OpenAI's cyber model has one price and no way to lower it

OpenAI's API pricing page has four rate tables. They are headed Standard, Batch, Flex and Fast, and between them they set out every way the company will bill you for a token.

Search all four for gpt-5.6-cyber and you get nothing.

The model is on the page. It sits in a separate grouped table headed "Cyber models", described as "Our latest Daybreak models", with one set of rates and no alternative to them: $12.50 per million input tokens, $1.25 cached, $15.625 for cache writes, and $75.00 output.

Every other model OpenAI sells has at least two ways to be billed. This one has one.

What OpenAI published#

In OpenAI's own words, GPT-5.6 Cyber is "our most advanced cybersecurity model for authorized vulnerability research and security testing", and "an alias for our most advanced purpose-trained cybersecurity models, for approved defenders conducting advanced, authorized vulnerability research, exploit validation, and security testing."

Those three numbers are worth a second look, because they are exact: 272,000 plus 128,000 is 400,000. The window is not a round marketing figure with some usable fraction inside it. It is the maximum prompt plus the maximum answer, and nothing else.

The three lanes that are missing#

Batch, Flex and Fast are not marketing tiers. They are the three ways OpenAI lets you trade something for a different rate: latency for a discount on Batch and Flex, money for speed on Fast.

The clearest measure of what their absence costs is the model sitting directly above GPT-5.6 Cyber in the same Daybreak table. GPT-5.6 Sol is the general-purpose model in the same program, and it appears in all four tables.

This is not an oversight in the table. The model page carries its own endpoint list, and against v1/batch it says, in as many words, "Not supported". The same list rules out Chat Completions, Assistants, Realtime, fine-tuning and embeddings. GPT-5.6 Cyber runs on the Responses API and nowhere else.

For a team already integrated against Chat Completions, that is a migration before the first billable token.

Why the long-context columns are empty#

OpenAI's pricing tables carry four extra columns for long context, and for most recent models they hold a second, higher set of rates. For gpt-5.6-cyber all four are dashes.

The rule those columns price is quoted on the model page: "Prompts with >272K input tokens are priced at 2x input and 1.5x output for the full request."

Maximum input on this model is 272,000 tokens. The surcharge starts above 272K. A prompt that triggers it is a prompt the model will not accept, so the second rate has nothing to price, which is what the empty columns are saying.

That is a real difference from the rest of the range rather than a technicality. On GPT-6 Astra the same threshold is very much reachable, and crossing it reprices the entire request. Here the ceiling and the threshold are the same number, so the headline rate is the only rate. If you are new to why any of this is billed per million tokens in the first place, tokens are the unit and they are not words.

What it actually costs#

Take a security engagement that runs 2,000,000 input tokens and 200,000 output tokens through a model. Round numbers, and the arithmetic is the point rather than the scenario.

The same tokens cost $40.00, $12.00 or $6.00 depending only on which model and lane they go through.

OpenAI puts a narrower comparison on the model page itself, against its two previous general-purpose flagships:

Those multiples are unusually clean. GPT-5.6 Cyber is exactly 2.5 times GPT-5.5 and exactly 5 times GPT-5.4, on input and on output alike. Because the ratio is identical on both, your input and output mix does not change it. Whatever your workload looks like, the cyber model costs two and a half times what GPT-5.5 would have cost to run it.

The price has not moved in a generation#

The Cyber models table lists three entries. Two of them are priced identically.

The capability moved from GPT-5.5 Cyber to GPT-5.6 Cyber. The rate did not. The only new figure is the cache write rate, and it is not a new charge so much as a newly published one: $15.625 is exactly 1.25 times $12.50, which is the multiplier the documentation already stated.

gpt-5.4-cyber is listed with no price in any column at all. The page gives no reason.

Caching is the one lever you still have#

With Batch and Flex unavailable, the cached input rate is the only published way to pay less. It is $1.25 against $12.50, so a cached token costs a tenth of a fresh one.

The catch is that writing to the cache costs more than not caching: $15.625 against $12.50, an extra $3.125 per million. Each later hit on that prefix saves $11.25. So a prefix that gets reused even once has already paid for its own write, and everything after that is the tenth-rate. If your engagements share a long standing preamble, caching is worth setting up properly here more than on any other OpenAI model, because it is the only discount on offer.

You cannot simply start using it#

The model page is direct about the gate: "This model requires separate approval and provisioning, you can apply to join the Daybreak program."

Two aliases sit alongside it. gpt-daybreak-blue-latest points at gpt-5.6-sol and gpt-daybreak-red-latest points at gpt-5.6-cyber, and OpenAI says both will be repointed at newer models as the program releases them, "with pricing adjusted to match each underlying model". The documentation never defines what blue and red mean, so this piece will not guess either.

The tool list explains the gate better than the naming does. Through the Responses API the model can use web_search, file_search, image_generation, code_interpreter, hosted_shell, apply_patch, skills, computer_use, mcp and tool_search. A model purpose-trained for exploit validation, holding a shell and a computer, is not something you hand out on signup. If you are wiring any model to tools like these, the prompt injection surface is the thing to understand first.

What to do with this#

The single most useful decision here is a routing one, and it is worth more than any prompt tuning you could do inside the expensive model.

Budget GPT-5.6 Cyber as a per-engagement line item rather than a platform default, and route to it only the step that genuinely needs a purpose-trained security model. Everything around that step, the reconnaissance summarising, the report drafting, the triage of what is worth looking at, is general-purpose work that GPT-5.6 Sol does on the Batch lane at $2.00 input and $10.00 output.

That is 6.25 times cheaper on input and 7.5 times cheaper on output than sending the same tokens to the cyber model. No amount of prompt compression inside a $12.50 model recovers a multiple like that. If you are putting a real number on this before you commit, the back of envelope framework for model costs is the right shape of estimate.

What this piece is not claiming#

These are OpenAI's published rates on OpenAI's own pricing and model pages, read on 5 September 2026. Vendor rates change, and this one has a documented history of not changing, which is not a promise.

The absence of Batch, Flex and Fast rows is what the page shows today. OpenAI may add them, and nothing in the documentation says it will not.

The documentation also gives no reason for the absence, and there is no point inventing one. Approval-gated capacity and a 24-hour discount queue are not an obvious fit for each other, but OpenAI has not said that, so it stays a question rather than an explanation.

Finally, the framing throughout is OpenAI's. This is a model the company describes as being for approved defenders doing authorized work, and the gating, the pricing and the tool list are all consistent with that. Nothing above should be read as a description of an offensive product.