Token cost is a big topic right now, and the pressure is building across the industry: the hyperscalers funding much of the compute have watched their combined free cash flow fall by nearly 90% in 18 months, and that pressure trickles down to your coding bill. We also called the $100k/year-per-developer AI bill, and cost isn’t slowing down. The teams that are ahead don’t cut back on AI use, but rather leverage visibility into their spend, and route across frontier and open models.
That’s the whole premise behind the four cost-saving levers we’ve written about before: model choice, observability, governance, and time/context. This last focus week, our team put serious work in the first two, model choice and observability, and shipped a batch of things that make your Kilo spend easier to see and control.
Here’s what’s live now.
You now know which model actually did the work
With the launch of Auto Efficient, model routing is driven by the session itself. When you use it, your session is first classified by the kind of work it is, and then, using that classification together with our public benchmarks, KiloBench, Kilo picks the right models for it. On KiloBench, that means Auto Efficient delivers 71% of published frontier completion at 72% lower cost. That’s been live for a couple of weeks. What’s new is that the router gives you more visibility during the session.
You can now see, turn by turn, which model handled your request. Select Auto Frontier or Auto Efficient and you’ll see the actual models that ran live during the session. If you’ve ever wondered what happens under the hood when you use Auto Model routing, now you can just look.
Users could already see which model ran from the usage dashboard, but now that visibility lives right in the task itself. That matters beyond just trust: seeing which model handled which kind of work is also how you learn what actually needs a frontier model and what doesn’t. We think full visibility into cost and models used is the only way to build real AI capability without getting locked into a single vendor.
Underneath that transparency, we also ran a fresh batch of benchmarks on KiloBench, our continuously-updated coding benchmark. More models and more recent data feeding the routing decisions. That means when Auto Efficient hands your task to a model, that call is backed by benchmark numbers that are days old, not months old.
You can compare your inference options
Model choice isn’t just “which model.” It’s also “how am I paying for this”: pay-as-you-go, a coding plan, your own API key, a subscription bundle, or a combination of all of these. Until now, figuring out which combination made sense meant some guesswork.
We shipped a single page for that: ** kilo.ai/models/inference**. It lays out coding plans, subscriptions, and Bring Your Own Keys (BYOK) options side by side, so you can see what you can actually bring into Kilo and what each path costs before you commit to one.
Once you know your options, the next question is provider-level: two providers can serve the exact same model at meaningfully different cost and reliability. That’s what the ** Kilo Gateway inference provider leaderboard** is for. It shows how inference traffic was distributed across providers last week for the 20 most-used models, ranked by total token volume, and for each one you can see cost, cache efficiency, and error rates side by side.
If you want the savings math done for you rather than eyeballing a leaderboard, the new ** Cost Savings Calculator** does that: toggle the features you’d actually use, and it projects your savings using KiloBench and live pricing data. Since coding plans went live in Kilo with MiniMax as our launch partner, we’ve expanded the MiniMax plans available to purchase directly in Kilo. Same idea as before: buy with the credits already on your Kilo balance, no separate subscription or invoice to manage.
You can see exactly where your balance went, everywhere you work
None of the above matters if you still can’t answer “where did my money actually go this week.” So we went after that directly.
The new ** credits page** gives you full credit history and deductions: every debit, in order, instead of one number that changes without explanation.
Your Kilo Balance now shows up everywhere you work. CLI, VS Code, and JetBrains all surface your current balance and Kilo Pass status directly, not just the web dashboard. If you live in your editor or terminal all day, you shouldn’t have to shift to a browser tab to know if you’re about to run low.
We also improved the model picker itself: auto-model grouping and highlights right in the extension, so the model picker doubles as a cost-visibility surface instead of just a dropdown of names.
And for the moment when visibility needs to become a hard stop: Session Cost Alerts now let you set a max-cost threshold in CLI and VS Code, with an alert when you cross it. Here’s what that looks like in practice:
The pattern underneath all of it
None of these ship in isolation. Auto Efficient tells you which model ran. The comparison page and calculator tell you what your inference options cost before you buy. The credits page, balance-everywhere, and cost alerts tell you where the money’s already gone. Put together, that’s helping you optimize your AI coding spend.
The pricing pressure across the industry isn’t going away. Usage-based billing is becoming the default, not the exception. The teams who come out ahead of that aren’t the ones spending less, but rather the ones who leverage model freedom and aren’t locked into a vendor.
If you want to go deeper on the model side, we also put together landing pages comparing Auto Efficient against frontier models with the full KiloBench data and one-shot prompt comparisons, if you’re evaluating this for a team and want something to share around. And one thing shipping soon: granular cost insights and recommendations on paths forward, so you don’t have to dig into the data yourself to optimze your spend and understand your usage.
Try Auto Efficient · See the inference comparison · Try the Cost Savings Calculator · Check your credits