GPT-5.6 pricing, captured September 3, 2026
OpenAI prices the GPT-5.6 family per million tokens, in three models that share one billing shape: Sol at $4.00 input / $20.00 output, Terra at $2.00 / $12.00, and Luna at $0.20 / $1.20 per 1M tokens — for requests whose input stays at or below 272,000 tokens. Push the input context past that line and every model moves to a second, higher band: Sol jumps to $8.00 / $30.00, Terra to $4.00 / $18.00, Luna to $0.40 / $1.80. The input rate doubling at the 272K boundary is the single most expensive mistake in GPT-5.6 budgeting, and it is the part most third-party pricing pages get wrong.
This breakdown is a snapshot, not a quote: every number below was captured on September 3, 2026 from OpenAI’s official API pricing page and the Modellix LLM model page, and the two agreed to the cent for all four models in this article. We run Modellix, an API aggregator that also serves these models through its LLM gateway, so we have a commercial interest when our route appears below. This family changed prices twice in six weeks during 2026 — treat any article without a capture date as suspect.
The table shows the standard rates, both bands, for the three current models and the previous generation:
| Model ID | Input ≤ 272K in / out | Input > 272K in / out | Cached read (≤ 272K) | Context window |
|---|---|---|---|---|
openai/gpt-5.6-sol |
$4.00 / $20.00 | $8.00 / $30.00 | $0.40 | 1,050,000 |
openai/gpt-5.6-terra |
$2.00 / $12.00 | $4.00 / $18.00 | $0.20 | 1,050,000 |
openai/gpt-5.6-luna |
$0.20 / $1.20 | $0.40 / $1.80 | $0.02 | 1,050,000 |
openai/gpt-5.5 (previous gen) |
$5.00 / $30.00 | $10.00 / $45.00 | $0.50 | 1,050,000 |
Rates per 1M tokens, standard processing. Sources: OpenAI API pricing page and the Modellix /llm page, both captured September 3, 2026.
The official OpenAI standard rates for the three GPT-5.6 models across both input-context bands, captured from developers.openai.com on September 3, 2026. The three rows are outlined in purple.
OpenAI lists the lower band under “Short context” and the upper band under “Long context”; the Modellix page labels the same split “Input Context ≤ 272K” for the lower tier. Cache write adds a 1.25× premium on uncached input in the new GPT-5.6 caching model: Sol $5.00, Terra $2.50, Luna $0.25 per 1M cached-write tokens at the lower band (GPT-5.5 has no cache-write rate — OpenAI lists “—“). Cache read keeps a 90% discount, at the rates in the table. We are not going to tell you how to game cache boundaries — treat caching as a billing detail to model, not a lever to chase.
The 272,000-token band is not the context window
The most common GPT-5.6 pricing error is conflating the two “272K” facts. The context window of all four models is 1,050,000 tokens — how much a request can hold. The input-context pricing band is a separate billing rule: per-request input token count decides which rate applies, and the boundary sits at 272,000 tokens. The number of pricing bands is independent of the context-window size, as Modellix’s docs note, and the same split exists whether you call OpenAI directly or through an aggregator.
Concretely: a request with 250,000 input tokens on Sol bills input at $4.00/1M; the same request with 300,000 input tokens bills input at $8.00/1M. The output rate moves too ($20 → $30), which is why a “doubled bill” is realistic for long-horizon agent runs that carry a large shared context across turns — the input side alone doubles, and the output premium stacks on top. Terra and Luna follow the same geometry at lower absolute prices.
The Modellix /llm cost calculator in its default state on September 3, 2026: GPT 5.6 Sol, tier “Input Context ≤ 272K”, with estimated cost equal to the listed original price — the page currently shows no discount banner on OpenAI models.
If you search “gpt 5.6 context window”, the number you will find is 1,050,000 tokens — that is the window. The 272K figure that keeps appearing next to prices is the pricing-band boundary, not a window limit. Two different facts, one shared threshold.
Sol vs Terra vs Luna: what each price buys
The three GPT-5.6 models are one family at three capability-and-cost points, all released to general availability on July 9, 2026. Choosing between them is mostly a cost question once you know the quality ladder; the pricing table above is the whole story on rates.
