GPT-6 Astra pricing: what it costs, and the threshold that doubles it

What OpenAI released, and when

OpenAI released GPT-6 Astra on 3 September 2026. The model id is gpt-6-astra. It carries a 1,050,000 token context window, accepts up to 922,000 input tokens in a single request and returns up to 128,000 output tokens.

Access did not open to everyone at once. Enterprises in OpenAI's Trusted Access Programme received it first, with API access and the Plus, Pro, Business and Enterprise plans following over the days after launch. If you are reading a benchmark comparison written in the first week, check which tier the author was actually testing on, because the rollout order means early numbers came from a narrow population.

The capability claims are large and OpenAI states them plainly. On SRE-Bench, an internal evaluation of site reliability tasks, Astra solved 88.0% of tasks on the first attempt and 99.2% within four, against 55.9% and 68.7% for GPT-5.6 Sol. It is also the first model OpenAI has classified at the Critical level for cybersecurity capability under its Preparedness Framework, a point we return to below because it changes the procurement conversation rather than the engineering one.

None of that is what this article is about. Capability claims are argued over for months and settle slowly. The price list is published, unambiguous, and takes effect on the first invoice.

What it costs against the model it replaces

The standard rates, taken from OpenAI's published pricing page, with GPT-5.6 Sol alongside for scale. All figures are per million tokens.

The consistency is the striking part. Every rate is exactly two and a half times its predecessor, including the cached and long-context rates. This is not a repricing of particular workloads, it is a uniform step up, which means no usage pattern escapes it by being shaped differently. If your bill is $4,000 a month on Sol and you move the same traffic to Astra unchanged, budget $10,000.

The comparison is against a temporary number

GPT-5.6 Sol is on promotional pricing. OpenAI's own documentation states it is available at least through 21 November 2026, wording that promises a floor on the duration and nothing about what follows. The $4 and $20 figures are a discount on Sol's earlier rates, described as a 20% reduction on input and 33% on output.

So the 2.5x gap is real today and may not be the gap you plan against. If Sol reverts to roughly $5 and $30 after November, the multiple narrows to about 2x and 1.7x. A migration business case built on today's spread should say which number it assumed, because the answer changes in under three months.

What that looks like on a monthly bill

Abstract multiples are easy to wave through. Take a support agent handling 50,000 conversations a month, each sending 8,000 input tokens of context and returning 800 output tokens. That is 400 million input tokens and 40 million output.

The same workload spans $2,400 to $12,000 depending on two choices, and only one of those choices is the model. Astra on batch costs less than Astra on standard by more than the entire Sol bill, which is the sort of thing that decides whether a migration is affordable at all.

The 272,000 token cliff

This is the part that catches teams out, and it is not prominent in the launch coverage.

Requests that exceed 272,000 input tokens are charged at the long-context rates: double the input price and 1.5 times the output price. Crucially, the higher rate applies to the entire request, not to the tokens above the threshold. There is no marginal band. One token over and the whole call is re-rated.

Two requests, five thousand tokens apart:

A 1.9% increase in input length produces a 99% increase in cost. Nothing about the request changed except which side of a line it landed on.

Why this matters more than it looks

The advertised context window is 1,050,000 tokens and the maximum input is 922,000. The threshold sits at 272,000. That means roughly seventy per cent of the usable context window is on the expensive side of the cliff, and the headline price only describes the first quarter of it.

Retrieval pipelines are the obvious exposure. A system that assembles context from a vector store will produce request sizes that vary with what it retrieves, and if the median lands near 250,000 tokens then some ordinary fraction of traffic crosses the line and doubles. The bill does not scale smoothly with usage, it steps, and the step is invisible in an average.

What to do about it

Instrument input token counts per request and look at the distribution rather than the mean. If a meaningful share sits between 250,000 and 300,000, a context budget that caps assembly just under the threshold is worth more than most prompt optimisation. Trimming a request from 280,000 to 270,000 tokens does not save 3.6% of the input cost, it saves half of it.

