1. Put the signup after the first success, and the pitch on the claim page

The quarter’s one movement took four launches to show its shape. Railway let anyone deploy without an account (opens in a new tab) on August 17: 60 minutes to build, 24 hours to claim, a $3 cap on LLM credit. Neon shipped Claimable Neon (opens in a new tab) on September 10: an agent reads neon.com/auth.md, runs neon claim create, and gets a Postgres project with no human account, capped at 100 MB, 1 GB of transfer and 72 hours. Railway followed with ssh railway.new (opens in a new tab) on September 25. And on October 2 Neon resized its free plan (opens in a new tab) to 1 GB per project across 100 projects, because it is “creating new projects at a rate of more than one per second”.

Each week filed one of those as a feature. Together they are a new shape for the free tier: not a signup, but a cap, a clock and a claim.

Why now? Because the first reader changed. Across the docs Mintlify hosts, agents were 66% of July’s traffic (opens in a new tab), 213 million agent requests against 105 million human page loads, up from 15.2% at the start of the year (Mintlify’s own hosts). Stack Overflow’s 2026 survey (opens in a new tab), published this week from 30,903 developers, has 70% asking an AI agent when they are stuck and 53% learning from AI tools. The agent reads your docs first, and now it can build on you before anyone has typed an email address.

Here is the strongest case against the quarter’s thesis, and it is a good one. The buyer did not change. In G2’s July survey of more than 1,000 B2B software buyers (opens in a new tab), 9% would let an agent execute a purchase within guardrails and 2% without pre-approval. And the developer behind the agent is not sold on it: in the Stack Overflow survey 6 in 10 cite a reason to avoid AI (opens in a new tab) at work, and 48% trust it only when they can check the answer themselves. So the agent-as-buyer, the story W28 opened in July, did not arrive. What arrived is narrower and more useful: the agent runs the trial, and a skeptical human signs.

That is exactly why the claim step is the right shape and the agent checkout is not. Neon says why it chose anonymity over an allowlist: “Which agents discover the file? What do they request? How far do they get?” A signup form counts people. A claim counts agents, and then counts the people who kept what the agent built. It is the first activation event a machine reader produces that a vendor can report.

Am I affected? If an agent can call your API and step one is “create an account”, yes. What does the agent get for free? Not a card. Three numbers and a deadline. What does the human get? A page that says what was built, what it costs after the clock, and why to keep it. That page is where your buyer meets you for the first time, and today it is a login form.

The trade-off is real. Anonymous provisioning is an abuse surface, and the caps are the abuse control. Railway’s $3 cap is a pricing decision written as a security one. Make yours the same way, and publish it.

Your move:

  • Write the unclaimed tier as three numbers on the pricing page: storage, transfer, hours. Add the dollar cap on anything metered inside it.
  • Publish what an agent may do with no key. /auth.md if you can, a paragraph in llms.txt if you cannot.
  • Treat the claim page as a landing page. What the agent built, what it costs from tomorrow, in plain text. Count discovered, created and claimed as one funnel, and publish the claim rate.

Takeaway: for a human the signup was the front door. For an agent it is the exit. The page at the exit is your new home page.

2. Price the unit the buyer counts, and ship the cap a quarter before the meter

The pricing thread moved three times in thirteen weeks, and each move was away from the token. On July 20 GitHub priced Code Quality (opens in a new tab) at $10 per active committer plus a meter for the AI parts, the seat-plus-meter shape the quarter started with. On September 1 Atlassian metered automation steps (opens in a new tab), a unit with no model behind it: 400 a month per Jira Standard user included, $0.50 per 1,000 above that (opens in a new tab), billing from December 3, with the usage view and the admin limits shipped the day of the announcement. On September 16 Sourcegraph billed Agentic Batch Changes per merged changeset (opens in a new tab), nothing for a diff the team rejects, and Canva merged 50-plus pull requests in the beta.

Then the model vendors stopped quoting the rate. Anthropic cut Opus 5.5 to $4 and $20 per million tokens (opens in a new tab) from $5 and $25, and led with “40% less on typical workloads”. OpenAI halved GPT-6 Sol to $2 and $10 (opens in a new tab), confirmed permanent, and pitched $0.27 per task on one benchmark. DigitalOcean wrapped raw agent meters in plans (opens in a new tab): $5 of credit with no card, $50 a month at 15% off usage, $200 at 20% off, “no seat charges”.

