

Bottom line

OpenAI's API pricing is more navigable than it looks. The line items that matter are: which tier you default to, whether caching is enabled, whether batch is used where eligible, and whether your prompts are sized like production code or like prototypes. Teams that get those four right typically run at 30–50% of the cost of teams that do not — at the same quality and the same vendor.

Run your real prompts through the calculator to see what your bill should be.