OpenAI API Pricing 2026: Every Model, Ranked by Cost

OpenAI’s API catalog now spans a 100x price range: $0.10 per million input tokens at the bottom, $10 at the top. Which model you pick matters more than how much you use — the same workload costs $1.80 a month on one model and $180 on another. Below is the complete ladder: every OpenAI model we track, at official list prices, ranked by one standard workload.

Prices are OpenAI’s official API list prices only — the same verified dataset behind our calculator and our OpenAI provider page. Gateway resellers and ChatGPT subscriptions are separate schedules and are not mixed in. Workload used for ranking: 10M tokens per month, split 8M input / 2M output, no caching, standard tier.

Disclosure: this article contains no affiliate links. Where CostMyTokens does use affiliate links elsewhere, they never affect order — every table is sorted by price only. Read our disclosure.

The full ladder

Source for all rows: OpenAI’s official pricing page, cross-checked against our dataset on the dates shown. Batch halves the input and output rates of every model in this table except chat-latest, which has no Batch support. Cached input is the exception to “exactly half”: OpenAI’s Batch table lists a halved cached-input rate for the GPT-6, GPT-5.6 and GPT-5 Mini rows, but no Batch cached-input rate at all for GPT-4.1, GPT-4.1 Mini, GPT-4o and GPT-4o Mini — see batch API discounts explained for how batch windows work.

The GPT-6 family

Five of the fourteen rows are GPT-6-branded models, and they are not priced alike. The current lineup is GPT-6 Luna ($0.1 / $0.5) for high-volume classification, extraction and chat where cost per call dominates; GPT-6.1 Terra at $1 / $5 (cached $0.05) as the step between Luna and Sol, at exactly half Sol’s rates; GPT-6.1 Sol at $2 / $10 as the mainstream tier; and GPT-6 Astra at $10 / $50 as the frontier tier. GPT-6 Sol remains listed at the same sticker as GPT-6.1 Sol. All five share a 1.05M-token context window, so choosing between them is a price (and generation) decision, not an architecture decision.

One easy row to misread: GPT-6 Sol and GPT-6.1 Sol have identical sticker prices ($2 / $10, $36.00 at this workload) but different cached-input rates — $0.2 vs $0.1 per 1M. On a cache-heavy workload that reuses a long system prompt, GPT-6.1 Sol reads the cache at half the price. For the full GPT-6 breakdown — tiers, worked examples and where Astra earns its premium — see our GPT-6 API pricing guide.

The GPT-5.6 family

The 5.6 line uses the same Luna / Terra / Sol names one generation back: GPT-5.6 Luna at $0.2 / $1.2, GPT-5.6 Terra at $2 / $12, and GPT-5.6 Sol at $4 / $20. Watch GPT-5.6 Terra’s output rate (not to be confused with the cheaper GPT-6.1 Terra): at the same $2 input as GPT-6.1 Sol, its $12 output makes output-heavy work cost more — $40.00 vs $36.00 at the standard mix, and the gap widens as the output share grows. Input price alone would have told you these two models cost the same.

The value tier

Five models land under $7 a month at the standard workload: GPT-6 Luna ($1.80), GPT-4o Mini ($2.40), GPT-5.6 Luna ($4.00), GPT-5 Mini ($6.00) and GPT-4.1 Mini ($6.40). The practical difference between them is the output rate and the caching discount. GPT-5 Mini’s output ($2) is more than triple GPT-4o Mini’s ($0.6), so chatty workloads drift toward the Minis; the GPT-6 and 5.6 Lunas pair the lowest input rates with deep caching discounts (cached input 90% off list: $0.01 and $0.02 per 1M). A workload with a stable, reused prompt is cheaper on a Luna even when its sticker price looks similar to a Mini’s. The next step up is GPT-6.1 Terra at $18.00 — ten times Luna, but half of GPT-6.1 Sol and well under the legacy GPT-4.1 ($32.00).

The legacy lines: GPT-4o and GPT-4.1

GPT-4o is still listed at $2.5 / $10 — and on every column of the table, GPT-4.1 is cheaper: $2 input, $8 output, and cached input at $0.5 instead of $1.25. GPT-4o’s caching discount is also among the shallowest in the catalog (cached input only 50% off, a rate matched only by GPT-4o Mini, vs 75–95% on the newer lines). If an integration is still pinned to GPT-4o for behavior reasons, that is a quality decision; on price alone there is no row where it wins against GPT-4.1. The Mini versions follow the same pattern in reverse order of age: GPT-4.1 Mini ($0.4 / $1.6) costs more than GPT-4o Mini ($0.15 / $0.6), so for the Minis the older model remains the budget pick.

Three billing traps

• Service tiers multiply the sticker price. OpenAI sells speed as separate Fast and Ultrafast tiers on top of standard rates — Ultrafast GPT-6 Astra is $60 / $300 per 1M, six times standard. These tiers are easy to enable in a console and hard to notice on an invoice; check which tier your keys are actually on.
• Long prompts reprice. Prompts above roughly 272K input tokens enter a long-context segment where input and cached rates roughly double (output rises about 1.5x on the GPT-6 / 5.6 lines). A workload that is cheap in testing can quietly cost much more in production once conversations grow.
• Caching is the counterweight — if your prompt is stable. Cached input runs from $0.01 (GPT-6 Luna) to $1.25 (GPT-4o) per 1M, a 90–95% discount on the GPT-6 and 5.6 lines. The same fixed ranking reshuffles once caching enters the picture — our prompt caching guide works through the math.

What is not in this table

OpenAI’s open-weight models (such as GPT OSS 120B) carry no official OpenAI API price — they are priced by the gateways that host them, which is why they appear in our cheapest LLM API ranking as gateway rows instead. ChatGPT subscriptions (Plus, Pro, Team) are a separate product with usage limits rather than per-token prices, and are not comparable line-by-line with this table. The one ChatGPT-branded API model, chat-latest (row 13), is in the table: it points to the Instant model currently used in ChatGPT, costs $5 / $30, has no Batch support, and changes snapshot over time — OpenAI itself recommends GPT-6 Astra for production. If you searched for the “ChatGPT API”, our ChatGPT API pricing explainer covers what that term really means and worked app costs. Models OpenAI lists but we do not yet track — the GPT-5.5, GPT-5.4, GPT-5.2, GPT-5.1 and GPT-5 SKUs (including Pro and Nano variants), GPT-4.1 Nano, the o-series, and specialized models such as gpt-5.3-codex — are omitted until they land in our pricing dataset.

Run your own numbers

This ladder uses one fixed mix. Your input/output split, cache hit rate and call volume will move models up and down it — plug your real numbers into the CostMyTokens calculator or open any model’s page above to cost that model on its own.