GPT-6 Astra bills $3.26 per Index task at $10 in / $50 out

Ten dollars in, fifty dollars out. Three dollars and twenty-six cents per Index task. GPT-6 Astra is OpenAI’s new API model, code gpt-6-astra, shipped September third, twenty twenty-six. The sticker is the menu price per million tokens, the text you send and the text the model writes back. Cost per task is what Artificial Analysis actually gets billed to finish one of its tests. OpenAI’s Standard table, read September eleventh, lists ten dollars in and fifty dollars out, per million tokens. Cached input is one dollar per million tokens. Cache writes are twelve dollars and fifty cents per million tokens. GPT-5.6 Sol on the same table is four dollars in and twenty dollars out, per million tokens. Astra’s sticker is two and a half times Sol’s sticker. Claude Fable 5.1, Anthropic’s model at the same sticker, prints the same ten dollars in and fifty dollars out. Artificial Analysis invoices Fable at seven dollars and sixty-three cents per Index task.

You can hand GPT-6 Astra about a million tokens at once. OpenAI’s page says one million fifty thousand tokens of context. OpenAI caps the prompt at nine hundred and twenty-two thousand tokens. OpenAI caps the reply at a hundred and twenty-eight thousand tokens. OpenAI lists GPT-6 Astra at ten dollars per million input tokens and fifty dollars per million output tokens. Artificial Analysis clocks GPT-6 Astra on OpenAI’s API at fifty-four point one tokens a second. The wait before the first answer token is three hundred and thirty-one point five four seconds, including thinking time. The license is proprietary. OpenAI publishes no public weights. Batch and Flex cost half the Standard rates. Fast mode costs twice the Standard rates. Fast is unavailable when the data stays in the EU.

Artificial Analysis scores GPT-6 Astra fifty-three on Intelligence Index version 4.3, at max effort. The Intelligence Index is a single score Artificial Analysis builds by running the same tests on every model itself. Artificial Analysis billed three dollars and twenty-six cents per Intelligence Index task. Claude Fable 5.1 at max with fallback scores the same fifty-three index points. Fable bills seven dollars and sixty-three cents per Index task. GPT-5.6 Sol at max scores forty-seven index points, at one dollar and ninety-nine cents per Index task. OpenAI’s launch post still prints Intelligence Index sixty-one point two for Astra and sixty-five point seven for Fable. OpenAI’s launch table is Index version 4.1.1. Artificial Analysis shipped Index version 4.2 on September fourth. Artificial Analysis’s live Index is version 4.3.

At max effort, Artificial Analysis measures about twenty-seven thousand output tokens per Index task for GPT-6 Astra. Output tokens per task are how many words-worth of tokens the model writes to finish one Index test. More writing raises the invoice at the same sticker. Claude Fable 5.1 at max with fallback writes about seventy-eight thousand output tokens per Index task. Both models score fifty-three on Intelligence Index version 4.3. Both models list ten dollars in and fifty dollars out, per million tokens. The billed unit is tokens. At fifty dollars per million output tokens, less writing is a cheaper invoice. The stickers match. Ten dollars in and fifty dollars out then bills three dollars and twenty-six cents for Astra. Fable bills seven dollars and sixty-three cents per Index task.

Artificial Analysis also prints a blended dollars-per-million-tokens figure. A blended rate is one dollar figure for a mixed bill: seven parts cached prompt, two parts new prompt, one part output. Astra’s blend is seven dollars and seventy cents per million tokens. Fable’s blend is seven dollars and seventeen cents per million tokens. The two blends sit close. The blend skips how many output tokens the model spends on a task. Cost per Index task counts those tokens. Artificial Analysis bills three dollars and twenty-six cents per Index task for Astra. Fable bills seven dollars and sixty-three cents per Index task. OpenAI’s model page steps the price once the prompt is longer than two hundred and seventy-two thousand input tokens. OpenAI bills twice the input and cache rates and one and a half times the output, for the full request. That step is twenty dollars in, two dollars cached, twenty-five dollars for cache writes, and seventy-five dollars out, per million tokens. OpenAI applies the higher rate to the whole request. GPT-5.6 Sol remains cheaper per task at one dollar and ninety-nine cents. Sol scores forty-seven index points.

GPT-6 Astra at max sits at fifty-three index points and three dollars and twenty-six cents per Index task. Claude Fable 5.1 at max with fallback sits at the same fifty-three index points. Fable bills seven dollars and sixty-three cents per Index task. GPT-5.6 Sol at max sits at forty-seven index points and one dollar and ninety-nine cents per Index task. Fable’s Index invoice is more than twice Astra’s Index invoice.

For an Index task, GPT-6 Astra is fifty-three index points at three dollars and twenty-six cents per Index task. Fable is fifty-three index points at seven dollars and sixty-three cents per Index task. The gap is tokens written. If the bill has to stay near Sol, GPT-5.6 Sol is forty-seven index points at one dollar and ninety-nine cents per Index task. Sol lists four dollars in and twenty dollars out, per million tokens. If the prompt will exceed two hundred and seventy-two thousand input tokens, plan twenty dollars in and seventy-five dollars out for the full request. Artificial Analysis scores Astra sixty-two on its coding-agent run in Codex. Fable scores sixty-two in Claude Code. Astra costs seven dollars and nine cents per coding task. Artificial Analysis ran Terminal-Bench version 4.0, a terminal-agent test. Astra scored fifty-nine point one percent on Terminal-Bench. Fable scored fifty-two percent. Sol scored thirty-nine point nine percent.

The ten dollars in and fifty dollars out from the open is the menu. The three dollars and twenty-six cents is the Index invoice. GPT-6 Astra writes about a third of Claude Fable 5.1’s output tokens per Index task.