OpenAI API Pricing Explained (2026)

By Gia Gray · Updated September 2026 · 8 min read

OpenAI's lineup has turned over again: the current generation is GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna, and the GPT-4 / o-series models are now legacy. I still see teams accidentally running deprecated models, paying for caching they never enabled, and reaching for the top-tier model on tasks a cheaper tier handles identically. Here's what I actually use to pick a model and estimate what it'll cost — with real numbers.

OpenAI's Current Model Lineup and Pricing

Source: openai.com/api/pricing, September 2026. GPT-4o and GPT-4o mini are still callable but legacy; the o-series, GPT-4-class, and prior GPT-5.4/5.5 models have been folded into the GPT-5.6 line. Always verify before production use.

GPT-5.6 Terra: The Workhorse for Most Production Use Cases

GPT-5.6 Terra is the model most teams should default to — fast, capable, multimodal, with a 400K context window, at $2.00/$12.00 per million tokens. For most text-based production workloads it's the sensible high-quality default.

Output pricing ($12/M) is where cost accumulates. If your app generates long responses — document drafting, detailed code explanations, multi-step reasoning — it adds up fast. A workload generating 500 tokens per response at 50,000 requests/month is already $300 in output alone.

Terra is worth paying for when response quality directly affects the user experience, when you're doing complex reasoning or instruction-following, or when multimodal input is part of the pipeline.

GPT-5.6 Luna: The Right Default for High-Volume Work

GPT-5.6 Luna at $0.20/$1.20 per million tokens is one of the best deals in the OpenAI lineup — roughly 10× cheaper than Terra and strong at classification, extraction, summarization, basic Q&A, translation, and simple code completion.

The common mistake is treating Luna as "Terra but worse." It's better framed as a different category — good enough for the large majority of tasks, drastically cheaper for the ones where quality differences don't reach users. Teams that route aggressively to Luna often cut their bill 50–70% with no meaningful user-facing impact.

GPT-5.6 Sol and Reasoning / Thinking Tokens

GPT-5.6 Sol ($5/$30) is the frontier tier for the hardest coding, math, and multi-step agentic work. Like the whole current generation, it leans on thinking tokens — internal reasoning generated before the visible answer, billed at the output rate. On hard problems those can dominate the cost of a single request.

That shifts the metric that matters from cost-per-token to cost-per-correct-answer: a pricier model that gets it right the first time can be cheaper end-to-end than a cheap one you re-run three times. Reserve Sol for tasks where the quality difference is real, and budget for thinking tokens explicitly on reasoning-heavy workloads.

Prompt Caching: ~90% Off Repeated Input

OpenAI caches repeated prompt prefixes (a long system prompt or document sent across many requests) and charges cache hits at roughly 10% of the normal input rate — a ~90% discount, in line with Anthropic and Google.

This matters a lot for large, consistent system prompts. A 2,000-token system prompt across 100,000 requests/month is 200M input tokens. On GPT-5.6 Terra that's $400 uncached vs roughly $40 on cache hits. Caching kicks in automatically for prompts sharing a common prefix of at least 1,024 tokens — just structure prompts so the stable parts come first.

Batch API: 50% Off Non-Realtime Requests

OpenAI's Batch API processes requests asynchronously (within 24 hours) at half price. For workloads that don't need immediate responses — bulk document processing, overnight analysis, bulk content generation — it halves your cost with no quality tradeoff.

Batch pricing on GPT-5.6 Terra: $1.00 input / $6.00 output per million tokens. On GPT-5.6 Luna: about $0.10 / $0.60. Low rates for capable models.

What Model Should You Use?