How much does OpenAI cost?

Two Price Tags, One Company: Why OpenAI Costs Confuse Everyone

Ask five people what OpenAI costs and you'll likely get five different answers — and all of them can be correct. The reason is structural: OpenAI sells the same models through two completely separate channels, and the pricing logic of one has almost nothing to do with the other.

The first channel is the ChatGPT subscription: a flat monthly fee per person. The tiers span Free ($0, ad-supported), Go ($8, ad-supported), Plus ($20), Pro ($100 or $200), Business Standard ($20/user/month annual) and Business Premium ($100/user/month annual), with Enterprise priced custom via sales. According to a detailed pricing breakdown, these tiers were restructured repeatedly in 2026 alone — the Pro tier split into $100/$200, an $8 ad-supported Go plan went global, and Business Premium was added at 5x Standard price for 5x usage.

The second channel is the API, billed per token with no subscription at all. Here's the part that trips up nearly everyone new to OpenAI:

• Paying for ChatGPT creates no API credits. As one pricing aggregator puts it, subscriptions authenticate a person, not an application — the two products never mix.
• Subscriptions authenticate a person, not an application — so chatbots, pipelines and agents have no subscription path at all.
• API spend scales with traffic, not headcount — a pipeline that doubles its volume doubles its bill, and nobody approves the increase.
• The API key itself is free to create; only usage is billed, with new accounts sometimes receiving trial credits.

The stakes for getting this wrong are real. Vendor research from Zylo found that 78% of IT leaders have experienced unexpected charges tied to AI and consumption pricing, and 60% lack full visibility into generative AI usage across their organizations. Because automated workloads generate consumption without any user interaction, spend can climb even when headcount is flat — which is exactly the scenario most finance teams aren't watching.

The confusion is compounded by how differently the two channels behave. Subscriptions are predictable: $20 is $20, with usage limits and bundled features. The API is a meter: output tokens cost more than input tokens, reasoning tokens are billed as output, and surcharges (Fast mode at 2x, long context at 2x input) can quietly stack on top of list prices. One analysis pegs the break-even around $150–250/month of API spend — below that, the meter is usually cheaper; above it, flat seats often win for interactive work.

If you're building agents or automated workflows on the API side, the model you pick matters more than the plan you pick — and the skills you attach to those agents matter just as much. SkillzBench is building a free directory of vetted agent skills plus a free MCP connector, so you can see which skills actually help before they touch your token bill. For now, the practical takeaway is simple: decide which of your OpenAI workloads are human and which are automated, because each belongs on a different price tag.

What the API Actually Charges: Tokens, Model Tiers, and the 100x Spread

If you're building anything automated on OpenAI's API, the single biggest cost decision isn't how many requests you make — it's which model you pick. The same family of models spans a roughly 100x price range, and most teams discover this only after their first invoice.

The API bills by the token, not by the request. A token is about three-quarters of an English word, so a short sentence runs 10–20 tokens, according to pricing analysis from Zylo. Two mechanics surprise newcomers: output tokens cost more than input tokens, and reasoning tokens — the "thinking" a model does before answering — are billed at the output rate, per OpenAI's pricing structure breakdown.

Model choice, though, dwarfs all of that. Within the GPT-6 family, input prices run from $0.10 per million tokens (Luna) to $10.00 (Astra) — a 100x spread for the same workload. Across the full catalog, BenchLM's pricing tracker puts the range closer to 300x, from GPT-6 Luna at $0.10/$0.50 per million input/output tokens up to GPT-5.5 Pro at $30/$180.

Here's what that spread means in practice, using a real monthly workload of roughly 25 million tokens (20M input, 5M output):

• GPT-6 Luna: ~$2–10/month — the same token volume costs single-digit dollars on the smallest tier
• GPT-5.6 Terra: ~$100/month for the identical workload, per Data Studios' comparison
• GPT-6 Astra: ~$200/month — the flagship runs roughly 100x the small model's bill

The counterintuitive part: cheap doesn't always mean cheap. "A lower-cost model that requires multiple requests can generate higher total spend than a higher-capability model that completes the task in a single step," as Zylo's analysis puts it. A $0.10-per-million model that fails twice and retries three times costs more per successful outcome than a $2 model that nails it in one shot.

