GPT-6 Astra: staged access and a 272K price cliff

OpenAI announced GPT-6 Astra on September 3, 2026. benchr rechecked the launch announcement and live API model documentation on September 5, preserving the staged-access language and exact benchmark variants.

Availability is a rollout, not a binary flag

The launch began with a limited set of organizations. OpenAI said ChatGPT Plus, Pro, Business, and Enterprise access, the OpenAI API, Azure, and AWS Bedrock would expand over the coming days; Enterprise workspace access starts off by default. Confirm the model in the target account and region before promising a migration date. The documented endpoint is evidence of a product contract, not evidence that every tenant was enabled on September 3.

A 272K prompt reprices the whole request

Below or at 272K input tokens, Standard costs $10 input, $1 cached input, $12.50 cache writes, and $50 output per million. Above 272K, the entire request—not only the excess—uses 2x input and cache rates and 1.5x output: $20, $2, $25, and $75. Split large corpora, retrieve only relevant passages, and measure the exact prompt envelope before accepting a 1.05M-window design.

The launch table is provider evidence, not a benchr score

OpenAI reports 96.0 on GPQA Diamond, 95.0 on ARC-AGI-2, 57.9 on Terminal-Bench 4.0, 74.1 on DeepSWE v1.1, and 72.6 on an offline partial-score OSWorld 2.0 set. The announcement says GPT evaluations use the maximum available effort and may differ from production ChatGPT because harnesses, prompts, and tools differ. Keep each variant label intact and replay the work that matters to your team; benchr has not reproduced these numbers.

The tool surface is broad; the media contract is not

The model page lists streaming, function calling, structured outputs, web and file search, image generation as a tool, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search in the Responses API. The model itself accepts text and image input and returns text; audio and video input are not supported, and fine-tuning is not offered. Do not infer a native media modality from the presence of an image-generation tool.

Migration changes more than the model string

OpenAI's model guide says Astra does not support the none or minimal reasoning levels, custom temperature or top_p, or log probabilities. Tool calling should move to the Responses API, and Fast mode is unavailable with EU data residency. Replace an old identifier only after checking these request parameters, tool behavior, cache settings, safeguards, and regional processing requirements.