Why did the rate stop carrying the claim? Because it stopped holding still. Vercel’s gateway put the average price per token down 23.2% in August (opens in a new tab), the third monthly fall in a row, with open-weight models at 56% of tokens and 14% of spend (Vercel’s own traffic). And HarnessTax (opens in a new tab), from UC Berkeley and Arena, ran seven models through three coding-agent harnesses and found the same model costing up to five times as much per resolved task while success rates moved by about two points. A token rate is someone else’s price list, multiplied by a harness your customer did not choose.

Am I affected? If your pricing page carries a token, a credit or a compute unit, yes. What replaces it? The unit your buyer already reports upward. Intercom’s Fin is $0.99 per outcome (opens in a new tab), with the outcome defined on the page in three lines and one charge per conversation. A merged pull request is the coding version. A resolved ticket is Atlassian’s.

Two trade-offs, and the second is the one the quarter exposed. Outcome pricing means you eat every failed attempt while the customer decides what “merged” means, so the rate has to carry the merge ratio. That suits an enterprise contract and not a $20 seat, which is why Sourcegraph’s number is not on its pricing page (opens in a new tab): “Starting at $16K. Includes credits for AI features.” It was not there in September and it is not there today. An outcome price a buyer cannot read is a token meter with better copy.

The second trade-off is sequence. Atlassian shipped the limit three months before the bill. GitHub shipped the bill first: every Copilot Business and Enterprise seat has been prepaid since October 1 (opens in a new tab), with the promotional credits that hid the meter gone a month earlier and prices unchanged. The first week was quiet. One public bill-shock post by the end of October decides which order was honest.

Your move:

  • Pick the unit your buyer reports upward, merged PRs, resolved tickets, deployed apps, and price one SKU on it by year end, with the rate and the definition of the unit in the same sentence.
  • Put one worked bill on the pricing page: what your quickstart’s first hour costs at list, with the workload written out. Restate every rate cut as its effect on that bill.
  • If anything you sell will meter in the first half of 2027, ship the usage page and the admin limit this quarter, and date the bill.

Takeaway: a rate is the vendor’s number. A bill is the buyer’s. The cap is the promise that makes the meter a price instead of a surprise.

3. We told you to ship llms.txt three times. Here is what it bought

This is the call the quarter corrected, and the corrected read changes what you do with a week of someone’s time. W28 put llms.txt beside OpenAPI as a machine-readable surface to publish. W31 narrowed that to “an afternoon’s insurance for the agent surface” after Google said in writing that it “doesn’t use them” (opens in a new tab). W35 kept the file and threw out the proof, after a satirical cats.txt collected every citation a vendor shows you. Three issues, three narrowings, and none of them said what the file is for.

The quarter’s last fortnight answered that. Gauge’s instrumented sessions opened a site’s llms.txt in 36.3% of build tasks and 0.5% of vendor-selection tasks (opens in a new tab) (Gauge sells agent analytics and states no session count). A week later its crawler logs showed that of roughly 600,000 requests carrying the ChatGPT-User agent across more than 500 million bot events, none asked for the file (opens in a new tab) (no period or site count stated). The crawler that produces citations never reads it. The agent that already chose you does.

And the citation itself has a clock. Profound followed 883,000 pages through seven engines for a year: the median page loses half its citation share within 11 days (opens in a new tab) of its peak, 78% do so within two weeks, and the engines barely agree with each other, with correlations of 0.03 to 0.09 (Profound sells the tracker). Pages cited by five or more engines were 5.5 times as likely to still be above half their peak after eight weeks.

So what should the July move have been? The quarter says: the docs page that explains your bill. Profound asked six engines 7,600 pricing questions (opens in a new tab) about Cloud 100 companies. The company’s own pricing page came first in 12% of answers. Vendr came first in 18.7%, Reddit in 18.6%, G2 in 15.9%. Only 57 of 77 public pricing pages were fully readable to a bot. The exception was Plaid, whose billing docs were cited in 70% of its answers against 64% for its pricing page, and those docs publish no prices. They define the billing model and say which numbers are not public. The same firm’s August study of 158,000 brand claims (opens in a new tab) found 54% of brands with an inaccurate claim citing their own site, and pricing was 12% of claims and 24% of the inaccuracies. The engine quotes your forgotten page with confidence.

Am I affected? If your plan for AI answers is a file and one content sprint, yes. The file serves one reader at one step. The pitch belongs on the pages the engines cite, and those pages need renewing.

