OpenAI API Pricing 2026: Cost Per Million Tokens

Quick Answer

OpenAI API pricing 2026: GPT-6 Astra $10 input / $50 output per million tokens on standard short context; $20/$75 long-context past 272K input tokens; cached input $1; batch and Flex $5/$25; Fast mode $20/$100; cache writes $12.50. GPT-5.6 Luna at $0.20/$1.20 is the cheapest flagship-tier model; GPT-5.6 Sol at $4/$20 is the cheapest mid-tier reasoning model with promotional pricing guaranteed through November 21, 2026.

Last verified: Sep 16, 2026.

At a glance

• GPT-6 Astra standard: $10 input / $50 output per 1M tokens (≤272K input)
• GPT-6 Astra long context: $20 input / $75 output past 272K input (whole request)
• Cached input: $1/MTok on standard; $0.50 on Batch/Flex
• Cache writes: $12.50 (5-minute TTL) / $20 (1-hour TTL, not published for Astra)
• Batch and Flex: $5 input / $25 output (50% off)
• Fast mode: $20 input / $100 output (2x); unavailable under EU data residency
• GPT-5.6 Sol promo: $4/$20, guaranteed through Nov 21, 2026
• GPT-5.6 Luna: $0.20/$1.20 (cheapest flagship-tier)

Four token rates, not one

OpenAI bills four token categories separately for every GPT-6 Astra request, and ignoring any of them can swing a budget by 5-10x.

The four billed categories are uncached input, cached input, cache writes, and output. On the standard tier at short context, those map to $10, $1, $12.50, and $50 per million tokens respectively. OpenAI documents these as separate line items on the API pricing page because most teams do not realise cache writes happen any time a prompt prefix is longer than 1,024 tokens — which is the trigger for OpenAI to cache the prefix (cloudzero.com, September 2026).

For a 5,000-token system prompt plus a 20,000-token user prompt with a 2,000-token answer, a single request incurs 25,000 uncached input tokens at $0.25, a cache write for the 5,000-token prefix at $0.0625, and 2,000 output tokens at $0.10. Total: $0.4125 per request. Multiply by 1,000 turns/day and the bill is $412.50/day, of which $62.50 is purely the cache write — money teams routinely forget to model (cloudzero.com, September 2026).

The 272K long-context cliff

GPT-6 Astra prices the entire request at the long-context rate when input tokens exceed 272,000 — not just the portion above the threshold.

Standard input rises from $10 to $20 per million and output rises from $50 to $75 per million for the whole request once the threshold is crossed. The 272K figure is a pricing boundary, not a context-window boundary — GPT-6 Astra's actual context window is 1,050,000 tokens, with up to 922,000 input and 128,000 output. A request at 273,000 input tokens costs roughly twice what it would at 271,000, even though only 2,000 tokens moved (laozhang.ai, September 2026).

This is the single biggest cost cliff in the OpenAI catalogue and the reason most production systems are designed around the 272K threshold. For long-document RAG workflows, it is cheaper to chunk and run two separate requests than to combine the full document into one that trips the threshold. The cliff applies identically on Azure AI Foundry, AWS Bedrock (when Astra ships there), and OpenRouter (laozhang.ai, September 2026).

Batch, Flex, and Fast: three different discounts

OpenAI runs three non-standard pricing modes that each cut cost in a different way.

Batch returns results within 24 hours at half the standard rate: GPT-6 Astra at $5 input / $25 output per million tokens. Flex returns within 2 hours at the same $5/$25 rate but with looser SLAs and the option to retry-on-failure at no charge. Fast mode cuts latency to single-digit seconds at $20 input / $100 output per million tokens, doubling the rate (OpenAI pricing page, September 2026).

For workloads that can wait, Batch is the cheapest OpenAI route to a frontier model. For workloads that cannot wait, Fast mode is expensive but is the only path that beats Flex's 2-hour SLA. The 50% Batch and Flex discount applies across the catalogue — including the GPT-5.6 family — making Batch the natural fit for nightly RAG ingestion, bulk classification, and offline evaluation (tokencost.app, September 2026).

The full GPT-5.6 family

OpenAI kept the GPT-5.6 family alive at promotional rates through at least November 21, 2026 to ease the transition from GPT-6 Astra.

GPT-5.6 Sol is OpenAI's mid-tier reasoning flagship at $4 input / $20 output per million tokens promotional. GPT-5.6 Terra is the cost-optimised model at $2/$12. GPT-5.6 Luna is the cheapest flagship-tier model at $0.20/$1.20. GPT-5.6 Cyber is the safety-classified model at $12.50 input per million. All four run at the same 50% Batch discount, the same 2x Fast rate, and the same 1.25x cache-write pricing as GPT-6 Astra (cloudzero.com, September 2026).

GPT-5.3 Codex at $1.75 per million input tokens is the cheapest coding-optimised model — useful for code-generation workloads that do not need reasoning. The 2.5x premium for Astra over Sol is justified only when the model materially reduces retries, completion time, or failed runs. For routine workloads that Sol already handles reliably, the Astra premium does not pay back (linkmodel.ai, September 2026).

What enterprise buyers should do next

Three actions for organisations evaluating GPT-6 Astra in 2026.

1. Model cost per successful task, not per token. A 2.5x token-rate premium (Astra vs Sol) is irrelevant if Astra reduces retries by 50%. Build a small evaluation harness and measure cost per successfully completed task before committing to a model.
2. Audit cache strategy. 5-10x cost differences between workloads often come down to whether the system prompt is cache-eligible. Keep system prompts above 1,024 tokens and as stable across requests as possible to amortise the $12.50/MTok cache write.
3. Set up a guardrail at the 272K threshold. Either chunk requests under the threshold or accept the long-context rate. Most production systems need explicit code to enforce this.

What to watch next

Three near-term datapoints. First, the November 21, 2026 deadline for GPT-5.6 Sol's promotional pricing — OpenAI has not yet announced what happens to the $4/$20 rate on that date. Second, GPT-6 Astra availability on AWS Bedrock and GCP Vertex AI — both are listed as 'coming days' in the September 2026 launch announcement; once shipped, regional pricing uplifts may apply. Third, the OpenAI pricing for the Daybreak enterprise coalition members, which is gated-access and not published — competitive intelligence on those rates will come from procurement reports over Q4 2026.