OpenAI API Pricing: The GPT Lineup Compared

OpenAI fields 6 models at list prices from $0.20/$1.20 to $5.00/$30.00 per million tokens in and out (August 2026). On a real workload the gap between the ends is 25.0×.

The GPT lineup

OpenAI runs the widest lineup of the three and the widest spread between its cheapest and most expensive model. That breadth is useful — there is almost always a rung that fits — but it also means the naming does not track the pricing, and two models with adjacent version numbers can differ several-fold in cost.

The output multiplier stays in a narrow band across this lineup, 6.00× to 6.25×, so moving up the ladder costs roughly the same proportion whatever the shape of your traffic. (What the multiplier means and how to use it.)

What each step up the ladder actually costs

Priced on one real workload — a support chatbot handling 1,000 conversations a month, about 12M tokens in and 800K out:

• GPT-5.6 Luna → GPT-5.4 Nano: $3.36 → $3.40 per month, a 1.01× step
• GPT-5.4 Nano → GPT-5.6 Terra: $3.40 → $33.60 per month, a 9.88× step
• GPT-5.6 Terra → GPT-5.4: $33.60 → $42.00 per month, a 1.25× step
• GPT-5.4 → GPT-5.5: $42.00 → $84.00 per month, a 2.00× step
• GPT-5.5 → GPT-5.6 Sol: $84.00 either way — no step at all, these two are priced identically

The expensive step is GPT-5.4 Nano → GPT-5.6 Terra at 9.88× — more than a doubling for a single rung. End to end, GPT-5.6 Sol costs 25.0× what GPT-5.6 Luna does on this workload — $80.64 a month more for the same thousand conversations.

The same six jobs, across the GPT lineup

Six concrete jobs at OpenAI's rates, cheapest highlighted. (All eight jobs, all 17 models.)

Read across a row rather than down a column. On this lineup the spread is much the same whatever the job — 25.0× for drafting against 25.0× for summarising — so the rung you pick moves your bill by a predictable factor regardless of traffic shape.

On GPT-5.6 Luna an 80% cache hit takes the chatbot workload from $3.36 to about $1.63; on GPT-5.6 Sol, from $84.00 to about $40.80 — which is below what GPT-5.5 costs uncached. In other words, caching on the top rung of this lineup beats dropping to the rung below it without caching.

Price twins inside the lineup

GPT-5.5 and GPT-5.6 Sol are priced identically, at $5.00 in and $30.00 out. Every cost figure on this page applies to both members of that pair without adjustment, which means there is no cost argument to be had between them. Choose on capability, latency, context window or rate limits instead. Cost re-enters the decision only if one of them caches or batches better for your particular traffic shape — worth measuring rather than assuming.

Where OpenAI sits against the other providers

Within each tier OpenAI competes with the same workload priced on Google and Anthropic. On the chatbot job:

The cheapest model tracked here overall is Gemini 2.5 Flash-Lite from Google, at $1.52 on this workload — 2.2× less than OpenAI's cheapest. If a workload is genuinely mechanical and high-volume, that gap is worth a look before defaulting to a single provider.

Standard-tier list prices, short context. Last checked against OpenAI's own published pricing page on 17 August 2026. Cached-input rates, the full 17-model table and the price changelog · all eight benchmark jobs.

Where to draw the line in OpenAI's lineup

OpenAI publishes 6 models here, spanning 25.0× from GPT-5.6 Luna at the bottom to GPT-5.5 at the top on the chatbot workload. That spread is what makes routing worth the engineering here; the general rules are on the tracker, and what follows is where OpenAI's own line falls.

The step worth arguing about is GPT-5.4 Nano → GPT-5.6 Terra. Every other rung on this ladder is a 1.0×/1.3×/2.0×/1.0× move; that one is 9.88×, or $30.20 a month on the chatbot workload. Test whether your hard requests actually need GPT-5.6 Terra before making it the default, because that single decision costs more than every other choice in this lineup combined.

One OpenAI model wins both shapes: GPT-5.6 Luna is cheapest for generation-heavy work ($0.0017) and for context-heavy work ($0.0031) alike, so there is no workload where a different OpenAI rung is the cheaper answer. Ranking within OpenAI is therefore stable — the only crossings on this site happen between providers, where output multipliers differ, and that is what the tier table above is for.

Estimate your own mix with the AI API cost calculator, or price a specific piece of text with the token calculator.