OpenAI's API price list got simpler in 2026. The input rates now form a short ladder: $4.00 per million tokens for flagship GPT-5.6 Sol, $2.00 for the balanced Terra, and $0.20 for the fast, cheap Luna, with output rates of $20, $12, and $1.20 per million tokens respectively. [1] [2] [3] Two dated cuts built that ladder — Luna's input price fell 80% on July 30 and Sol's fell 20% on August 24 — while Terra slid from its launch price of $2.50/$15 to $2/$12 without a dated cut announcement of its own. [1] [2] The sticker, though, only sets the rung: five billing mechanisms behind these rate rows multiply it into the actual bill.

Before you pick a model, map your workload's shape: how much it generates, how often it repeats context, how deeply it reasons, how fast you need responses, and how long the inputs run. The input rate sets the base; those five factors set the multiplier on top of it, and each maps onto a cost component below — generation volume onto the output rate, repeated context onto cache hits, reasoning depth onto billed reasoning tokens, speed onto Fast mode, and long inputs onto the 272K re-tiering threshold.

Every figure in this article is in US dollars per million tokens. The sticker and cache rates on the models tracked by the LLM Abacus price tracker carry an October 11, 2026 verification against OpenAI's official pricing page; Terra's $2/$12 rests on an independent gateway's official-rate listing (see the table footnote), and the long-context and Fast rows carry earlier tracker verification dates — Sol's September 28, 2026, Luna's August 2, 2026, GPT-5.5's processing tiers July 31, 2026, with Terra's rows recorded in the July 30 adjustment note. [1] [2] [3] [4] Re-check every rate on the official pricing page before you commit a budget.

The price card (USD per 1M tokens)

Model Input Output Cached input Over 272K input (in/out) Fast mode (in/out)
GPT-5.6 Sol $4.00 $20.00 $0.40 $8.00 / $30.00 $8.00 / $40.00 [1]
GPT-5.6 Terra $2.00* $12.00* — $4.00 / $18.00 $4.00 / $24.00 [1] [3]
GPT-5.6 Luna $0.20 $1.20 $0.02 $0.40 / $1.80 $0.40 / $2.40 [2]
GPT-5.5 $5.00 $30.00 $0.50 $10.00 / $45.00 — (Batch/Flex ≈ half price; Priority $10/$45) [4]
GPT-5.1 $1.25 $10.00 $0.13 — — [5]

*Terra's standard rate comes from an independent gateway's official-rate listing, cross-checked against the tier multiples on the pricing tracker; its launch price was $2.50/$15. [1] [3] [6] A dash means the listings cited above do not state that rate — read it as unknown rather than zero. Context windows and generations: the GPT-5.6 models carry 1.05M-token context and 128K maximum output with a February 2026 knowledge cutoff; GPT-5.5 (launched December 2025) and GPT-5.1 (launched November 2025, cutoff September 2024) carry 400K-token context and the same 128K output cap. [1] [4] [5]

What changed this year

OpenAI launched the GPT-5.6 series on July 9, 2026 as a limited preview — invite-only, through the API and Codex — at $5/$30 for Sol, $2.50/$15 for Terra, and $1/$6 for Luna. [1] [6] On July 30 it cut Luna by 80% — a cut the Luna tracker listing describes as permanent — to $0.20/$1.20, and dropped the cached-input read to $0.02; the same adjustment introduced the Fast processing tier across the 5.6 lineup, replacing Priority Processing. Contemporary Chinese-language coverage reported the same 80% figure and the $0.20 price. [2] [7] On August 24 Sol's rates fell — input down 20% from $5 to $4, output down a third from $30 to $20, cached input from $0.50 to $0.40 — with the promotional price reported as guaranteed until at least November 21, 2026. [1]

Two consequences matter when you compare articles. First, anything written before late August quotes dead prices; the July launch figures no longer describe the current bill. Second, the naming scheme OpenAI has described separates generations from capability: from GPT-5.6 on, the number marks the generation, and Sol, Terra, and Luna mark capability tiers that can each update on their own schedule. [6] So “GPT-5.6” alone tells you the vintage, not the price.

The five costs behind every API bill

The sticker table has five moving parts. None outranks the tier choice, but each multiplies the rate you pick, and together they decide how far the final bill lands from the sticker.

Output tokens. Output rates run 5–8× the input rate across the lineup — Sol 5×, Terra, Luna and GPT-5.5 6×, GPT-5.1 8×. [1] [2] [3] [4] [5] A generation-heavy workload pays the expensive side of the table on every call, which is why the cheapest-input model is not automatically the cheapest bill.

