I read the billing page before it applied to me. Salesforce had posted a transition notice for Agentforce Vibes, and buried in it was the line that mattered: after the first of June 2026, free limited access ends for orgs that aren’t Developer Edition. I read it, worked out what it meant for mine, and closed the tab. Nothing dramatic happened that day. I just knew the door would shut, and roughly when.

Illustration of a person standing before a closed shutter with a thin orange opening.
The Process Is Closed / by GPT

Which makes what I did next harder to defend. I had the warning, I understood the warning, and I still didn’t do the one thing it should have prompted me to do: measure what I was using while I could still use it.

Under An Hour

The free ceiling was 110 requests, or 1.5 million tokens, per month.

In writing, that sounds like a fair trial, and for ordinary use it probably is. You ask a model to write a class, it writes a class, and you get a hundred and ten of those before anyone asks you for money.

Agentic work doesn’t spend like that. I was reshaping a nested Salesforce structure, several object nodes deep, into the output format an OmniStudio DataRaptor expects. That work has the agent reading the source shape, reading the target contract, and reconciling the two, which is a great deal of looking before any of it becomes a line of code. With plan mode running and subagents spawning, a single instruction from me isn’t a single request. It’s a fan-out, and every branch of it spends from the same bucket while I’m deliberately not watching, because not watching each step is the entire point of the pattern.

I burned the month in under an hour. By the time I thought to check the counter, there was nothing left to check it with.

A transparent cylinder outlined in black, with an orange mark near one end.
The Unreached Measure / by GPT

That hour is the only measurement I own, and it’s crude. I don’t even know which ceiling I hit first, the request count or the token count, and that distinction matters enormously to anyone trying to plan around it. So here’s the claim I’d defend: you can’t evaluate agentic Salesforce tooling on your own any more, because the trial ends before the measurement does. Not because the tool is bad. Because the window is shorter than the thing you’re trying to measure.

The obvious objection is Developer Edition, which stayed free and still works, and it’s a fair objection because I run one. But a DE org isn’t the one you’re evaluating. It has no real data volume, no accumulated automation, and no metadata sprawl that makes agentic work expensive in the first place. What you measure there tells you something about the tool and almost nothing about the bill.

Page 255

In August, the Winter ‘27 release notes arrived. On page 255, inside a section called Headless 360, Salesforce documented a new reference topic for estimating Flex Credit consumption before committing to metered billing. It covers Prompt Usage Types and how the multipliers work. It carries measured token data for Opus 4.8, Sonnet 4.6 and GPT-5.4 from a single-session example. It includes a monthly estimator you can feed team size into.

That’s the exact document I needed in May, published in September, on the far side of the paywall I’d already fallen out of.

I want to be fair because publishing a consumption estimator provides more transparency than most vendors offer for agentic tooling, and the release is still in preview. Hence, none of it is generally available. But look at the sequencing. The cost model shows up once you’re a paying customer with a team to size, not while you’re deciding whether to become one. That’s the opposite of the order a buyer actually needs it in.

The Rate Isn’t The Cost

Suppose you decide to pay anyway. You still can’t predict the bill, and that's worth getting right because a lot of people have it wrong.

Salesforce’s pricing page states that using different AI models does not affect Flex Credit or Flat Fee pricing. People read that and conclude model choice is free. The sentence is doing something more precise than that. What it means is the rate doesn’t move: a credit costs half a cent whether you run Sonnet or Opus, because Flex Credits sell at five hundred dollars per hundred thousand. What it doesn’t mean is that the same task costs the same.

Opus charges no premium per unit. It spends more units getting to the same place, because it thinks for longer, reads more, and produces more tokens on the way. The price per unit is flat and the unit count isn’t, and your invoice is the product of the two. A flat rate on a variable meter still results in a variable bill.

There’s a second option that makes the same point from the other direction. Salesforce sells flat-fee access at $125 per user per month and markets it as a predictable monthly cost. Think about what that tier actually is. It isn’t a discount and it isn’t better models, since both tiers run the same models. You’re buying out the variance, and a vendor only sells insurance against a number when that number is hard enough to forecast that customers will pay to stop trying.

My Own Opinion Fails The Test

I moved to Claude Code, and I think it’s better than Agentforce Vibes for my work.

Now watch me fail my own argument. My comparison runs against the Vibes I used before June 2026, and Salesforce has since described Vibes 4.0 as a full re-engineering on their coding agent platform, rebuilt with Claude handling reasoning and Mastra handling orchestration, shipping plan mode, subagents, custom skills and MCP support. That isn’t a patch on the thing I formed an opinion about. It’s a different product wearing the same name.

So my judgment is stale, and it’s stale for exactly the reason this piece is about. I can’t refresh it. I’m holding a comparison I have no way to retest, which makes me the clearest example I can offer of the problem: a developer with a strong view about a tool, locked out of the only evidence that could correct it.

If you have access to 4.0, your opinion on this is worth more than mine. Mine is a photograph of a version that no longer exists.

A sketch of a building on a curled sheet of paper held by an orange pin.
How It Used to Work / by GPT

What You Can Still Own

One more thing shipped on that same page, and I’ll keep it short because it’s a different argument than the one I’m making. Agentforce Vibes now applies AI governance policies configured in ALM Control Center. Admins set them, Vibes fetches them for every session, and developers can’t modify or turn them off. I think that’s correct design; an enterprise needs somewhere to say what its agents may do. It does mean that by the time you can afford to evaluate the tool, you’re inside an org that has already decided how you’ll use it.

I’ve been building a compound engineering loop since December 2024, and it kept running through June because it doesn’t depend on anyone’s meter. That isn’t a victory lap. The loop is worse than Salesforce’s own plugin at knowing my org, and I’ve said that publicly because they own the compiler and I don’t. On the thing they built, they win.

What the loop gives me is the thing I lost everywhere else: I can measure it. I can run it a hundred times this week and change my mind about it on the evidence.

So if you’re weighing agentic Salesforce tooling for a team, price it before you pilot it, because the pilot won’t survive first contact with real agentic work; an hour of it, in my case. Developer Edition is the only free path left, and it’s worth setting one up purely as an instrument. And whatever else you build on, keep one part of the stack you can still test whenever you want, because an opinion you can’t refresh stops being knowledge fairly quickly. Mine did.

Thanks for reading today’s piece, and to continue reading more, don’t forget to subscribe and share with your mates who might find it useful. Cheers, Sangamesh!