Sol is the flagship: OpenAI’s stated frontier for coding, knowledge work, cybersecurity, and scientific research, and the most expensive of the three at $4/$20 (≤272K). It is the model you reach for when a task is hard enough that a better answer pays for the difference.
Terra is the balanced option at $2/$12 — half Sol’s input price, ten times Luna’s. If you are coming from GPT-5.5, Terra is the natural drop-in to evaluate first: it costs 40% of GPT-5.5’s per-token input rate while OpenAI positions its benchmark performance against the previous flagship.
Luna is the fast, low-cost tier at $0.20/$1.20 — built for high-volume, latency-sensitive work: classification, extraction, routing, first-pass drafting, lightweight agent loops. Its price makes it the default “try it and see” model, and it is the model you can afford to call a lot.
For the full 28-model Modellix price table with every provider’s discount column, our LLM API pricing comparison is the sibling page to this one; on the Anthropic side, Claude API pricing covers the other main frontier vendor.
Why every article quotes different GPT-5.6 prices
If you have been reading GPT-5.6 pricing pages, you have probably seen $5/$30 for Sol, or $1/$6 for Luna. Those numbers were real — on earlier dates. OpenAI cut prices twice within weeks of launch:
- July 30, 2026 — OpenAI reduced Luna by 80% and Terra by 20% (its price-performance announcement and the GPT-5.6 release page both record the change).
- August 21, 2026 — OpenAI dropped GPT-5.6 Sol’s API and credit pricing by more than 20% for the next three months; OpenAI’s pricing page notes Sol’s promotional pricing holds at least through November 21, 2026.
Third-party pages that verified once and stopped now publish stale numbers — which is why this article dates every capture, and why you should re-check rates the day you commit spend. On the aggregator side, the same discipline applies: Modellix’s own list prices for these models moved with OpenAI’s cuts, and a page captured a week apart can differ.
Direct, aggregator, or OpenRouter: what parity means
Because the current OpenAI list price and the Modellix page price match to the cent for all four models in this article, the honest comparison is not “who is cheaper” — on September 3, 2026, nobody is undercutting anyone on these rates. Sol costs $4/$20 whether you call OpenAI directly or through Modellix. (OpenRouter lists the same $0.20/$1.20 for GPT-5.6 Luna, though its per-provider routing can add its own terms.) That parity is a snapshot, not a promise: the discount banners on the Modellix page have changed before, and aggregator margins can appear and disappear without a model release.
So the route decision rests on everything around the unit price:
- OpenAI direct is the right answer if you want one vendor’s billing, org-level verification, and their full model surface including fine-tuning and Fast mode (the 2×-price, up to 2.5×-speed tier that replaced Priority Processing on July 30, 2026).
- Modellix is a distribution layer, not a model owner — it does not operate its own GPT-5.6. What a shared key buys you here is one integration across nine provider labels and three wire protocols (OpenAI Chat Completions and Responses, Anthropic Messages), per-call cost logs you can attribute to each request, and the same 1,050,000-token windows through a single bill. For the wider picture of routing many models through one integration, see our unified AI API overview.
- Modellix is not an official OpenAI channel. Nothing on this page implies a partnership or reseller agreement; it is one API aggregator carrying the same models.
A note for anyone comparing keys while reading: Modellix API keys are pay-as-you-go — you add credit and pay per token used. There is no free GPT-5.6 key tier on either side of this comparison.
GPT-5.6 on the Modellix LLM Gateway
See the models and pricing page and API reference for calling gpt-5.6-sol, terra, and luna over one key.
View DocsBudget math across the 272K line
The doubling is easiest to see on a real agent shape: one run that streams a long codebase context plus a large generated answer. Compare the same workload just below and just above the boundary, input-heavy (roughly 1.2M input tokens total across the run, 60K output), on Sol:
| Run (Sol) | Input band | Input cost | Output cost | Total |
|---|---|---|---|---|
| Context stays ≤ 272K per request | $4.00/1M | 1.2M × $4 = $4.80 | 60K × $20 = $1.20 | $6.00 |
| Any request crosses 272K | $8.00/1M | 1.2M × $8 = $9.60 | 60K × $30 = $1.80 | $11.40 |
Illustrative arithmetic at September 3, 2026 rates; real totals depend on per-request input counts, output length, and cache hits. The point is the ratio: the input side doubles and output rises 50% once any request in the run crosses the band.
