The model id does not choose the effort, the speed tier, or the cache lifetime. On 18 November 2026 OpenAI opened GPT-6.2 Astra and GPT-6.2 Sol in ChatGPT Work, Codex, and the API. The ids are gpt-6.2-astra and gpt-6.2-sol. A request that sets reasoning effort to none is rejected. Ultrafast is service_tier ultrafast. A cheaper overnight run sets background to true and background_options.completion_window to 2h or 8h, and that discount requires service_tier default.
This page is dated 18 November 2026 from the launch post, the two model pages, the pricing page, the reasoning guide, the Ultrafast guide, and the background guide. Yesterday’s note covered Qwen4-Max. The fields that change this request and the bill are reasoning.effort, service_tier, and background_options.
What opened on 18 November
Both models take text and images and return text. Audio and video are unsupported. Chat Completions and Responses are supported. Realtime, Assistants, fine-tuning, image generation, video, and speech routes are not. Batch is supported. The model page lists tools for the Responses API: web_search, file_search, image_generation, code_interpreter, hosted_shell, apply_patch, skills, computer_use, mcp, and tool_search.
Knowledge cutoff is 18 May 2026 on Astra and 1 June 2026 on Sol. The Responses API also lists discounted background completion windows in public beta on these two models.
| Limit | Tokens |
|---|---|
| Context window | 1,050,000 |
| Max input | 922,000 |
| Max output | 128,000 |
OpenAI’s launch post prices short-context Astra at $8 input, $0.40 cached input, and $40 output per million tokens: 20% below GPT-6 Astra on input and output, and 60% below GPT-6 Astra’s $1.00 cached input. Sol keeps GPT-6.1 Sol at $2, $0.10, and $10. Those three numbers leave out the cache-write rows. On Agents’ Last Exam, OpenAI reports Astra 5.2 points above GPT-6 Astra at less than a third of the cost, and 5.5 points ahead of Claude Fable 5.2 at 5% of the cost. Those scores are OpenAI’s. The leaderboard note says Fable 5.2’s result excludes nine tasks that used a fallback model.
The launch post says Astra is the first externally released OpenAI model at the High threshold for AI self-improvement. Both models meet the Critical cybersecurity threshold and the High biological and chemical threshold, and both ship with tighter safeguards on higher-risk cybersecurity requests. Daybreak currently offers GPT-6 Sol on Daybreak Blue, an alpha for GPT-6 Astra, and GPT-6 Cyber on Daybreak Red. The post says OpenAI will work to bring the 6.2 models to Daybreak. In two simulated workplace tests the models were more likely to leave out an obstacle, such as a missing file. OpenAI says instructions added in Codex substantially reduce those omissions.
Standard rate limits on both model pages are below. The Ultrafast guide tells you to check the organization’s own limits for these two models. It publishes default Ultrafast tokens per minute only for GPT-6 Astra and GPT-6.1 Sol. Ultrafast data residency for GPT-6.2 Astra and Sol is the United States only.
| Standard tier | RPM | TPM |
|---|---|---|
| Build | 5,000 | 1,000,000 |
| Launch | 10,000 | 4,000,000 |
| Grow | 15,000 | 40,000,000 |
reasoning.effort, and what 6.2 drops
On Responses, set reasoning.effort. On Chat Completions, set reasoning_effort. The reasoning guide says both GPT-6.2 models reject none. It does not publish a default effort for these two ids. The medium default, and the rejection of minimal, are stated for GPT-6.1 Sol, so put the effort on a 6.2 request yourself. The guide’s list is model-dependent and can include minimal, low, medium, high, xhigh, and max.
reasoning.mode is a separate field. standard is the default. pro asks for more model work. The guide says that work is billed at the selected model’s standard token rates, and that pro uses more tokens. It is not another model id. The guide’s pro example is gpt-6.2-sol with mode pro and effort medium.
To raise effort on a later turn without editing the cached prefix, append a configuration_update item and leave the request-level reasoning.effort as it was. The response still reports that request-level value. On GPT-6 and later, the guide limits this update to standard, single-agent mode.
{
"model": "gpt-6.2-astra",
"service_tier": "default",
"reasoning": { "effort": "low" },
"previous_response_id": "resp_previous",
"input": [
{
"type": "configuration_update",
"reasoning": { "effort": "high" }
},
{
"role": "user",
"content": "Now check the failure modes."
