What Does One Antigravity Credit Cost? Nobody Knows — Including Google.

I hit Antigravity's free tier on a Friday morning, about two hours into trying it out. Fair enough — it's the free tier, it's meant to run out. The interesting bit came next.

I tried to find out what I'd spent.

Not in pounds. In anything. Tokens, requests, agent turns, model calls. Anything that would tell me what the two hours had actually consumed and what each prompt had cost me.

What Google publishes

Credits exist. Credits cost $0.01 each. You can buy 20,000 of them for $199. Subscriptions include "built-in credits."

That's it. That's the documentation.

What Google does not publish

How many tokens one credit buys. How many credits a Claude Opus request costs versus a Gemini Flash request. How many built-in credits the $20 Pro plan or $100 Ultra plan actually include. Whether reading a 5,000-line file counts as one credit or fifty.

Their own docs include the disclaimer: "Specified rate limits are not guaranteed."

So the unit you're billed in has no documented exchange rate to anything you can measure. You can buy more of them, but you can't tell what you'd be buying.

This isn't just Antigravity

It's how subscription AI works now. The model providers are doing the same thing. Claude Max, ChatGPT Pro, Cursor's various tiers — all of them moved off transparent per-token pricing the moment they figured out users would pay flat fees for vague "high" or "generous" quotas. Opacity is the business model.

It's not malicious. It's just optimal pricing strategy when your customers can't measure what they're getting. Vague tier names ("Pro", "Ultra", "Max") with vague benefits ("higher limits", "priority access") are easier to raise prices on than published per-unit costs.

The bit that should worry your finance team

I'm a founder running multiple products, so I notice when I can't account for £20. Most people I talk to are running three to five AI subscriptions in parallel without thinking about it:

• ChatGPT Pro for general writing — about £16/month
• Claude Pro or Max for code review — £15 to £100/month
• Cursor or Antigravity for the IDE — £16 to £100/month
• An OpenAI or Anthropic API key for one custom integration — pay-as-you-go
• A team Notion AI add-on that came bundled with something else — £8 per seat

That's a £55 to £250 per-developer footprint per month, and nobody on the finance side has a single number for it. Worse, nobody has any idea what each subscription is actually doing for them. The IDE plan might be cannibalising the API spend. The Pro plan might be paying for capacity the team rarely uses. The team Notion AI might be unused entirely.

SaaS sprawl was the last decade's version of this. AI subscription sprawl is the next one, and it's growing faster because the marginal cost of adding "one more AI tool" feels invisible. Until you add them up.

Governance starts with what you can see

Here's the practical bit. With API-based AI tools — the ones where you bring your own key — you can route through a proxy and see everything. Every token, every model, every prompt, by team and by project. That's a solved problem.

With closed subscription tools like Antigravity, Cursor, ChatGPT Pro, and Claude Max, you can't. Google owns the routing. Anthropic owns the routing. OpenAI owns the routing. The most you can do is track the subscription line items in your accounting system and hope someone is using them.

So the practical advice is:

Default to BYOK where you can. Direct API access plus a governance layer beats a bundled subscription every time, if your team is technical enough to manage keys. You see the bill. You see the per-task cost. You can attribute spend back to the work.

Audit the closed subscriptions twice a year. Whoever owns each one needs to answer: how much did we use it, what did it produce, would we miss it if it went away? If the answer's vague, kill it.

Watch the stack-up. If a developer ends up with three AI tools running in parallel that overlap in what they do, you're paying three times for the same capability. Pick one.

The compliance angle I didn't expect

There's an edge to this I hadn't really considered until I hit the wall on Friday morning. When EU AI Act enforcement lands on 2 August 2026 — about 65 days from now — businesses need an audit trail of what AI did in their operations.

For the API-based stuff that's straightforward. You log every call. You know which model produced what output. You can answer the regulator's question.

For closed subscription tools where the provider holds all the telemetry, you have a gap. You know your developer used Antigravity. You don't know what it touched, what it read, or what it produced. That's not a problem today. It might be one in six months.

Where this leaves you

I'm still using Antigravity. It's a genuinely good IDE and the multi-model picker is excellent — switching between Claude Opus, Gemini 3 Pro, and Flash inside the same conversation is a real productivity unlock. I just won't pretend I understand what I'm paying for.

If you're noticing the same thing — that your AI bill is creeping up and nobody can really account for it — start with the half of the stack where you can see the spend. That's the API channel: every call your developers make to OpenAI, Anthropic, Google, Mistral, and the rest. Make that visible first. Then turn your attention to the closed subscriptions and ask whether you're getting your money's worth.

For the rest, all I've got is the same advice as everyone else: read the line items, ask hard questions, and don't trust opaque pricing to stay friendly forever.