That is a 90% total increase from a single threshold crossing — and because one long request often means many, the practical move is to profile input-context lengths per request before choosing a model, not after the invoice arrives. If the same workload on Terra fits the lower band, the numbers are $2.40 + $0.72 = $3.12 versus $4.80 + $1.08 = $5.88 past the line.
Using gpt-5.6-sol from a harness
The gateway route also matters when the model sits inside an agent harness. Modellix’s docs show exactly how to add openai/gpt-5.6-sol as a custom provider in DeepSeek Harness via $DSH_HOME/settings.yaml (use Modellix LLM with DeepSeek Harness):
1 | llm-pi-ai: |
The full string provider/name is what the gateway sends on the wire, so the model ID in your config must match the table above exactly. The dsh-modellix plugin also exposes the live Modellix model catalog in the Harness model selector; earlier 0.1.x builds used openai/gpt-5.6-luna as the fixed model behind the Design-view parameter assistant, until the 0.2.0 rewrite (August 31, 2026) moved media workflows to a chat-first design.
One alias trap for harness users: the Modellix page routes ~openai/gpt-latest to openai/gpt-5.6-sol as of September 3, 2026. The page’s own caution applies — “Targets change as each family evolves. The mappings on this page are current.” Pin the concrete model ID for anything you benchmark or budget against; an alias that silently moves to another family can land you on a different rate band.
Check Live GPT-5.6 Rates on Modellix
Log in to see current per-token pricing and per-call cost logs for Sol, Terra, and Luna on one key.
LoginFAQ
Which GPT-5.6 model is best?
Best depends on the workload, not the model. Sol is the flagship for hard agentic and reasoning tasks; Terra is the balanced default that OpenAI positions against GPT-5.5 performance; Luna is the high-volume, low-latency workhorse at 5% of Sol’s input price. Start from Terra or Luna and escalate to Sol only where its capability measurably pays for itself.
How much does GPT-5.6 Luna cost?
As of September 3, 2026, Luna costs $0.20 per 1M input tokens and $1.20 per 1M output tokens for input contexts up to 272K, and $0.40 / $1.80 above that. Cached reads are $0.02/1M and cache writes $0.25/1M in the lower band.
How much does GPT-5.6 Sol Ultra cost?
There is no separate “Sol Ultra” model ID or per-token rate. Ultra is a capability setting on gpt-5.6-sol that coordinates multiple agents across parallel workstreams; the model still bills at Sol’s standard rates (the API may charge for the extra compute, so check the current request options before assuming).
What is GPT-5.6’s context window?
1,050,000 tokens for all three GPT-5.6 models and for GPT-5.5 — the window size, separate from the 272,000-token input-context pricing band. Crossing 272K of input changes the per-token rate; it does not cap the request.
Is there a free GPT-5.6 API key?
No. OpenAI charges per token for GPT-5.6 API access, and Modellix is pay-as-you-go with no free model tier for these models. “GPT-5.6 free” results usually point at ChatGPT subscriptions, which are a different product from API billing.
How does GPT-5.6 pricing compare with GPT-5.5?
GPT-5.5 costs $5/$30 in the lower band and $10/$45 above 272K — meaning current Sol is 20% cheaper on input than GPT-5.5 was, and Terra is 60% cheaper on input. Migration math depends on your mix, but the previous generation is now the expensive option in most comparisons.
Do OpenAI and Modellix charge the same for GPT-5.6?
On September 3, 2026, yes — to the cent, on all four models in this article. Modellix’s value is the single key, per-call cost logs, and protocol coverage, not a lower unit price on this family on that date. Re-check both pages before committing budget, because aggregator discounts have changed without a model release before.
Why do GPT-5.6 prices keep changing?
OpenAI cut Luna and Terra on July 30, 2026 and Sol by more than 20% on August 21, 2026 (promotional through at least November 21, 2026). Pages that captured once and stopped now publish stale numbers; this article is dated September 3, 2026 for the same reason.
Rates and availability reflect public information as of September 3, 2026 and change frequently; OpenAI cut prices twice in six weeks during 2026. Validate against OpenAI’s live pricing page and the Modellix /llm page before committing spend. Modellix is an API aggregator with a commercial interest in the route comparison above and does not operate its own GPT-5.6 models. Access image and video models, including leading Chinese models, through a single API key at modellix.ai.
