01
How hard is this, really?
Not how long the prompt is: what it asks, how much of the codebase it reaches, and how the session is going. The same one line can be routine or risky.
Product · Token optimizer
Efficient Router
The right intelligence for every intention.
For every prompt in Cursor, Claude Code and Codex, the Efficient Router picks the most capable model at close to the lowest price for that exact task, and the effort it needs. Routine work goes light, hard problems go to the strongest model, and every switch shows where you work with a one-line reason.
Start free trialIn Cursor, Claude Code and Codex, on your machine.
About 100x
The price gap between the lightest and the strongest model, in every tool.
16 vs 12
Tasks of 20 solved by Opus 5.5 at medium effort ($0.46 a task) and at extra high ($0.94 a task).
1.6–14x
What max effort cost over a lower effort on the same model that solved as many tasks or more.
Prices are published list prices per million input tokens (Claude: Haiku 5.5 $0.10, Fable 5.1 $10; Codex: GPT-6 Luna $0.10, GPT-6 Astra $10). Task results are from our own 820 Terminal-Bench 2.0 runs through Claude Code and Cursor, October 2026: 40 configurations of model and effort, 20 tasks each (Claude Haiku 4.5 ran twice). In those runs, lower effort was never significantly worse than higher effort on the same model, and max effort scored worst or tied-worst for 5 of the 7 models that offer it. Twenty tasks is a small sample. Benchmarks: every model and effort we measured →
Four questions, before every prompt sends.
Answered on your machine, in the moment between pressing send and the prompt going out.
02
Which model is enough?
The most capable model at close to the lowest price for this task. Hard work is never moved to a weaker model to save money.
03
How much thinking does it need?
Effort is chosen with the model. More thinking is not always better: in our runs it often cost more and solved less.
04
Is a switch worth it?
When the saving would be small, or it isn't sure, it leaves your model alone. An unnecessary switch costs more than a missed saving.
Real recording
A small edit, moved off a frontier model.
Claude Opus 5.5 was selected for a one-line request. OptiFlume switched the model under the composer to GPT-5.2 before the prompt went out.
Before, and with OptiFlume.
Before
With OptiFlume
BeforeOpus on a typo fix, at frontier prices.
WithA lighter model for the typo and the strongest for the redesign, each with a one-line reason.
BeforeGuessing which model and effort to pick, prompt after prompt.
WithPicked before the prompt sends, and you can always keep yours.
BeforeMax effort on everything: slower, dearer, and in our runs no better.
WithThe effort the task needs, chosen with the model.
BeforeRoutine prompts burning through your plan on a frontier model.
WithRoutine work goes light, so the strongest model is there for the work that needs it.
BeforeA router that takes a cut of every token it sends.
WithOne flat price, no cut of your spend, and no reason to upsell you.
01
You have the last word
You can always choose not to route a prompt: it goes out on the model you chose. A switch never changes your saved default model, and picking a model yourself never turns routing off: the next prompt is checked again.
Pause it for a project, set it to favor quality over cost, or turn it off for one tool. Every way to steer it is in the docs.
02
What it saves
Your console counts every switch, in every tool. In Cursor, where the token counts are available, it also shows the spend avoided, at published prices.
And it has no reason to upsell you: OptiFlume charges one flat price and takes no cut of your spend.
What you see in each tool.
On as soon as OptiFlume is installed, with your own account and plan in each tool, and in every tool you can keep your own model for any prompt. Antigravity, Copilot and others come next.
Cursor
The model in Cursor's picker changes, a note above the composer says why and how the new price compares with yours, and the prompt sends on the new model.
Claude Code
In the VS Code panel, a line under your prompt and the model chip changes. With claude in the terminal, the line appears and the prompt sends on the new model.
OptiFlume: routine change → Sonnet 5 · lowCodex
In the VS Code panel, the model picker shows the model that ran and the reason sits under your prompt. With codex in the terminal, the line appears under your prompt.
OptiFlume: routine change → GPT-5.6 Terra · lowWon't a cheaper model write worse code?+
Hard work still goes to the strongest model; only routine work goes light, and when OptiFlume is unsure, nothing changes. Effort matters as much as the model: in our own runs, lower effort was never significantly worse than higher effort on the same model.
Does a switch cost me an extra request?+
No. OptiFlume decides on your machine before anything is sent, so the prompt goes out once, on the chosen model. When it doesn't switch, nothing changes at all.
Which models can it pick?+
Only models you can already use: the models your Claude account can run in Claude Code, your ChatGPT account in Codex, and the models you have turned on in Cursor's settings.
What happens on the free tier?+
OptiFlume names the better-suited model and your prompt goes ahead on yours; it is never stopped. Pro and the 30-day trial switch for you, in every tool.
How do I see what it saved?+
The console counts every switch in every tool. Dollar savings are counted in Cursor, where the token counts are available; Claude Code and Codex don't report them.
The other systems, in the same install
Context management & collaboration
Context Manager →
Long sessions keep their goals, decisions and dead ends through compactions and switches.
Safety, security and focus
Guardian →
Secrets, personal details, card and bank details kept out; destructive commands caught; loops called out.
Learns how you work
Adaptive Intelligence →
Reads your code and your session, and learns from every choice you overrule.
Start with your next prompt.
Pro is free for 30 days, no card. Install, keep working, and see what it did and why.
macOS 11+ · Apple silicon and Intel · Signed by OptiFlume, notarized by Apple · Never locks you out · Uninstall with one command