}
]
}
Starting with 6.2, the API also drops a reasoning item when the prefix before it changes (prefix_mismatch), or when safety_identifier differs between the request that stored the item and the request that replays it (end_user_mismatch). Sending the identifier on only one of those two requests counts as a difference. Deleting whole items without editing the items that remain, or appending items after the reasoning item, does not by itself raise prefix_mismatch. The launch post says to keep prompt prefixes unchanged and append new instructions. When an item disappears, read input_transformations on the response.
reasoning.summary set to auto opts in to a summary. Raw reasoning tokens stay hidden. They are still inside output_tokens, and output_tokens_details.reasoning_tokens is that count. max_output_tokens caps reasoning and the visible answer together. If that cap is hit, status is incomplete and incomplete_details.reason is max_output_tokens.
service_tier: Standard, Fast, Ultrafast
The Ultrafast guide sets service_tier to ultrafast on each Responses response.create, over a WebSocket or over HTTP. For an agent that calls tools in quick succession, it recommends keeping one WebSocket and passing previous_response_id. fast and the older name priority select Fast mode, at twice the Standard card for the same context length. The note that the response echoes priority is written for GPT-5.6 and earlier. default selects Standard. With auto, a discounted background window qualifies only when the API selects Standard.
Prices below are US dollars per million tokens for a short-context request, already multiplied. Copy the row for the tier you sent. Copying an Ultrafast row and then multiplying by 3 again counts Ultrafast twice. The Ultrafast multiplier is 3× through 28 February 2027 and 6× from 1 March 2027, Pacific time. Long context is more than 272K input tokens: input and cache rates double, and output is 1.5× the short-context output rate, for the whole request. The published long-context rows already include that change.
| Tier | Astra input / cached / 30m / 12h / output | Sol input / cached / 30m / 12h / output |
|---|---|---|
| Ultrafast through 28 Feb 2027 | $24 / $1.20 / $30 / $60 / $120 | $6 / $0.30 / $7.50 / $15 / $30 |
| Ultrafast from 1 Mar 2027 | $48 / $2.40 / $60 / $120 / $240 | $12 / $0.60 / $15 / $30 / $60 |
| Fast, 2× Standard | $16 / $0.80 / $20 / $40 / $80 | $4 / $0.20 / $5 / $10 / $20 |
| Batch, 0.5× Standard | $4 / $0.20 / $5 / $10 / $20 | $1 / $0.05 / $1.25 / $2.50 / $5 |
Batch is the Batch API at half of Standard. It is not a service_tier value on the Ultrafast page. This HTTP body is the interactive Ultrafast case. It does not set background.
{
"model": "gpt-6.2-astra",
"service_tier": "ultrafast",
"reasoning": { "effort": "low", "summary": "auto" },
"prompt_cache_options": { "ttl": "30m" },
"input": "Draft the patch and the test that should fail first."
}
background_options and the window
Discounted completion windows are in public beta on GPT-6.2 Astra, GPT-6.2 Sol, GPT-6.1 Terra, and GPT-6 Luna. Set background to true. The clock starts when the API accepts the initial request. A follow-up that reuses the window, the continuation id, the model, and the project keeps the discount. 2h and 8h require Standard processing, so set service_tier to default. Leave stream unset or set it to false. Input tokens on those two windows count against the Batch/Background limit at submission.
| completion_window | Model-token price | Rate limit |
|---|---|---|
| soonest, or omit the field | Standard, no discount | Ordinary request limits |
| 2h | 30% below Standard | Batch/Background limit |
| 8h | 40% below Standard | Batch/Background limit |
A Standard request outside those windows cannot reuse a prompt-cache entry that a 2h or 8h background request created, even when the prompt text matches. Applied to the short-context Standard card, 30% below is Astra $5.60 / $0.28 / $7 / $14 / $28 and Sol $1.40 / $0.07 / $1.75 / $3.50 / $7. Forty percent below is Astra $4.80 / $0.24 / $6 / $12 / $24 and Sol $1.20 / $0.06 / $1.50 / $3 / $6, in the order input, cached input, 30-minute write, 12-hour write, output. Those figures are the Standard card times 0.70 and 0.60. The pricing page states the discount and does not print a separate background table. Regional processing and FedRAMP add 10% for models released on or after 5 March 2026, which includes these two ids.
{
"model": "gpt-6.2-sol",
"service_tier": "default",
"background": true,
"background_options": { "completion_window": "8h" },
"reasoning": { "effort": "medium" },
"prompt_cache_options": { "ttl": "12h" },
"input": "Review the overnight document batch."
}
One tier, then the cache write
Omit prompt_cache_options.ttl and the lifetime is 30m. Set 12h when the next turn can be hours away. On this card a 30-minute write is 1.25× the input price and a 12-hour write is 2.5×. That multiplier is the whole write price. A cached read on both 6.2 models is 0.05× input: $0.40 on Astra’s $8, and $0.10 on Sol’s $2. GPT-6 Astra’s cached read is $1.00 on a $10 input price, which is 0.1×. The launch post’s 60% figure is $0.40 compared with that $1.00.