The second defence is to stop treating the context window as a resource to fill. A window of 1,050,000 tokens invites an architecture that retrieves generously and lets the model sort it out, and on the old pricing that was merely wasteful. Now it is a cost cliff with a hard edge at roughly a quarter of the window. Retrieval that returns the twenty best passages instead of the two hundred available is cheaper on both sides of the threshold and usually more accurate as well.

It is worth being clear about what this rule is not. It is not a penalty on long conversations as such, because the threshold counts input tokens in a single request. A hundred-turn conversation that stays under 272,000 tokens per call never triggers it. What triggers it is a single large assembly, which in practice means retrieval, document processing and codebase-wide analysis rather than chat.

Caching, which most estates underuse

Cached input on Astra is $1 per million tokens against $10 uncached. That is a ninety per cent reduction, the largest discount in the price list, and it applies to exactly what conversational and agentic systems do most: resending the same context on every turn.

The mechanism has a cost on the way in. Cache writes are charged at 1.25 times the uncached input rate, so writing a prefix costs $12.50 per million rather than $10. Whether it pays is simple arithmetic. A prefix costs 1.25 units to write and 0.1 units to read, against 1 unit to send uncached every time. Two reads and you are ahead. Three and it is not close.

Figures are per million tokens of prefix. An agent with a 20,000 token system prompt and tool schema, running a ten-turn conversation, reaches the third row inside a single session.

The practical obstacle is rarely the economics. It is that caching rewards a stable prefix, and prompts assembled by string concatenation often vary in ways nobody intended: a timestamp near the top, a user id interpolated before the instructions, a tool list whose order is not deterministic. Any of those invalidate the cache on every call while looking perfectly correct in a code review. If cached token counts are lower than you expect, something in the prefix is moving.

The two levers that move the rate

OpenAI publishes three service modes and the spread between them is wider than the gap between model generations.

Batch and Fast are a four-fold spread on identical tokens. That is a larger factor than the 2.5x between Astra and Sol, which means the mode you send work in can matter more to the invoice than the model you chose.

The distinction is latency, not capability. Anything that does not need an answer while somebody waits is a candidate for Batch: overnight enrichment, classification backfills, evaluation runs, document processing queues. Teams routinely leave this on the table because the synchronous path is what they built first and it keeps working.

Fast mode is the mirror image and deserves the same scrutiny. At $20 and $100 it is the most expensive way to buy Astra, and it is worth confirming that the workloads using it genuinely have a human waiting rather than having inherited the setting from a prototype.

Two smaller adjustments

Cached input is $1 per million against $10 uncached, a ninety per cent reduction on repeated context. Cache writes are charged at 1.25 times the uncached input rate, so caching pays back once a cached prefix is read roughly twice. For an agent that resends a large system prompt on every turn, that happens immediately.

Regional processing endpoints, used where data residency is required, carry a 10% uplift for models released on or after 5 March 2026. Astra is inside that window, so a residency requirement is a line item rather than a configuration choice.

Critical cyber capability, and why procurement will ask

OpenAI classified Astra at the Critical level for cybersecurity capability under its Preparedness Framework, the first model to reach that tier. In OpenAI's description the model can find previously unknown security flaws and develop ways to exploit them against well-defended systems, and during evaluation it discovered and used two previously unknown vulnerabilities.

OpenAI's response was to restrict rather than delay: additional safeguards against harmful cyber actions, and a staged rollout that put enterprises in a trusted access programme ahead of general availability.

For a buyer this is a procurement fact more than an engineering one. A model carrying a vendor's own highest risk classification will attract questions in a security review that GPT-5.6 Sol did not, and the honest answer to most of them is that the classification describes what the model can do rather than what your deployment permits. Expect to document which safeguards apply to your access path, and expect the review to take longer than the last one did.

It also sharpens a question worth asking anyway. If a workload does not need frontier capability, the classification is a reason to keep it on the cheaper model rather than a reason to wait.

