PANONDA
Últimas
← News
Ferramentas·4 min read

GPT-6.2 Astra got 20% cheaper to ask. It's still four times the price to let it write.

OpenAI launched GPT-6.2 Astra at $8 per million input tokens and $40 per million output. The cheaper Sol stays at $2/$10. For publishers, the gap only matters at volume — here's the math.

By PANONDA Newsroom

Foto: Markus Winkler · Pexels

What happened

OpenAI announced two models on November 17, 2026: GPT-6.2 Astra and the cheaper GPT-6.2 Sol.

Astra is priced at $8 per million input tokens and $40 per million output tokens. Cached input — repeated source material the model has already seen — runs $0.40 per million. Sol stays where the previous Sol sat: $2 per million input, $10 per million output.

The headline from the launch coverage is a roughly 20% price cut on Astra. That cut is on input. Output did not get cheaper in any way that changes the comparison.

The comparison that actually matters

Not new versus old. Astra versus Sol, on the work you actually do.

Astra's output price is four times Sol's, which means the model you pick matters only in proportion to how much text you let it generate.

Run the numbers on your own volume before you pick a side:

  • 100,000 generated words a month is roughly 130,000 output tokens. On Astra that's about $5.20 in output. On Sol, about $1.30.
  • 1 million output tokens is $40 on Astra versus $10 on Sol. A $30 difference.
  • 50 million output tokens — an automated operation, not a newsroom — is $2,000 versus $500.

If you are a one-person newsletter, the difference between these two models is lunch. Say that out loud before you spend an afternoon optimizing it. The pricing story is real, but at individual-creator volume it changes nothing.

It starts to bite when you run automated summaries, transcript cleanup at scale, or metadata generation across a back catalog. That's where output volume stops being a rounding error.

Where the cached rate changes the shape

The $0.40 cached input rate is the part worth testing. If you send the same 40-page source packet, the same style guide, or the same transcript back to the model across a dozen prompts, the repeated portion can price at a twentieth of standard Astra input.

That makes a split workflow plausible: large stable source material in as cached input, Astra reserved for the hard synthesis, Sol handling headlines, social copy and first-pass summaries.

What nobody has published is how often production prompts actually qualify for the cached rate. Treat it as a number to verify on your own account, not a number to plan around.

What this story is not

It is not an endorsement of Astra for drafting. Nothing in the launch material establishes benchmark results on publishing tasks — editing, fact extraction, headline writing. Frontier-grade pricing is not evidence of frontier-grade copy.

It is also not a complete picture. The exact context windows and rate limits for both models were not confirmed. Neither was whether the 20% reduction applies uniformly across every prior Astra tier or only to the cited API rates. If you're budgeting off those, ask OpenAI directly.

And for small teams, the API comparison may be beside the point entirely. A ChatGPT subscription with usage limits can be cheaper than any per-token math — until it isn't.

The two-week problem

One pricing guide lists Claude Sonnet 5.5, Gemini 4 Argon and GPT-6.1 Sol all at $2 per million input and $10 per million output — the same tier as the new Sol. The same guide says Google's Gemini 4 Carbon and Argon are scheduled to double to $4/$20 on December 1, 2026.

If that holds, a model you benchmark this week could cost twice as much before the month ends. That is a reason to delay a long-term commitment, not a reason to switch today.

Separately, Microsoft announced public-preview pricing for MAI Image 2.7 at about $35 per 1,000 images, with a Flash version around $10 per 1,000. Useful to log. Not enough published about quality or commercial terms to act on.

What to do this week

  1. Pull your actual output token count for last month from your API dashboard. Not an estimate.
  2. Multiply it by $40 and by $10 per million. If the gap is under $50, stop optimizing and pick on quality.
  3. Run the same reported brief through Astra and Sol. Compare the finished piece, not the per-token rate.
  4. Check whether your repeated source material qualifies for cached input. Measure it on a real week of prompts.
  5. Price one finished article end to end — input, output, retries — on both models.
  6. Do not sign an annual commitment before December 1, given the scheduled Gemini change.
  7. Ask OpenAI whether rate limits differ by account tier before you move production volume.

Sources

Everything above was checked against these pages. Open them and see for yourself — that is why they are here.

  1. 1testingcatalog.com
  2. 2ranthis.com
  3. 3microsoft.ai
Share

Get the stories by email

What changes in AI for people who publish — sourced, and to the point.

No spam. Unsubscribe anytime.