The 10-cent model has a 100-second catch

Independently measured, it is the cheapest model anywhere near its score. At the setting that earns that score it thinks for 100 seconds before it says a word. Here is what to move to it, and the 2...

Welcome to the Toolkit, the Friday half of The AI Debrief. Tuesday is what happened. Friday is what to do about it: one tool worth your time, one play you can run today, and nothing you have to buy.

Here's what we're getting into this week. On Tuesday OpenAI shipped two GPT-6 models below Astra, and the small one rewrote the bottom of the price list. Last Friday the cheap one and the good one were two models. This week they are one model with a dial on it, and the dial is the bill.

In this email

» GPT-6 Luna on the independent leaderboard: the score, the cost per task, and what "a hundredth the cost" actually measures

» The effort dial, setting by setting, with the score and the price of each

» A 20-minute routing test that moves your cheap jobs to the cheap model and leaves the rest alone

💰 The tool: the floor just moved

GPT-6 Luna

On September 22 OpenAI released GPT-6 Sol and GPT-6 Luna, the two tiers under Astra. Sol is $2 per million tokens in and $10 out. Luna is $0.10 in and $0.50 out, with cached input at one cent. The model page gives Luna a 1,050,000-token context window, 128,000 tokens of output and a reasoning dial with six positions from none to max. In the API it is gpt-6-luna. In ChatGPT it is in Work and Codex for Plus and up, and Free and Go users get it in the desktop app.

Artificial Analysis had both models on its board within a day, which is the part worth your time. At max effort, Luna scores 37 on the Intelligence Index and costs seven cents to run one of the index's tasks. GPT-5.6 Luna, the model it replaces, scores the same 37 at 18 cents. GPT-5.6 Sol, the model the launch post compares it to, scores 47 at $1.99. Of every other model within five points of it that carries a cost figure, the nearest is its own predecessor at two and a half times the price.

The price ladder · index score · cost per task

Claude Opus 5.5 max

$5.98

GPT-6 Astra max

$3.26

GPT-6 Sol max

$1.06

GPT-5.6 Sol max

$1.99

Gemini 3.8 Flash high

$1.24

DeepSeek V4.1 Flash max

$0.27

GPT-5.6 Luna max

$0.18

GPT-6 Luna max

$0.07

Scores and costs from Artificial Analysis, read September 24. Bars are cost, scaled to Opus 5.5. None of these numbers come from the vendors.

Now the sentence everyone quoted. The launch post says that "at higher effort levels it matches GPT-5.6 Sol at about a hundredth its cost." Read the sentence before it: that is OpenAI's internal factuality evaluation, built from real conversations where users flagged a mistake. It measures how often the model gets a fact wrong, not how well it reasons, and on the independent index it sits ten points below Sol. Both are true. The honest version: it makes things up about as rarely as last quarter's mid-tier, at a fortieth of the token price, and it is not last quarter's mid-tier.

The effort dial · GPT-6 Luna, one model, six settings

none

no thinking tokens
chat speed

low

under a cent
a task

medium

29 · $0.02
the default

high

32 · $0.03

xhigh

34 · $0.04

max

37 · $0.07
103s to first token

The catch

Every setting above none makes the model think before it answers, and OpenAI's reasoning guide is plain about the cost: reasoning tokens are invisible, they take up context, and they "are billed as output tokens." So $0.50 a million buys every token it thinks with as well as every token it writes, and at max it thinks for a while: Artificial Analysis measured 102.8 seconds to the first token, reasoning included, the slow end of reasoning models. Once it starts it runs at 131 tokens a second. The 37 is real. The 100-second wait is what it costs.

Who this is not for. Anything a person is waiting on in real time at high effort, and anything where ten index points is the difference between right and wrong: contracts, code you will not read, numbers that go in front of a customer. Those stay on Sol or Opus. Everyone else has been paying mid-tier prices for jobs a seven-cent model can do, and has not checked.

⚡ Steal this: the 20-minute routing test

Find out which of your jobs the cheap model can already do

What it replaces: guessing. Most people pick a model once, on the day they build the thing, and never look again. What it costs: a few cents of API credit and twenty minutes. No key? The free ChatGPT desktop app runs the same model, and step five is for you.

The routing test

20 minutes · five prompts · two settings

Pull your five most-run prompts. Not the clever ones: the one in a cron job, the one in a Zap, the one you paste every morning. Save each with one real input you already know the right answer to.

Run each one through gpt-6-luna twice with the snippet below, once at effort none and once at high, and once through whatever it runs on today. Keep the usage block from every call.

Grade pass or fail against what you actually needed, not against the expensive model's prose. If you could not tell which answer came from which, that is a pass.

Route by the result, using the three columns below. Change the model string and nothing else, then watch a week of output before you touch the prompt.

No API key? Open the ChatGPT desktop app, pick GPT-6 Luna, paste the prompt you use most. If the answer holds up, you have just learned what that task costs to run: about nothing.

from openai import OpenAI
client = OpenAI()
PROMPT = open("prompt_1.txt").read()

for effort in ["none", "high"]:
    r = client.responses.create(
        model="gpt-6-luna",
        reasoning={"effort": effort},
        input=PROMPT,
    )
    u = r.usage
    print(effort, "output:", u.output_tokens,
          "reasoning:", u.output_tokens_details.reasoning_tokens)
    print(r.output_text[:600])

Passed at none

Move it. Luna, effort none. It now costs cents and answers at chat speed.

Passed only at high

Move it if nothing is waiting. Luna, effort high. Expect the pause; batch it if it can wait an hour.

Failed at high

Leave it. It stays on the model it is on. Ten points is ten points, and you just proved it.

The usage block is the whole point. Ask a model whether it was the right model for the job and you get a paragraph. Ask the usage block and you get two integers: how many tokens it wrote and how many it thought with. At effort none the second one reads zero. When a prompt passes at none, you have found the cheapest correct answer on the market, and you have the receipt.

Batch is the second discount, and it stacks. OpenRouter's Batch API came out of beta on Tuesday: about half the per-token price on more than 70 models, and across 230,000 beta batches the median came back in seven minutes and the 99th percentile in ten hours. GPT-6 Luna in batch is five cents in and 25 out; OpenAI's own batch endpoint takes the same 50 percent off. A nightly summariser, a weekly classifier, a backfill of old tickets: none of those need an answer in seven seconds.

The test outlasts the model. Every time a price list moves, it is the same five prompts, the same two settings, the same two integers.

Two more numbers from Tuesday, for the weekend. GPT-6 Sol is $2 in and $10 out, and Anthropic shipped Claude Opus 5.5 the same day at $4 in and $20 out, claiming it costs 40 percent less to run than Opus 5 on typical workloads. On the independent board it is the top score at 58, and the cost per task is a shade above Opus 5, not below it. Tuesday's Debrief takes both apart.

If you do only one thing from this issue, run one prompt through Luna at effort none and read the usage block. Honest worst case: it fails, and you keep paying what you pay, now on purpose. Likely case: one job on your list has spent months on a model forty times more expensive than it needed.

Every test in this issue gets run, argued over and compared by the people in the free AI Academy.

Join the free AI Academy →

Free, no card. The tools and people behind this issue live here.

That's the Toolkit.

See you Tuesday for the Debrief.

From Drew and The AI Debrief team

P.S. What is the one task you would never hand to the cheap model, whatever the leaderboard says? Hit reply and tell us. We read every one.

P.P.S. New here, or skipped one? Every past edition is in the archive.