What a buyer should actually do

Four things, in the order they pay back.

Measure the distribution, not the average

Pull input token counts per request for a representative week and plot them. The single most expensive thing about Astra's pricing is a threshold, and a threshold is invisible in a mean. If traffic clusters near 272,000 tokens, capping context assembly below it is the highest-return change available and it requires no model decision at all.

Sort work by whether somebody is waiting

Batch is half price and Fast is double. Most estates have work in the standard or fast path that nobody is waiting on, usually because that is how it was first built. Moving it is a configuration change with a 2x to 4x effect on those tokens.

Do not migrate wholesale

Astra is 2.5 times Sol on every rate. The SRE-Bench gap OpenAI publishes is large, but it describes agentic reliability work, not summarisation or classification or extraction. Route by task. A retrieval pipeline that was accurate enough on Sol does not become more accurate on Astra, it becomes 2.5 times more expensive.

Put a date on the Sol assumption

Sol's promotional pricing is documented as lasting at least through 21 November 2026. Whatever you decide, write down which Sol price the decision assumed and revisit it in November, because a comparison against a promotional rate expires with the promotion.

Questions people ask

How much does GPT-6 Astra cost per million tokens?

$10 for input and $50 for output at standard rates, with cached input at $1. Requests over 272,000 input tokens are charged at $20 input and $75 output. Batch processing halves every rate and Fast mode doubles it.

How much more expensive is GPT-6 Astra than GPT-5.6 Sol?

Exactly 2.5 times, on every published rate. Sol is $4 input and $20 output against Astra at $10 and $50, and the same multiple holds for cached input and for the long-context rates. Note that Sol is on promotional pricing stated as running at least through 21 November 2026, so the gap is measured against a temporary number.

What happens at 272,000 tokens?

A request that exceeds 272,000 input tokens is re-rated at double the input price and 1.5 times the output price, and the higher rate applies to the whole request rather than only the tokens above the threshold. A 275,000 token request costs roughly twice a 270,000 token one. Because the context window is 1,050,000 tokens, most of the window sits above this line.

Is GPT-6 Astra worth 2.5 times the price?

It depends entirely on the task. OpenAI reports Astra solving 88.0% of SRE-Bench tasks on the first attempt against 55.9% for Sol, which is a large gap on agentic reliability work. That benchmark says nothing about summarisation, classification or extraction, where the cheaper model may already be accurate enough. Route by task rather than migrating an entire estate.

Why is GPT-6 Astra restricted rather than generally available?

OpenAI classified it at the Critical level for cybersecurity capability under its Preparedness Framework, the first model to reach that tier, and responded with additional safeguards and a staged rollout beginning with enterprises in its Trusted Access Programme. API and paid plan access followed in the days after the 3 September launch.

Does data residency change the price?

Yes. Regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026, which includes GPT-6 Astra. If a residency requirement applies to your deployment, treat it as a line item rather than a configuration detail.

Sources

Every price, limit and date above was checked against these pages on the day shown. Where a figure has since moved, the vendor’s own page is the authority and this one is a snapshot.

1. 1Pricing, OpenAI API documentation · OpenAICarries the standard, batch and fast rates for gpt-6-astra and gpt-5.6-sol, including the long-context rates.Checked 4 September 2026
2. 2GPT-6 Astra model reference · OpenAIModel id, the 1,050,000 token context window, 922,000 maximum input and 128,000 maximum output.Checked 4 September 2026
3. 3GPT-5.6 Sol model reference · OpenAISol rates, the 272K threshold rule, cache write multiplier, and the promotional pricing date of 21 November 2026.Checked 4 September 2026
4. 4GPT-6 Astra: a new generation of intelligence · OpenAIRelease on 3 September 2026, SRE-Bench figures and the rollout order.Checked 4 September 2026
5. 5Safety overview: GPT-6 Astra · OpenAIThe Critical cybersecurity classification under the Preparedness Framework and the safeguards applied.Checked 4 September 2026