AI coding costs are the reason a lot of people quietly give up on using AI tools every day — you hit a usage limit, or the bill creeps up, and you stop. One programmer just showed a way around that: stop reaching for the "smartest" AI model by default, and use a cheaper, faster one instead. The result, in his own numbers: roughly 20 times cheaper, with no drop in output.
What Happened
A developer who goes by Igumnov wrote up an experiment on the tech site Habr. He'd been defaulting to the most advanced, most expensive AI coding models — the ones that top every leaderboard — assuming "smarter" meant "better for me." Then he noticed a metric called cost per task: not how smart a model is in the abstract, but how much it actually costs, in tokens (the small chunks of text an AI reads and writes, which is how you're billed), to finish a real piece of work.
By that measure, a cheaper model he calls "Luna" came out on top. On a 0–100 intelligence scale, Luna scores 38 — genuinely one of the lower scores among current models. The top model he tested, "Astra," scores 53. But Luna finished the same kind of task for $0.18, while Astra cost $3.26 for a comparable job. That's about 18x more expensive for roughly 40% more "intelligence" on paper.
Think of it like hiring a competent handyman instead of a specialist consultant to fix a leaky faucet. The consultant is more qualified across the board, but for a well-described, well-understood job, the handyman gets it done just as well, much faster to book, and for a fraction of the price.
The practical payoff was immediate: on a $20-a-month plan, he stopped running into the usage caps that most AI coding tools impose (a 5-hour limit and a weekly limit, similar to a phone data plan that throttles you once you've used your quota). Switching to the cheaper model meant far more work fit inside the same subscription.
What It Means for You
You don't have to be a professional coder to benefit from this shift. The same logic — match the model to the task instead of always grabbing the fanciest one — applies anywhere you use AI.
Work: If your company gives you a seat on an AI assistant with a monthly usage cap, switching your default model to a cheaper option for routine requests (drafting emails, summarizing documents, cleaning up spreadsheets) stretches that cap much further before you hit a wall.
Business: If you're paying for AI tools across a team, this is a direct line to margin. A small agency running AI-assisted work for clients can cut its AI tooling bill by double digits just by rightsizing which model handles which task, instead of routing everything through the priciest option by habit.
Study: Students using AI to learn to code, write, or research often burn through free-tier limits fast. A cheaper model means more practice runs, more draft attempts, and fewer "come back tomorrow" limit walls mid-project.
Creativity: Hobbyists prototyping an app, a website, or a small tool don't need a research-grade model to get a working first draft. Cheaper, faster models let you iterate more — try ten small variations instead of three — because each attempt costs almost nothing. If you just need quick, everyday AI help — resizing an image, writing a product description, generating a short script — without any subscription at all, free tools like the ones at mykreatool.com cover a lot of that ground before you'd ever need a paid coding assistant.
Income: Freelancers and solo founders who bill clients but pay for their own AI tools directly feel this in take-home pay. Lower AI costs per task means more of what you charge stays as profit, especially on repetitive work like formatting, boilerplate code, or first drafts.
How to Try It Right Now
You don't need to switch tools to test this — most AI coding assistants already let you change which model handles your request. Free option first, then the paid comparison:
1. Start free. Go to OpenRouter and browse its model list — it includes free-tier access to a range of open models, sorted by price and speed, so you can see the cost gap for yourself before spending anything.
2. Check your existing tool's model picker. If you use OpenAI's Codex (its coding assistant/CLI), open its settings and look for the model selector. Instead of leaving it on the default top-tier option, pick one of the cheaper models on the list and run a real task with it for a day.
3. Test for speed separately. Sign up directly at DeepSeek's API platform and try the DeepSeek-V4.1-Flash model — it runs at roughly 300 tokens per second, compared with an average of 25–50 tokens per second on many hosted services. A task that used to take 5–15 minutes can finish in under a minute.
4. Write a clear task first, then compare. The "cost per task" trick only works if your request is specific — a vague, half-formed ask forces even a cheap model to guess, burning more tokens correcting itself. Write the task the way you'd hand it to a competent junior colleague, with the key details already spelled out.
5. Track what you spend. Note the price and time for the same type of task on the cheap model versus the expensive one for a week. The gap is usually bigger than people expect.
Upsides and What Changes
The headline change is financial: work that used to cost dollars per task can drop to cents, and subscription limits stop being the bottleneck they were. But there's a bigger shift underneath it. The race between AI companies to build the single smartest model appears to be leveling off — even the "dumb" budget models are now competent enough for most well-defined work. The same thing already happened with context windows (the amount of text an AI can "remember" at once): the race went from 65,000 tokens to a million, and then largely stopped mattering, because most tasks never needed that much anyway. The next competitive front, according to this experiment, is speed — how fast a model can turn a request into a finished result, not just how clever it is on a benchmark.
Limitations
This is one person's self-reported benchmark, not an independently audited study, so treat the exact dollar figures as directional rather than gospel — your own costs will depend on your provider, your task mix, and your region's pricing. It also doesn't apply evenly to everyone: people who aren't used to writing a clear, well-specified request (often newer or non-technical users) tend to rely on the smarter, pricier model precisely because it's more forgiving of a vague prompt — for them, switching to a cheaper model may mean more back-and-forth, not less. And "cheaper" only pays off if you're actually finishing the task correctly the first or second time; a cheap model that requires five retries can end up costing more than one expensive attempt.
Conclusion and One Action for Today
The core lesson isn't "always use the cheapest AI model" — it's that the most expensive, most advanced option is often overkill for tasks you can describe clearly. Do that today: open your AI coding or writing tool's settings, switch the default model to the cheapest one on the list, and run one real task on it before automatically reaching for the expensive model again. You'll likely find it finishes the job just fine — for a fraction of the cost.
Comments 0