Our team was burning through a ton of tokens each month. Our finance team said slow it down.
Our lead engineer, Rampey, built a tool to help us route our prompts more efficiently to the right model.
We were surprised by the results, saving 50%+, and decided to share it with the world.
for 
ChatGPT
FREE
The installer adds an Untitled Auto entry to the model picker in the ChatGPT/Codex app. Everything else stays exactly where it was.
~ % curl -fsSL https://tryuntitled.ai/install.sh | bash
__ _ __ __ __ __ ______ / /_(_) /_/ /__ ____/ / / / / / __ \ / __/ / __/ / _ \/ __ / / /_/ / / / // /_/ / /_/ / __/ /_/ / \__,_/_/ /_/ \__/_/\__/_/\___/\__,_/
Untitled Auto
Automatic model routing for Codex
prompt --> [Auto] --> the best available model
One picker. Smarter routing.
ChatGPT account connected
Read the account model list from Codex
Untitled installation registered
Local Untitled relay is ready
Untitled menu-bar status is ready
Untitled Auto is installed in the stock ChatGPT/Codex app.
Untitled Auto sits in the same model picker you already use. Quick edits go to a light model, hard problems to the strongest one, and your plan lasts longer.
A small status item in your menu bar shows the local relay is running, how far your plan now stretches, and when your usage resets. Switch back to stock Codex any time, in one click.
Real reviews from real testers. Want the inside scoop?
Join our Telegram group for beta testers.
A router for Codex. Pick Untitled Auto once, and every request goes to the lightest model that can do the job.
Run the command on macOS. A small local relay hooks into the Codex you already use, signed in with your existing ChatGPT login. It works in both the Codex CLI and the desktop app.
Quit Codex, open a new session and pick Untitled Auto from the model picker. Sessions that were already open, or that you resume, keep your old settings.
Untitled Auto sends every request to the lightest model that can handle it. Changes to existing code usually go to a lighter model; deploys, tooling and new projects step up to the flagship. Pick any other model and your requests pass through unchanged.
The banana in your menu bar shows how much usage you’ve saved and how much further your plan goes. The first figure appears after a few requests.
Codex keeps control of tools, permissions and approvals. Uninstall from the menu bar any time to go back to standard routing.
Testers’ real Codex work, measured against running the same tokens on Sol, the flagship model.
1.87×
more Codex from the same plan
46.49% of usage saved across the whole workload
DeepSWE: 113 real-world coding tasks, scored by whether the fix actually works.
Untitledtasks solved
Untitled Auto · 70.6%cost per task (log scale) →
Sol-high quality.
A third of the cost.
Every saving is measured against what the same work would have cost on Sol, and checked for quality, at our expense.
Untitled picks a random sample of prompts and runs each one again on Sol, the flagship model, using Untitled’s own paid account.
Both runs are priced at the same token counts. The saving is the gap between the two costs.
An agent compares the two answers without knowing which model wrote which, and judges which is better.
The menu bar shows the share saved and what it means for your plan:
extension = 1 ÷ (1 − share saved)
| Saved | 40% | 46% | 50% | 74% |
|---|---|---|---|---|
| Plan goes | 1.67× | 1.85× | 2.00× | 3.85× |
Untitled adds an Auto model to Codex. It sends each request to the cheapest model that can handle it, so your ChatGPT plan’s Codex allowance lasts longer. Easy edits go to a lighter model; deploys, tooling and new projects go to the flagship.
Across all testers, Untitled saved about 46% of Codex usage, roughly 1.87× more from the same plan. Your own number depends on what you work on: testers have seen anywhere from 8% to 74%.
In blind comparisons against Sol, the flagship, Untitled Auto scored the same or better on correctness. For harder tasks it moves up to a stronger model on its own.
On macOS, run the one-line command at the top of this page. Then quit Codex, start a new session and pick Untitled Auto. It works in both the Codex CLI and the desktop app.
You’re probably in an old session. Sessions opened before the install keep your previous settings. Quit Codex (or open a new terminal), start a new session on Untitled Auto, and savings appear after a few requests. If you still see a dash, run the install again to get the latest version.
Not yet. Switching models in the middle of a resumed session can fail, so start a new session.
The GPT models in your Codex picker, from GPT-6 Astra down to GPT-5.5. Lighter models handle most changes to existing code; stronger ones take harder work. If newer models are missing, run the install again.
Usually, yes. Untitled uses a stronger model only when the task needs one.
Operational telemetry excludes prompts and answers. Quality capture is a separate choice during installation; if you allow it, sampled prompts and responses are sanitized and may be retained to evaluate and improve routing.
No. Comparison reruns are billed to Untitled’s own account, never yours.
The script pulls our GitHub release, the same way tools like Foundry and Nix install. You can read install.sh before you run it.
macOS today, Windows next. Claude Code support is planned. Other harnesses such as Hermes could work but need extra integration, and the community has built a status-bar plugin for oh-my-pi.
Another app may be using the same port. Reinstall on different ports, then quit Codex and start a new session:
curl -fsSL https://tryuntitled.ai/install.sh | UNTITLED_LOCAL_PROXY_PORT=28787 UNTITLED_LOCAL_ENGINE_PORT=38787 bashOpen the banana menu in your menu bar and choose Uninstall Untitled. Codex goes back to standard routing.