GPT-6 Astra at $10/$50, and a correction: GPT-5.6 Sol has cost $4/$20 since August 21

OpenAI released GPT-6 Astra on September 3, 2026 at $10.00 in / $50.00 out per 1M tokens ($1.00 cached input), labelled promotional pricing with no end date — the same headline rate as Claude Fable 5.1. It is a new row in our cost calculator. The correction is the older news: on August 21 OpenAI cut GPT-5.6 Sol from $5 / $30 to $4 / $20 "at least through November 21, 2026", and our August 28 sweep did not catch it. We carried the old price for 17 days.

Verdict: OpenAI's new top model lands at $10/$50, and the flagship under it was quietly cut to $4/$20 two weeks earlier — a cut we missed

The take

Astra first, from OpenAI's own rate card (developers.openai.com, read September 7): input $10 / 1M, cached input $1 / 1M, cache writes $12.50 / 1M, output $50 / 1M. Prompts above 272K input tokens bill at 2× input and cache and 1.5× output for the whole request; Batch and Flex run at 50%; Fast mode at 2×. OpenAI calls the model "our most intelligent and aligned model yet" and its first at the Critical cybersecurity capability level — vendor framing, and the benchmark claims (an ARC-AGI-3 figure under OpenAI's own harness among them) are unreproduced by us. It is rolling out to Plus, Pro, Business and Enterprise ChatGPT plans within their existing usage limits, so the subscription rows in our calculator do not move.

The Sol cut is the part that touches our numbers. OpenAI's August 21 announcement: "Sol now costs $4 per million input tokens and $20 per million output tokens, 20% lower input pricing and 33% lower output pricing", applying to Fast mode, long-context, Batch and Flex too, while "Pro, Plus, and Business subscription usage remains unchanged." The rate card marks it promotional, "available at least through November 21, 2026." Our row moves from a $17.50 to a $12.00 blended per-million rate. We do not pre-schedule a revert — that is the mistake the cancelled Sonnet 5 rise taught us — but a November re-check is on the list.

Why we missed it: the August 28 sweep re-read the OpenAI pricing page for Luna and Terra (cut July 30) and did not re-quote Sol, because nothing had flagged it. That is the same failure shape as the Gemini 3.7 Flash miss in August — a vendor halved a price quietly and we kept the old figure — and, as then, the error flattered local hardware by overstating the cloud side. The fix this sweep: every row in the calculator gets re-quoted from its vendor page on every sweep, not only the rows that made news.

Two same-price version bumps landed alongside and are renames, not repricings: Claude Fable 5.1 replaced Fable 5 at the top of Anthropic's rate card at the identical $10 / $50 (cache hits drop to $0.25 / MTok from $1), and Gemini 3.8 Flash is now Google's top Flash tier at the same introductory $0.75 / $3.75 through December 31, 2026 that 3.7 and 3.6 Flash carry. There is still no Gemini 3.x Pro tier above 3.1 Pro Preview. Local picks are unaffected by all of this; the cloud side of every payback chart on this site now uses the corrected figures.

Where this fits

Models: DeepSeek V4-Flash · Kimi K3 (2.8T-A50B) · Qwen3.8-Max (2.4T-A95B)

Hardware: NVIDIA RTX 5090 · Mac Studio M3 Ultra 96 GB

Sources

Next step

Try this in the planner→