Cached input. When prompt-cache hits land, input drops to roughly a tenth of the standard rate on the 5.6 models — Sol from $4.00 to $0.40, Luna from $0.20 to $0.02. [1] [2] The tracker's GPT-5.1 notes, citing OpenAI, state that caches are kept up to 24 hours. [5] Repeated context is where this pays: long system prompts, document-search (RAG) pipelines, and multi-turn support conversations.

Reasoning tokens, billed as output. Thinking tokens bill at the output rate, and you can set reasoning_effort to none for the fastest, cheapest runs — per GPT-5.1's documentation as relayed in the tracker's notes. [5] The GPT-5.6 models are reasoning-capable too, so budget for output you never see in the response.

Fast mode, if you pay for speed. Fast is opt-in at call time and costs exactly double the standard rate — Sol Fast runs $8.00/$40.00. OpenAI's own speed claim, “up to 2.5× faster,” attaches only to Sol; Terra and Luna carry no published multiplier. [1] [2] GPT-5.5 predates Fast and keeps the older processing tiers instead: Batch and Flex at about half price with slower or shared capacity, and Priority at $10/$45 — double the input rate, 1.5× the output rate — for reserved capacity. [4]

Long-context re-tiering. A single request whose input exceeds 272,000 tokens re-prices the entire request at the higher band — Sol moves to $8/$30, Luna to $0.40/$1.80 — not just the overage. The price trackers above source this tiering rule to OpenAI's developer pricing documentation. [1] [2] [8]

What a typical call costs (hypothetical example)

The following are hypothetical arithmetic examples, not quoted bills. Assumptions: 2,000 input tokens, 500 output tokens, standard tier, no cache hits. One Sol call costs 2,000 ÷ 1,000,000 × $4.00 + 500 ÷ 1,000,000 × $20.00 = $0.018. The same call runs about $0.010 on Terra, $0.001 on Luna, $0.025 on GPT-5.5, and $0.0075 on GPT-5.1. [1] [2] [3] [4] [5]

Scaled to 100,000 such calls a month, Luna lands near $100 and Sol near $1,800 — an 18× spread on a light, input-heavy workload. In dollar terms the gap grows as output share rises, even though the multiple narrows, from 20× on an all-input call toward roughly 17× when output dominates. Caching narrows it from the other side: with the full 2,000-token context cached, the Sol call's input component drops from $0.008 to $0.0008, bringing the call to about $0.0108. [1] [5] The tier sets the rung in this example — the 18× spread — and caching trims the multiplier stacked on top of it.

Which tier fits which job

For complex software engineering, research analysis, and long agent runs, Sol is the capability-first tier — launch coverage describes it as the strongest of the three, built for precisely these workloads. [6] For everyday code, document analysis, and knowledge-base Q&A, Terra is the balance tier, described at launch as a step below Sol at half the calling cost — a ratio that still holds at $2 versus $4 on input. [6] [3] For classification, tagging, extraction, summaries, and support bots at volume, Luna is the speed-and-cost tier. [2] [6]

GPT-5.1, at $1.25/$10, fits budget-sensitive general work where you control reasoning depth per call. [5] GPT-5.5, at $5/$30, remains OpenAI's agent flagship — the tracker relays OpenAI's claim that it produces about 40% fewer output tokens than GPT-5.4 on Codex-class programming tasks, a vendor claim with no independent retest in the reporting we could access — and its Batch tier returns work within about 24 hours at roughly half price, which suits overnight bulk jobs. [4]

For context rather than verdicts: the third-party Artificial Analysis intelligence index put GPT-5.5 at 60, and Sol's score has drifted past the pages' printed ceiling of “about 60” — it measured 59 at the July launch and 61 in the tracker listings verified October 11, 2026. Read these as snapshot comparisons rather than grades against a fixed scale. [1] [4]

Check these before you rely on the numbers

OpenAI's official pricing page and its developer docs — which ship as Markdown pages under developers.openai.com — are the source of record. The sticker and cache figures above trace to trackers citing that page with an October 11, 2026 verification date, the tier and processing rows carry the earlier dates noted up top, and automated access to the official pages was blocked when this was checked in mid-November 2026. [1] [8] Verify three things on the official page before you sign anything: current sticker rates, whether Sol's promotional $4/$20 survived past its reported November 21, 2026 guarantee window, and Terra's cached-input rate, which no listing we could access states. [1] [3]

One warning about third-party listings: some aggregators surface a Luna price of $0.10/$0.60 — exactly half the sticker — which matches batch-style rates rather than the standard tier. Two independent listings corroborate $0.20/$1.20 as the standard rate, so treat any half-price figure as a data-source quirk to investigate, not a better deal. [2] [9] [10]

Pick the tier by workload shape; it sets the rung. Then manage the five components — output volume, cache hits, reasoning effort, speed premiums, and the 272K re-tiering threshold — which set the multiplier on top.