The Standard card below is US dollars per million tokens. A long-context row replaces the short-context row for every token in that request once input is over 272K. gpt-6.2-astra-law is a separate id on the pricing page, and its Standard short and long rows match gpt-6.2-astra. service_tier does not switch the id.
| Model | Context | Input | Cached input | 30m write | 12h write | Output |
|---|---|---|---|---|---|---|
| gpt-6.2-astra | Short, ≤272K | $8 | $0.40 | $10 | $20 | $40 |
| gpt-6.2-astra | Long, >272K | $16 | $0.80 | $20 | $40 | $60 |
| gpt-6.2-sol | Short, ≤272K | $2 | $0.10 | $2.50 | $5 | $10 |
| gpt-6.2-sol | Long, >272K | $4 | $0.20 | $5 | $10 | $15 |
| gpt-6-astra | Short, ≤272K | $10 | $1.00 | $12.50 | $25 | $50 |
Ordinary input is input_tokens minus cached_tokens minus cache_write_tokens. cache_write_tokens equals ttl_30m_tokens plus ttl_12h_tokens. reasoning_tokens is already inside output_tokens. On the sample below, short-context Standard Astra bills 3,000 ordinary tokens at $8, 12,000 cached tokens at $0.40, 5,000 twelve-hour write tokens at $20, and 1,800 output tokens at $40, per million. That is $0.024 + $0.0048 + $0.100 + $0.072 = $0.2008. The same counts at Ultrafast through 28 February 2027 are 3× that, $0.6024. The same counts on an 8h background window are 0.6× the Standard total, $0.12048.
{
"usage": {
"input_tokens": 20000,
"input_tokens_details": {
"cached_tokens": 12000,
"cache_write_tokens": 5000,
"cache_write_tokens_details": {
"ttl_30m_tokens": 0,
"ttl_12h_tokens": 5000
}
},
"output_tokens": 1800,
"output_tokens_details": {
"reasoning_tokens": 900
}
}
}
Adding cache_write_tokens on top of ttl_12h_tokens adds another $0.10 to this sample. Adding reasoning_tokens on top of output_tokens adds another $0.036. A request whose input_tokens are over 272K uses the long-context row instead of this short-context arithmetic.
JSON to verify before launch
modelisgpt-6.2-astraorgpt-6.2-sol, orgpt-6.2-astra-lawwhen that configuration is the one you want.reasoning.effortorreasoning_effortis present, and the value is notnone.service_tiermatches the card you will multiply. A2hor8hbody usesdefault,backgroundtrue, andstreamabsent or false.prompt_cache_options.ttlis30mor12h. The prefix matches the previous turn, andsafety_identifieris the same on both turns or absent from both.- The bill uses one tier,
cached_tokensonce, and the two ttl fields once.
Check the JSON locally with the JSON validator, JSONPath ($.usage.input_tokens_details.cache_write_tokens_details.ttl_12h_tokens), and JSON Diff against the previous request. Nothing is uploaded.
FAQ
What happens if reasoning effort is none?
The API rejects the request. The reasoning guide names that rejection for both GPT-6.2 models and does not publish a default effort for them, so set the effort on the request. The medium default belongs to the sentence about GPT-6.1 Sol.
Can Ultrafast and an 8h window go on one request?
A discounted window requires service_tier default. auto qualifies only when the API selects Standard. ultrafast is a different value: 3× Standard through 28 February 2027 and 6× from 1 March 2027, Pacific time.
Should the bill add cache_write_tokens to the ttl fields?
cache_write_tokens is already ttl_30m_tokens plus ttl_12h_tokens. Multiply those two counts by the 30-minute and 12-hour write prices. reasoning_tokens is already inside output_tokens.
Does only the portion above 272K move to the long-context price?
The pricing page marks short context as ≤272K input tokens and long context as >272K. Over that line, input and cache rates double and output is 1.5× the short-context output rate for the full request.
Does editing an earlier tool result keep the reasoning item?
On 6.2, a changed prefix removes the reasoning item with reason prefix_mismatch. Append a configuration_update to change effort, and append new instructions, instead of rewriting the prefix. Keep safety_identifier identical on both requests, or omit it from both.
Is Sol the same price as GPT-6.1 Sol?
The launch post and the pricing page keep Sol at $2 input, $0.10 cached input, and $10 output per million on the short-context Standard card, with a $2.50 30-minute write and a $5 12-hour write. Those write rows match gpt-6.1-sol. Astra does not keep the GPT-6 Astra card.
Summary
The 18 November 2026 ids are gpt-6.2-astra and gpt-6.2-sol. Set reasoning.effort, because none is rejected and the guide does not name a 6.2 default. Set service_tier to ultrafast, fast, or default, and keep a 2h or 8h background window on default. Bill one tier, then cached_tokens and the two ttl write fields.