FULL STORY

GPT-6 Astra Pricing and Reviews: Pricey but Capable?

GPT-6 Astra's pricing leaked at $10/$50 per million tokens. OpenAI engineers argued token pricing is misleading as reviews showed major capability gains amid doubts about official benchmarks.

2026-09-04 ~ 2026-09-04 · 3 episodes · 14 posts

Episode 1 · GPT-6 Astra pricing leaks: $10 input / $50 output per million tokens (2026-09-04, 7 posts)

On September 4, multiple community posts and the official pricing page revealed OpenAI's API pricing for the new GPT-6 Astra model: $10 per million input tokens and $50 per million output tokens, matching the rates of Fable 5.1. The output price is notably higher than current flagship tiers, placing it in the premium range for frontier models, and the community quickly began debating its positioning and value. The model is not yet broadly available.

Confirmed

  • Pricing: $10/million input tokens, $50/million output tokens, corroborated by multiple posts and the official pricing page (link shared by kimmonismus)
  • The rates match Fable 5.1 (relayed by Bindu Reddy and koltregaskes)

Unconfirmed

  • Bindu Reddy says the model is not yet GA and will open up broadly in a few days; this timeline is a personal disclosure pending official confirmation
  • A tweet cited by koltregaskes claims Astra will use a technology replacing compression (toggleable, default-on in the future), relying on notes, chat history, and tool outputs; this comes from a single source
  • A repost by thesaraharminta mentions benchmark results such as ARC-AGI-3 at 98.6% and FrontierMath, but the original post was truncated and the full figures can't be verified

Why it matters

At $50 per million output tokens, Astra enters the most expensive frontier-model tier, directly affecting developers' cost calculations and model choices. If the rumor about dropping compression technology is true, its context management and billing could change structurally—worth watching after the general release.

Episode 2 · OpenAI Engineer: Token-Based Pricing Is Broken, Cost Should Be Measured Per Task (2026-09-04, 4 posts)

OpenAI engineer Steven Heidel argued token pricing is misleading because GPT-6 Astra's efficiency makes it cheaper per task than models 13x cheaper per token, and amplified findings that Anthropic quietly raised Opus 4.7 costs 30% via a less efficient tokenizer.

Episode 3 · GPT-6 Astra reviews: strong gains but benchmark claims questioned (2026-09-04, 3 posts)

Early reviews of GPT-6 Astra report major capability gains—such as shrinking 33,000 lines of SQL to 1,800—and halved per-task costs, though chain-of-thought monitoring broke. Critics note gaps between official benchmarks (97.6% on FrontierMath Tier 4) and independent tests (61), with doubled pricing for roughly parity.