Rollover Minutes Are Back. Except Now They're Tokens.

Rollover Minutes Are Back. Except Now They're Tokens.

I've been joking about this for weeks. Holding up an imaginary sign: will work for tokens. Offering increasingly desperate favors for another thousand context window. The memes are already out there — the internet figured out that tokens are the new currency of getting things done, and the humor writes itself. But something just happened that isn't funny. It's actually significant. And it's one of those things that sounds like a minor product update until you recognize the pattern it belongs to.

OpenAI just introduced banked resets for Codex.

Instead of rate limit resets firing on a fixed schedule whether you need them or not, eligible users on Plus and Pro plans can now accumulate their resets and deploy them whenever they actually need them. Save them up. Use them when a big project hits. Let them stack when you're light. That's rollover minutes. That's exactly what that is.

If you're old enough, you remember the cellular minutes era.

You paid by the minute. Long distance was a separate category of financial violence. Roaming charges were something people warned you about like a natural disaster. The phone companies had perfected the art of selling you a plan that almost covered what you actually needed, then extracting a premium for every minute you went over.

Then SMS showed up. At first it was a curiosity. Then it was how teenagers communicated. Then it was how everyone communicated. And the phone companies, with a straight face, began charging twenty-five cents per text message while simultaneously selling you the phone at a loss. They had found the new scarcity. The new metered resource. The new thing they could sell you a plan for and penalize you when you exceeded it.

The minutes became the loss leader. The texts became the business.

Rollover minutes came along when the minutes stopped being enticing. Suddenly AT&T was very interested in letting you bank the thing you were already paying for. It was a retention play dressed up as generosity. They had already shifted the extraction mechanism to texts and data — the minutes were cheap to give back.

Then texts went unlimited too. And the whole cycle shifted to data.

Every time one resource gets commoditized, the industry finds the next scarce thing to meter, sells you a plan just tight enough to create anxiety, and eventually offers you rollover as a loyalty feature once the real money has moved somewhere else.

We are in the cellular minutes era of AI.

Tokens are the minutes. Context windows are the data plan. Rate limits are the overage charges. And the memes — will work for tokens, trading favors for compute — are the 2026 version of joking about your phone bill.

Codex just crossed three million weekly active users. That's not a product in early adoption. That's a product with a user base large enough that inflexible rate limit resets became a loudly documented complaint across developer communities for three straight months. OpenAI didn't introduce banked resets because they're generous. They introduced banked resets because the friction was becoming a competitive problem and the feature costs them very little.

That's exactly how rollover minutes worked. The minutes were already paid for. Letting you roll them over didn't cost AT&T money — it cost them a retention argument they no longer needed because they'd already moved the real business to texts and data.

What has OpenAI moved the real business to?

Longer context. More capable models. Enterprise contracts. API usage at scale. Agent-driven automation that runs whether you're watching or not. The per-session rate limit on Codex isn't where the money lives anymore. Letting you bank resets is the move you make when the constraint you're releasing isn't the constraint you're actually monetizing.

Here's where I think this goes.

The token economy right now looks like early mobile pricing — complicated, opaque, full of artificial scarcity that mostly creates frustration without serving a real engineering purpose. Every platform has its own unit. OpenAI has tokens and credits and resets. Anthropic has message limits. Google has queries. Each one is calibrated to create just enough friction that you notice the constraint without quite being angry enough to leave.

That's not a stable equilibrium.

Mobile pricing got simpler when competition forced it to. Unlimited minutes became the standard not because carriers suddenly grew a conscience but because the unit of scarcity shifted and the old metering structure became a competitive liability. The companies that held on to per-minute billing the longest looked like they were clinging to something customers had already decided they deserved for free. The same pressure is building in AI.

Open-weight models are getting better fast. Local inference is becoming viable for an increasing number of use cases. The argument that you need to pay per token to access a frontier model is going to get harder to make as the frontier moves and last year's frontier becomes this year's open-source release. Platforms that have already moved the real value proposition — deep integration, enterprise workflow, agent orchestration, proprietary data advantages — will be fine. Platforms that are still primarily monetizing raw token access are going to feel this. The banked reset is a small thing. But small things that mirror historical patterns are worth watching.

The part that I find genuinely interesting isn't the feature itself. It's the referral mechanic attached to it. OpenAI paired the banked reset launch with a program where existing users can refer up to three friends, and when a new user completes their first Codex interaction, both the referrer and the new user earn an additional reset. The conversion event isn't a subscription. It's a first use.

That's not a discount. It's not a credit toward a future purchase. It's a unit of the scarce resource itself — the thing you actually want — used as the incentive to expand the network. That's the moment when a resource starts behaving like currency.

Rollover minutes were valuable because minutes had value. Banked resets are valuable because compute access has value. Using them as referral incentives is the move you make when you know your users feel the scarcity acutely enough that the resource itself is more compelling than a discount on something they may or may not buy.

The joke about working for tokens isn't really a joke anymore. It's an accurate description of how people are thinking about AI compute access. And when the industry starts issuing the resource as a referral reward, you're watching something become currency in real time.

We've been here before. The details are different, the underlying dynamic is identical.

Scarce resource gets metered. Plans get designed to create friction just below the rage threshold. Power users complain loudly. The platform introduces a flexibility feature — rollover, banking, pooling — that costs them little but reduces the most vocal source of churn. The real monetization has already moved to the next layer.

The question worth asking now is what the next layer is. Because by the time the token plans go unlimited — and eventually they will — the industry will have already moved the real money somewhere else. If I had to guess: it's the agent that acts autonomously on your behalf, continuously, in the background, across systems you couldn't coordinate manually. Not the token you spend asking a question. The subscription to the thing that works while you're not watching.

Tokens are the minutes. Agents are the data plan.

And we're somewhere around 2003, trying to figure out why our text messages keep costing us twenty-five cents each.

---

Ad: Use code: RICH99 for a discount

Rich Washburn is a technologist and strategist working at the intersection of AI, infrastructure, and capital. He is Managing Partner and Chief AI Officer at Eliakim Capital and CIO of Data Power Supply.