GPT got cheap for two years. Then the bill came back.

GPT got cheap for two years. Then the bill came back.

The default model raced toward pennies. Astra is priced like the old flagship, on purpose.

For two years the headline about OpenAI was a price cut. GPT-6 Astra, out last week, is the first flagship in that stretch whose output token costs $50 per million. That is the same list Anthropic put on Fable 5.1. The cheap era did not end. A second price list grew on top of it.

This is the bill behind Opus losing the crown and behind the $50 security tier.

What a million tokens cost

Standard rates, input and output, per million tokens. The first row is the flagship as this window opened. The last row is the flagship now.

Two lines are hiding in that table.

The default flagship fell. GPT-4o launched at $5 / $15, was $2.50 / $10 by August 2024, and GPT-5 in August 2025 was $1.25 / $10. Luna, the current cheap GPT, is $0.20 / $1.20. From the May 2024 flagship to Luna, output tokens dropped from $15 to $1.20.

The ceiling did the other thing. o1 opened a reasoning price at $15 / $60 in September 2024 and kept it in December. GPT-4.5's preview in February 2025 was a one-off spike at $75 / $150, and it did not become the default. o3 launched at $10 / $40 and was cut to $2 / $8 within two months, which is the old pattern: ship high, then join the race to the floor. Astra did not join that race. At $10 / $50 it costs eight times GPT-5 on input and five times GPT-5 on output. Next to Sol, the previous expensive GPT at $4 / $20, Astra is still more than double.

The cut was the product, until it wasn't

From autumn 2024 through summer 2025, a price cut was how OpenAI told you a model had graduated. o3's 80% reduction was the announcement. GPT-4o's cut from $5 to $2.50 was the announcement. Developers learned to wait a quarter and budget the second number.

Astra breaks the habit. It is not a preview with a footnote that the price will collapse in June. It sits on the enterprise rate card beside Luna and Sol, as its own tier. Fable 5.1 sitting at the same $10 / $50 means the price is no longer an OpenAI quirk. Two labs looked at the cost of a top agent model and printed the same rate.

Under them, the floor is intact. Luna at $0.20 / $1.20, Gemini 3.8 Flash at $0.75 / $3.75 through December, Grok 4.7 at $2 / $6, Sonnet 5 at $2 / $10. A review platform that sends every diff to Astra is choosing the 2024 reasoning price after the industry spent two years abolishing it.

What to budget

Split the invoice the way the menu is already split.

• Volume belongs on the floor. Luna, Flash, Grok, Sonnet. A pull request comment is output-heavy, so the $1.20 versus $50 gap is the whole decision. On output tokens, one Astra review costs about as much as forty Luna reviews.
• The hard general diff can sit in the middle. Sol at $4 / $20, Opus 5 at $5 / $25, GPT-4.1's successor prices if that is the model you pinned. You are paying for a stronger pass, not for a cyber program.
• Astra and Fable are the new top, and they earn it on the work described in the security piece: long agent runs, and the fenced security tasks. They do not earn it as the default reviewer.

GPT-4.5's $150 output token is the cautionary row. An expensive model that is not clearly better than the one under it does not stay. Astra will be judged the same way, against Fable on agent scores and against Grok and Flash on whether the extra dollars changed the finding.

ThinkReview keeps that choice in Model selection. The cheap models are the ones to leave on. The $50 model is the one you switch to, on GitHub, GitLab, Azure DevOps, and Bitbucket, when the diff is worth the old price of a flagship.

GPT-4o, o1, GPT-4.5 preview, GPT-4.1, o3, and GPT-5 rates are the published API prices at each launch or cut. Luna, Sol, and Astra are on OpenAI's current enterprise rate card. Fable and Opus rates are Anthropic's. Review on the tier you mean to pay for with ThinkReview.