Reading the OpenAI price list without going cross-eyed
OpenAI ships more API models than anyone else — my pricing feed counts over forty active GPT variants — but the lineup collapses into a simple shape once you see it: one current generation (5.5), one workhorse generation (5.4) in three sizes, and a long tail of older 5.x models that mostly exist for backward compatibility.
Every price is quoted per million tokens, input and output billed separately. The per-message column in the table assumes a realistic chat turn: 1,500 tokens in, 400 out. One thing to watch with OpenAI specifically: the output multiplier varies by model — 6x on GPT-5.5 and 5.4, about 8x on GPT-5.2, 6.25x on nano. Claude, by contrast, holds a uniform 5x across the lineup. The multiplier decides which model wins for output-heavy work, a dynamic I unpack in the output multiplier piece.
Every price is quoted per million tokens, input and output billed separately. The per-message column in the table assumes a realistic chat turn: 1,500 tokens in, 400 out. One thing to watch with OpenAI specifically: the output multiplier varies by model — 6x on GPT-5.5 and 5.4, about 8x on GPT-5.2, 6.25x on nano. Claude, by contrast, holds a uniform 5x across the lineup. The multiplier decides which model wins for output-heavy work, a dynamic I unpack in the output multiplier piece.
GPT-5.5 and 5.5 Pro: the frontier tier
GPT-5.5 ($5 / $30, roughly 1M context) is OpenAI's mainstream flagship and scores 74 on the Arena leaderboard — statistically tied with Claude Opus 4.8, which costs the same on input and $5 less per million on output.
GPT-5.5 Pro is the price tag that startles people: $30 / $180 per million, six times the cost of regular 5.5. A single long agentic session that consumes 2M input and 200K output tokens runs $96 on Pro versus $16 on standard 5.5. My rule from pricing these out: Pro is for tasks where a single better answer is worth dollars, not cents — deep research synthesis, hard one-shot reasoning — and almost never for anything running in a loop. If you're tempted by Pro for volume work, run the numbers in the calculator first; the monthly delta is usually a salary line, not a rounding error.
GPT-5.5 Pro is the price tag that startles people: $30 / $180 per million, six times the cost of regular 5.5. A single long agentic session that consumes 2M input and 200K output tokens runs $96 on Pro versus $16 on standard 5.5. My rule from pricing these out: Pro is for tasks where a single better answer is worth dollars, not cents — deep research synthesis, hard one-shot reasoning — and almost never for anything running in a loop. If you're tempted by Pro for volume work, run the numbers in the calculator first; the monthly delta is usually a salary line, not a rounding error.
The 5.4 family: where production workloads should start
GPT-5.4 ($2.50 / $15, 1M context) scores 73 on Arena — one point behind GPT-5.5 at exactly half the price. That one point costs you double. For most production work the quality difference is invisible and the bill difference is not.
GPT-5.4-mini ($0.75 / $4.50, 400K context) is the volume workhorse: drafting, summarization, structured extraction, support replies. GPT-5.4-nano ($0.20 / $1.25) is for classification, routing, and anything where you'd otherwise reach for a fine-tuned small model — at under a tenth of a cent per message, the API call is cheaper than the engineering time to avoid it.
The pattern worth internalizing: within a generation, each size step down costs roughly 3-4x less. Cascading — try mini first, escalate to 5.4 or 5.5 only when a confidence check fails — captures most of the quality at a fraction of the spend. I sketched that architecture in the model routing piece.
GPT-5.4-mini ($0.75 / $4.50, 400K context) is the volume workhorse: drafting, summarization, structured extraction, support replies. GPT-5.4-nano ($0.20 / $1.25) is for classification, routing, and anything where you'd otherwise reach for a fine-tuned small model — at under a tenth of a cent per message, the API call is cheaper than the engineering time to avoid it.
The pattern worth internalizing: within a generation, each size step down costs roughly 3-4x less. Cascading — try mini first, escalate to 5.4 or 5.5 only when a confidence check fails — captures most of the quality at a fraction of the spend. I sketched that architecture in the model routing piece.
Caching and batch: OpenAI's two big discounts
Cached input — the repeated prefix of your prompts (system message, tool definitions, few-shot examples) — bills at a steep discount off the regular input rate once OpenAI recognizes it. For chat products that resend the same scaffolding every turn, effective input costs often land far below the list price. The general technique is the same as Anthropic's, covered in the caching guide.
The Batch API gives a flat 50% off both meters for jobs that can wait up to 24 hours. Evals, embeddings backfills, nightly classification runs — anything asynchronous should be batched by default. Between caching and batch, a pipeline that does both routinely lands at a quarter of naive list price.
The Batch API gives a flat 50% off both meters for jobs that can wait up to 24 hours. Evals, embeddings backfills, nightly classification runs — anything asynchronous should be batched by default. Between caching and batch, a pipeline that does both routinely lands at a quarter of naive list price.
What real workloads cost on OpenAI
Scenarios I've priced at June 2026 rates.
A support chatbot, 10,000 conversations a month (about 9,000 cumulative input tokens and 1,200 output per conversation): roughly $122/month on GPT-5.4-mini, $405 on GPT-5.4, $810 on GPT-5.5.
Classifying 1M short records (300 tokens in, 20 out each) on nano: about $85 per run at list prices, $42.50 with batch.
An agentic coding workflow burning 2M input and 150K output tokens per session: $7.25 on GPT-5.4, $14.50 on GPT-5.5, $87 on 5.5 Pro. Multiply by your team's session count before picking a default; the spread is the whole story. More scale math in the 1M-requests cost breakdown.
A support chatbot, 10,000 conversations a month (about 9,000 cumulative input tokens and 1,200 output per conversation): roughly $122/month on GPT-5.4-mini, $405 on GPT-5.4, $810 on GPT-5.5.
Classifying 1M short records (300 tokens in, 20 out each) on nano: about $85 per run at list prices, $42.50 with batch.
An agentic coding workflow burning 2M input and 150K output tokens per session: $7.25 on GPT-5.4, $14.50 on GPT-5.5, $87 on 5.5 Pro. Multiply by your team's session count before picking a default; the spread is the whole story. More scale math in the 1M-requests cost breakdown.
OpenAI vs the field in June 2026
Head-to-head at the frontier: GPT-5.5 ($5 / $30) versus Claude Opus 4.8 ($5 / $25) versus Gemini 3.1 Pro ($2 / $12). Same input price as Anthropic, but Google undercuts both — Gemini 3.1 Pro scores 75 on Arena at 40% of GPT-5.5's input price.
At the budget tier the fight is GPT-5.4-mini ($0.75 / $4.50) against Claude Haiku 4.5 ($1 / $5) and Gemini 3.5 Flash ($1.50 / $9) — mini wins on price, Flash punches hardest on quality score (73, frontier-adjacent). And below all of them sits DeepSeek V4 Flash at around $0.10 / $0.20, the subject of the DeepSeek pricing guide.
The takeaway: OpenAI is competitive but no longer automatically the value pick at any single tier. Check the live comparison table before defaulting.
At the budget tier the fight is GPT-5.4-mini ($0.75 / $4.50) against Claude Haiku 4.5 ($1 / $5) and Gemini 3.5 Flash ($1.50 / $9) — mini wins on price, Flash punches hardest on quality score (73, frontier-adjacent). And below all of them sits DeepSeek V4 Flash at around $0.10 / $0.20, the subject of the DeepSeek pricing guide.
The takeaway: OpenAI is competitive but no longer automatically the value pick at any single tier. Check the live comparison table before defaulting.