OpenAI's API rates in November 2026 run from $0.05 to $10 per million input tokens and from $0.40 to $50 per million output tokens across the 23 models in the November 3 snapshot of the LiteLLM price list [1]. Cached input bills at half to a twentieth of the standard input rate, depending on the row [1]. Batched requests bill at exactly half standard on the 22 rows the snapshot prices for batch [1]. Microsoft's Azure channel adds a third billing mode, reserved throughput bought as Provisioned units on monthly or annual terms [2].
The November 3, 2026 price ladder
The snapshot's own sort runs from cheapest typical request to most expensive, GPT-5 nano first and GPT-6 Astra Law last, and every figure is a standard rate in US dollars per million tokens, meaning non-cached and non-batch API calls [1].
| Model | Input, USD per million tokens | Output, USD per million tokens | Typical request, USD (1,000 input, 500 output) |
|---|---|---|---|
| GPT-5 nano | 0.05 | 0.40 | 0.00025 |
| GPT-6 Luna | 0.10 | 0.50 | 0.00035 |
| GPT-4o mini | 0.15 | 0.60 | 0.00045 |
| GPT-5.6 Luna | 0.20 | 1.20 | 0.0008 |
| GPT-5.4 nano | 0.20 | 1.25 | 0.000825 |
| GPT-4.1 mini | 0.40 | 1.60 | 0.0012 |
| GPT-5 mini | 0.25 | 2.00 | 0.00125 |
| GPT-5.4 mini | 0.75 | 4.50 | 0.003 |
| GPT-6.1 Terra | 1.00 | 5.00 | 0.0035 |
| GPT-4.1 | 2.00 | 8.00 | 0.006 |
| o3 | 2.00 | 8.00 | 0.006 |
| GPT-5 | 1.25 | 10.00 | 0.00625 |
| GPT-5.1 | 1.25 | 10.00 | 0.00625 |
| GPT-6 Sol | 2.00 | 10.00 | 0.007 |
| GPT-6.1 Sol | 2.00 | 10.00 | 0.007 |
| GPT-4o | 2.50 | 10.00 | 0.0075 |
| GPT-5.6 Terra | 2.00 | 12.00 | 0.008 |
| GPT-5.2 | 1.75 | 14.00 | 0.00875 |
| GPT-5.4 | 2.50 | 15.00 | 0.01 |
| GPT-5.6 | 4.00 | 20.00 | 0.014 |
| GPT-5.5 | 5.00 | 30.00 | 0.02 |
| GPT-6 Astra | 10.00 | 50.00 | 0.035 |
| GPT-6 Astra Law | 10.00 | 50.00 | 0.035 |
GPT-5 nano holds the floor at $0.05 input and $0.40 output per million tokens, and the site's typical request of 1,000 input and 500 output tokens costs $0.00025 on it [1]. GPT-6 Astra and GPT-6 Astra Law hold the snapshot's ceiling at $10 and $50, and the same request costs $0.035 there, 140 times the nano figure [1]. The middle of the ladder is a band from $1 to $2.50 input that carries 11 of the table's 23 rows, GPT-6.1 Terra at $1 and $5, GPT-5 and GPT-5.1 at $1.25 and $10, GPT-5.2 at $1.75 and $14, and o3 and GPT-4.1 at $2 and $8 [1].
Input and output are billed separately
Every row prices input and output as two meters, and output is always the steeper one, between four and eight times the input rate across the table [1]. The typical-request column prices a fixed 1,000 input and 500 output tokens so rows compare on one number, and it is the basis for the table's sort order [1]. A real workload moves the bill toward whichever meter dominates it, since a long-output job pays closer to the output rate and a long-prompt job closer to the input rate. The snapshot's footer adds its own boundary, that providers bill by their own token counts and that readers should check the provider's pricing page before relying on a number [1].
Cached input bills at half to a twentieth of standard
All 23 rows carry a cached-input price, and on every row it sits below the standard input rate, from half on the GPT-4o rows to a twentieth on the newest ones [1]. GPT-6.1 Terra prices cached input at $0.05 against a standard $1, GPT-6.1 Sol at $0.10 against $2, GPT-5.1 at $0.125 against $1.25, and GPT-6 Astra at $1 against $10 [1]. The older rows discount less, GPT-4o mini at $0.075 against $0.15 and o3 at $0.50 against $2 [1]. The snapshot also carries a cache-write column, and where the column has a price it runs at 1.25 times the standard input rate, $1.25 per million on GPT-6.1 Terra and $12.50 on GPT-6 Astra [1]. As a general mechanism, cache pricing rewards workloads whose requests repeat the same prompt block, and the cache-write charge covers putting that block into the cache.