That's why the smart play is cost per outcome, not cost per token. Measure what you pay for a completed task — a correct summary, a working code change, a resolved ticket — not the raw token rate. This matters especially for agent builders, where multi-step workflows multiply every inefficiency. If you're wiring up agents, SkillzBench (skillzbench.com) is building a free directory of agent skills with security scanning and a free MCP connector, so you can test what actually improves outcomes before committing tokens to it.

One more thing worth knowing: prices are falling fast. OpenAI cut GPT-6 Sol from $4/$20 to $2/$10 and Luna from $0.20/$1.20 to $0.10/$0.50, with a spokesperson confirming the new rates "carry no expiration date," per Yahoo Finance. The cheap tier keeps getting cheaper — re-check the pricing page before you budget.

Discounts and Surcharges: The Multipliers That Change Your Real Bill

The sticker price on OpenAI's rate card is rarely the price you actually pay. Between caching discounts, batch processing, and a stack of surcharges, the same model can bill anywhere from a fraction of list price to four times it.

On the savings side, three levers matter most. Cached input costs just 10% of the standard input price — a 90% discount on cached token reads that OpenAI has publicly highlighted, per its September 2026 price-cut announcement. The Batch API and Flex processing mode both take 50% off, rewarding you for tolerating slower, asynchronous completion, according to pricing analysis from OptimNow.

The surcharges cut the other way:

• Fast mode doubles the standard rate (2x) when latency matters more than cost.
• Long-context requests over 272K input tokens bill at 2x for input and 1.5x for output.
• Regional processing adds 10% on models released on or after March 5, 2026.

Stack the worst cases and the math turns ugly. A Fast-mode Astra request running over 272K tokens bills at $40 per million input tokens — 4x the $10 list price — before any regional surcharge, per the same OptimNow analysis. For agent builders pushing large system prompts and tool definitions into every call, long-context multipliers can quietly become the default rather than the exception.

Two broader trends reshape the picture. First, prices keep falling permanently: OpenAI cut GPT-6 Sol from $4/$20 to $2/$10 per million tokens in September 2026, with a spokesperson confirming the new rates "carry no expiration date" — part of a pattern that includes an 80% cut to GPT-5.6 Luna in July, attributed to efficiency gains in models, inference systems, and serving infrastructure.

Second, older models no longer save money. BenchLM's pricing comparison shows GPT-5.6 Sol ($4/$20) and GPT-5.4 ($2.50/$15) both costing more than the newer GPT-6.1 Sol ($2/$10) — so defaulting to a previous-generation model now means paying a premium for less capability.

If you're wiring agents to these APIs, the surrounding workflow matters as much as the rate card. SkillzBench is building a free directory of security-scanned agent skills plus a free MCP connector, so you can equip and test the tooling around your model calls without adding another line to the bill.

Subscription or API? Finding Your Break-Even Point

The mistake teams make most often is treating ChatGPT and the OpenAI API as one budget line. They're not: paying for ChatGPT doesn't create a pool of API credits, because subscriptions authenticate a person while the API authenticates an application. Automated workloads — chatbots, pipelines, agents — have no subscription path at all.

So the decision comes down to workload type. Use subscriptions for interactive human work, where bundled features like deep research, files, and Codex add real value. Business plans add managed workspaces, admin controls, security, and analytics that raw API tokens can't replace. Use the API for repeatable backend tasks, where consumption-based billing means you pay only for what runs.

