Remember when we said flat-fee pricing for access to AI models can be summarized as: Painless now, feel the real pain later?
Well, today is "later" for Cursor and Kiro users. As of September 15, 2025, Cursor continues its tradition of arbitrarily rolling out changes that are worse for the community, and Kiro has also decided that it wants some of your cash now. Cue the surprised Pikachu memes.
What happened? In short:
Cursor
Rejigged "unlimited" Auto mode for individual Pro users to now be “competitively priced”
Switched the Teams plan from per-request pricing to variable token costs
Kiro
Raised their pricing after already having decreased the number of requests from 3,000 to 700 4 weeks ago
September 15 update: At the last possible moment Kiro changed their pricing again, combining the spec and vibe request types into ‘credits’. The unpredictable complexity factor is still there.
Here’s why this matters.
"Competitive" Pricing That’s Significantly More Expensive Than the Competition’s
Cursor claims its auto mode will contribute to your included monthly usage at “competitive token rates.” Translated from corporate speak, that means "You're now paying per token, just like everyone else."
In terms of numbers, here’s what “competitive” means in Cursor-land:
Input + cache write: $1.25 per 1M tokens
Output: $6.00 per 1M tokens
Cache read: $0.25 per 1M tokens
Meanwhile, here’s the pricing for Grok Code Fast 1:
Input: $0.20 per 1M tokens (6.25x cheaper than Cursor)
Output: $1.50 per 1M tokens (4x cheaper than Cursor)
Cached: $0.02 per 1M tokens (12.5x cheaper than Cursor)
More for Less
Cursor’s auto mode dynamically selects AI models based on availability and cost-efficiency, meaning it’ll give you whatever model is cheapest, not the one you want. Sure, it might save on fees, but at the cost of your control.
Imagine if Netflix said "unlimited streaming!" but randomly switched your 4K movie to 480p whenever their bandwidth costs got too high. That's auto mode.
None of that is “competitive.” Not the pricing nor the features.
Cursor Teams Plans Shift to Variable Pricing
Cursor's Teams pricing also changed on September 15:
Before September 15:
250 Sonnet requests per month
Additional usage at $0.08 per request
After September 15:
$20 of agent usage per user
Ability to purchase additional usage
All usage is consumed at the publicly listed API prices of the underlying model + Cursor’s markup fee of $0.25 per million total tokens (Kilo doesn’t charge a markup on tokens for teams users)
Cursor explains the change:
“Using the same model, a single difficult prompt can consume an order of magnitude more tokens than a simple one.”
Tellingly, they themselves admit, “Fixed costs per request aren't suited to this new reality.” That's exactly why subscription pricing for AI never made sense in the first place.
Meanwhile at Kiro: "Vibes" Pricing
While Cursor realized they have to pull back on the unlimited promise, Kiro – the closed-source latecomer with the suspiciously familiar name (amazing what one letter can do!) – has decided it wants that cash and has introduced its own custom pricing model. The Amazon-backed tool decided to reinvent the wheel with something that’s, once again, willfully opaque. Here’re the terms they’ve introduced:
Spec requests are when you execute tasks within Kiro's structured development workflow
Vibe requests cover any agentic operation that does not involve spec requests
Easy, right? But dig a little deeper, and it gets funky. Just look at how they explain it:
We’ll give them the benefit of the doubt and assume they’re doing their best to give us some understanding, but this convoluted structure reminds us of the fine print in an insurance policy. Why not just, for example, charge what the usage costs?
Likewise, their overage pricing – $0.20/spec request or $0.04/vibe request – might sound reasonable until you realize you have no way to predict how many of each you'll need because all of this is based on average complexity. What does that entail, exactly?
Here’s the fundamental problem: proprietary units don't translate directly to tokens, making it difficult to compare costs across platforms. It's the same subscription trap Cursor fell into, just wrapped in different packaging, proving that being backed by Big Tech doesn't mean you understand the economics of AI any better than the rest.
The Pattern Keeps Repeating
We’ve seen this same thing play out before:
The hook: "Unlimited AI coding for just $20/month!"
The adoption: Developers integrate it into their workflow
The reality check: Power users consume $1,000+ worth of compute daily
The squeeze: Introduce "rate limits," "usage pools," or "competitive pricing"
The backlash: "Why, as an active paying customer, didn't I get any e-mail from Cursor with clear explanation of upcoming changes"
The apology: "We didn't handle this pricing rollout well"
Three months ago, we predicted it. Then we watched Cursor’s June-July pricing disaster unfold. Then Anthropic did it with Claude Code.
Anthropic, the company that makes Claude, realized it couldn’t sustainably offer unlimited, unthrottled access to their own models for a flat fee. If the creator and operator of the model can’t land this offer, how can any wrapper around their API do the same?
Simple: they can't. We’re not sure why Kiro’s trying a similar playbook.
We weren't being prophetic. We were just doing basic math. And developers are doing the math too.
The Bottom Line
September 15, 2025 marked another milestone in the AI coding subscription game.
Cursor’s users are learning what we've been saying for months: there's no sustainable way to offer unlimited access to expensive AI models at consumer price points.
The future isn't subscriptions with hidden limits and sudden changes. It isn’t arbitrary units based on vibes either.
It's transparent, usage-based pricing that respects both the economics of AI and the intelligence of developers.
That’s what we’re about at Kilo Code. Because when you align incentives properly, everyone wins:
You get the exact model you want
You pay only for what you use at no markup (our money comes from Teams and Enterprise plans)
You can optimize costs in real-time
You never get surprised by a pricing change
Because being opaque with your developer community isn't just bad ethics – it's bad business.
Welcome to the honest side of AI coding. Your tokens are waiting, no subscription required. And we launched Autocomplete on September 4. It isn’t as good as Cursor’s yet but it will continue to improve this month and next.
P.S. To every Cursor or Kiro user reading this who's tired of the pricing gymnastics: we've got your back. Join our Discord where you can even talk about competitors (yes, really). First $5 in credits on us (plus $20 more if you top up at least $5), no strings attached.
Nicely said