Batch processing halves the rate for a day of patience
Every row with batch prices bills exactly half standard on both meters, 22 of the 23 rows, GPT-5.1 at $0.625 input and $5 output against $1.25 and $10, GPT-6 Astra at $5 and $25 against $10 and $50 [1]. The exception is GPT-6 Astra Law, whose batch columns are blank in the snapshot [1]. The condition is patience, and Microsoft's own description of the Batch API on Azure states that it returns completions within 24 hours for a 50% discount on Global Standard Pricing [2], the same half-price structure the snapshot's columns show. A workload that can wait a day bills at half rate without changing models.
Reserved throughput is the third billing mode
Steady, predictable load has a third option on the Azure channel. Microsoft prices the same OpenAI model IDs under a Standard on-demand mode billed per token, under Provisioned capacity in PTUs where monthly and annual reservations reduce overall spend, and under the Batch API [2]. Microsoft states that these models are billed through the Azure subscription, covered by Azure service-level agreements and supported by Microsoft [3]. Azure prices those models on its own page, so the rates above describe the OpenAI direct API meter, and a team comparing channels needs Azure's figures for the Azure meter.
A month at scale, as a hypothetical example
Hypothetical example, built only from the snapshot's rates. A team sends 1,000,000 requests in a month through GPT-6.1 Terra, each one the typical request of 1,000 input and 500 output tokens. The month carries 1,000 million input tokens and 500 million output tokens, which bills to $1,000 of input at $1 per million and $2,500 of output at $5 per million, $3,500 in total, or $0.0035 per request, the figure the snapshot's typical-request column shows for that row [1]. Run through batch pricing, the same month bills $1,750, because batch halves both meters [1]. If the same 1,000-token prompt repeats with every request and bills at the snapshot's cached rate of $0.05 per million, input costs $50 for the month against $1,000 at standard, output stays at $2,500, and the month totals $2,550 before any cache-write charge, which the snapshot prices at 1.25 times the standard input rate [1]. The same arithmetic reruns on any real volume, requests multiplied by input and output tokens at each rate.
Where the two snapshots disagree
Every rate figure above comes from tracker snapshots compiled from the LiteLLM price list, and the dates attached to those snapshots are part of the data [1]. OpenAI's own pricing pages returned HTTP 403 on every fetch made for this article, which is why the snapshots are the source of record for the rates here. The November 3 page states its upstream as OpenAI's pricing documentation via the LiteLLM price list [1]. A second tracker, AI Cost Check, shows its OpenAI rows last updated September 22, 2026 [4], and prices the GPT-5.6 family differently, GPT-5.6 Sol at $5.00 input and $30.00 output, GPT-5.6 Terra at $2.50 and $15.00, and GPT-5.6 Luna at $1.00 and $6.00 [4]. The November 3 table lists GPT-5.6 Luna at $0.20 and $1.20, GPT-5.6 Terra at $2 and $12, and a $4 and $20 row named simply GPT-5.6, with no separate Sol row [1]. Microsoft's own catalog currently lists gpt-5.6-sol, gpt-5.6-terra and gpt-5.6-luna among the Foundry models it sells, so the family is real on both sides of the comparison and the trackers disagree about what it costs [3].
Where the two trackers overlap on other rows they agree, GPT-6 Astra at $10 and $50 on both, GPT-5.2 at $1.75 and $14 on both, and GPT-5 nano as the cheapest input on both [1] [4] [5]. Their censuses differ, 23 models in the November table against 36 on AI Cost Check's OpenAI page [1] [4]. A census reflects a tracker's own inclusion list, and neither census by itself establishes OpenAI's current lineup. GPT-5.3 Codex, for example, sits outside the November table while AI Cost Check prices it at $1.75 input and $14 output [4] and Microsoft lists gpt-5.3-codex among its current models [3]. All rates here are API rates, and ChatGPT's consumer subscriptions are a separate product outside this table.
The budget takeaway
Patience buys exactly half on any batch-capable row, a repeated prompt buys input at half to a twentieth of standard, and the model choice then sets the remaining spread from $0.05 to $10 per million input tokens as of November 3, 2026. The GPT-5.6 rows are the least settled part of that range, and the two trackers' figures for that family still disagree, so any budget built on those rows carries both sets of numbers.