The financial crossover matters too. One analysis of the two channels puts the break-even at roughly $150–250 per month of API spend — below that, metered billing usually wins; above it, a flat seat often costs less for equivalent usage — per ProxyLLM's cost comparison. Treat that figure as a planning estimate, not a guarantee: API prices have been falling fast, with GPT-6 Sol and Luna rates dropping more than 50% in September 2026 according to Yahoo Finance coverage of the announcement.

Most organizations land on a hybrid architecture:

• Subscriptions for people — Plus at $20/month for daily interactive work, with Pro tiers ($100 for 5x usage, $200 for 20x) only for genuinely heavy users.
• Metered API for backend automation — scheduled jobs, agents, and integrations that run without human involvement.
• The API as the safety valve — when a person's workload outgrows their seat's limits, overflow work routes to metered billing rather than forcing a tier upgrade.

That last point matters because API spend scales with traffic, not headcount. A pipeline that doubles its volume doubles its bill with no approval gate, and Zylo's research found 78% of IT leaders have experienced unexpected AI-related charges. Continuous monitoring, not annual budgeting, is what keeps the meter honest.

Finally, audit your subscription seats the same way you audit API keys. A $200 Pro seat delivers 20x a Plus seat's usage — but if someone uses it like a $20 Plus account, downgrading pays for itself in the first month. The same logic applies to Business Premium at $100/user/month versus Business Standard at $20/user/month.

And once you've settled your channel strategy, agent skills can automate the rest of the workflow. SkillzBench is building a free directory of vetted agent skills plus a free MCP connector, so you can browse, scan, and test skills before they ever touch your metered API budget — join the waitlist to be first in line.

Keeping OpenAI Spend Under Control — and What It Means for Agent Builders

The fastest way to overspend on OpenAI isn't a surprise invoice — it's picking the wrong model for the job. Within a single model family, input prices can span a 100x range, and the same token volume that costs around $2/month on a small model like GPT-6 Luna runs roughly $200/month on GPT-6 Astra, according to one pricing analysis.

That spread is your biggest lever. The practical playbook, echoed across model-selection guidance, is simple: start with the cheapest tier and upgrade only when your own evals show a quality gap. A more capable model isn't automatically cheaper per result — a low-cost model that needs multiple retries can outspend a premium model that finishes in one step. Judge cost per outcome, not cost per token.

Before you raise budgets, use the discount levers OpenAI already offers:

• Cached input — repeated prompts bill at roughly 10% of standard input price
• Batch API — 50% off for asynchronous jobs that don't need instant responses
• Flex processing — another 50% discount on non-urgent workloads
• Watch the surcharges — Fast mode bills at 2x, and requests over 272K tokens trigger 2x input and 1.5x output pricing

Monitoring matters just as much as model choice. API bills scale with traffic, not headcount — a pipeline that doubles its volume doubles its bill, and nobody approves the increase. That's why 78% of IT leaders report unexpected AI-related charges, and why consumption pricing demands real-time governance rather than an annual budget line, per vendor research. Set usage alerts, log token counts per workflow, and review them weekly, not quarterly.

There's a reason prices keep falling while discipline still matters: OpenAI has cut rates repeatedly — GPT-6 Sol and Luna dropped more than 50% in September 2026 alone, with the new rates carrying, per an OpenAI spokesperson, "no expiration date." Cheaper tokens reward builders who automate more, not those who monitor less.

That's the agent angle. If you're building agents on these APIs, your cost problem compounds: multi-step workflows multiply token consumption, and the skills you plug into those agents multiply it further — for better or worse. Published research found that roughly 26% of agent skills carry at least one vulnerability, so which skills you install is a security and quality decision, not just a convenience one.

That's exactly the gap SkillzBench is building for. It's creating a free, independent directory of agent skills — every version pinned and security-scanned, every skill benchmarked in a sandbox with a published Lift score — plus a free MCP connector that lets you choose exactly which skills your agent can see. Everything on the platform will be free: the directory, the benchmarks, the connector. If agent skills can automate the rest of your workflow, browse and test free ones on SkillzBench and join the waitlist to get verified skills from day one.