Price the merge, not the token

Bill the unit the buyer already counts

Sourcegraph shipped Agentic Batch Changes (opens in a new tab) to general availability on Tuesday with one pricing rule. You pay per changeset merged into your codebase. Not per token, not per seat, not per attempt. Open a pull request the team rejects and the bill is zero. Beta customers merged more than a thousand changesets; Canva merged 50-plus in one library migration.

Why now? Because the token stopped being a price you own. Vercel’s gateway index put the average price per token down 23.2% in August (opens in a new tab), the third monthly drop in a row, with open-weight models at 56% of tokens but 14% of spend (Vercel’s own traffic). And UC Berkeley’s HarnessTax study (opens in a new tab) ran one model, GPT-5.6 Sol, through three coding-agent harnesses: $1.540 per resolved task on Claude Code, $0.441 on Pi, and no quality difference that survived correction across 42 comparisons. A token meter is a pass-through of someone else’s price list, and the harness on top of it moves the number 71% by itself.

Am I affected? If your pricing page has a credit, a token or a compute unit on it, yes. What replaces it? The thing the buyer was going to count anyway. Intercom’s Fin costs $0.99 per outcome (opens in a new tab), with the outcome defined on the page. Atlassian put AI agent resolutions on a meter on September 1. A merged pull request is the coding version.

Here is the catch, and it is a marketing one. Sourcegraph’s rate is not on its pricing page (opens in a new tab). Enterprise starts at $16K a year and “includes credits for AI features”. The per-merge price lives behind sales. An outcome price you cannot read is a token meter with better copy.

The trade-off is real. Outcome pricing means you eat every failed attempt, and the customer decides what “merged” means. A team that merges three of ten proposals pays for three while you paid for ten. The rate has to carry that ratio, which is why Sourcegraph can afford it on an enterprise contract and a $20 seat cannot.

Your move:

• Find the unit your buyer reports upward: merged PRs, resolved tickets, deployed apps. Price one thing on it this quarter, even as an add-on.
• Put the rate and the definition of the unit in the same sentence on the pricing page. Say what a failed attempt costs: zero, or the number.
• If you keep the token meter, publish the harness you run and the model you default to. The buyer now knows those move the bill more than the model does.

Takeaway: a diff is effort. A merge is an outcome. Price what the buyer would have counted without you.

Quick hits

• GitHub rewrote Copilot’s runtime in Rust. (opens in a new tab)
  About 430,000 lines of TypeScript became 832,000 of Rust in 14.5 weeks, mostly one developer using Copilot, with 158 unsafe blocks and the regressions both listed by category. The number that cuts against the story is what makes “wasn’t affordable before agents” a claim you can check rather than a slogan.
• Cloudflare split “train on me” from “find me”. (opens in a new tab)
  A “Disallow AI Training” setting keeps a site in AI search while blocking training, and Apple, Google, Microsoft, Amazon, Anthropic, Meta and OpenAI already run the two crawlers apart. It is off by default on every plan, so the accidental citation block is now one checkbox you have to go and set.

Open questions

1. Does Sourcegraph publish the per-merge price by October 31?
   Open sourcegraph.com/pricing (opens in a new tab) and look for a dollar figure per merged changeset on the public page: a number makes outcome pricing a pricing-page claim, and “includes credits” left as it is means a sales motion wearing the words.
2. Does the first Copilot bill-shock post land by October 31?
   Every Copilot Business and Enterprise seat is billed up front (opens in a new tab) from October 1, so search GitHub Community and Hacker News for “Copilot bill”: one public story means GitHub shipped the meter without the cap Atlassian shipped first, and a quiet month means the meter was explained.