The counter-case is that nobody has a conversion number for any of this, and that is still true. Val Town’s July datum (opens in a new tab), most new Pro signups with a known source arriving via AI, mostly Claude, is the only attributed figure any vendor published all quarter, and it is a self-reported share, not a rate. Which is the point: fund the pile you can attribute, and treat the rest as hygiene.

Your move:

  • Strip the pitch out of llms.txt. Make it a routing index for an agent already building on you: the quickstart, the reference, the billing page.
  • Write the billing page in the docs. Plain HTML, one question per heading, every public number in the text, one plain line for each number that is not public. No tabs, nothing rendered client-side.
  • Put a refresh date on every page that carries a citation and republish its facts monthly. Run the free check on all seven engines, and protect the pages that show up on five.

Takeaway: the file serves the agent that already picked you. The citation is a lease. The docs page that explains your bill is what gets read at the pick.

4. Proof is a launch format now: a date, a number and the script

In July proof meant a dogfooding post. By October it had a shape with three parts, and the fourth vendor to use it shipped the part the first three did not.

Vercel put up to $1 million (opens in a new tab) on breaking its sandbox, two weeks on HackerOne, $50,000 cap per report. The tally reached the press on September 15: 1,285 reports (opens in a new tab), one critical, seven high, 15 medium, 49 low, 19 informative, about $325,000 committed, and the two most serious bugs in the Linux kernel’s networking stack rather than Vercel’s code. Docker rewrote its homepage to “Trust Docker for the Agents You Don’t”, still the hero today, and shipped Cloud Sandboxes billed by the second (opens in a new tab) with its Kit spec under Apache 2.0 and a promise to the CNCF. Postman ran its own monitor against a coding agent on one laptop from August 12 to September 9 and published the count against itself: 22 secrets across 33 requests to 18 hosts (opens in a new tab), while the project’s .env held four.

Then GitHub shipped the part the others left out. Its security agent found 24 Android vulnerabilities (opens in a new tab), and the post includes the taskflows, one command to run them on your own repository (./scripts/audit/run_mobile.sh myorg/myrepo), an estimate of an hour or two on a medium project, and the cost: a Copilot licence and a lot of premium requests. The receipt ships with the script.

Why does this format work on this audience? Because the audience validates by habit. Stack Overflow’s survey has 48% trusting AI only when they can check the answer. A number a reader can rerun is the only claim that survives that reader. A number without a method is an adjective with digits, and the quarter produced those too: a “97% of users pay us $0” with no denominator, a “48% fewer tokens” resting on one task run twice.

The trade-off is that proof is necessary and not sufficient. Reputation filled the room for Block’s Buzz in July and the room still said no. And a wager is only half a campaign until the result ships: Vercel’s program page promised a write-up of the attack techniques, the CVEs are pending, and as of today that post does not exist.

Your move:

  • Put one proof asset on next quarter’s launch calendar with a date and a number: a pen test, a postmortem, a migration with its costs named.
  • Ship the method with it. The script, the rubric or the sample, so a reader can rerun the number instead of taking it.
  • Print the number that cuts against you beside the one that flatters, with your reading of why it is fine.

Takeaway: a claim that costs you something reads as proof. A claim a reader can rerun reads as fact.

Quick hits

Watch

  1. A claim rate is published by December 31. Neon or Railway publish discovered, created and claimed for the anonymous tier, or a third vendor ships a claim-later tier, checkable by fetching /auth.md (opens in a new tab) from Supabase, PlanetScale and Vercel by October 31 as W37 asked. A rate confirms the agent free tier as a category and gives the referral loop its second datum. No number by year end means two vendors ran an experiment, and the loop closes.
  2. The outcome unit reaches a public price by December 31. A dollar figure per merged changeset on sourcegraph.com/pricing (opens in a new tab), or a second coding agent billing per merge. Either makes outcome pricing a pricing-page claim. Silence leaves it a sales motion wearing the words, and a quiet October for Copilot’s prepaid seats would say the meter was explained.
  3. Vercel publishes the techniques write-up by November 30. The program page (opens in a new tab) promised it, the kernel fixes are under private review, and the CVEs are pending. A published report completes the campaign and sets the bar for every attack-me launch after it. Nothing by the end of November means the wager was the campaign.

The quarter in issues

The 12 issues this quarter is made of, each with its own record of the window it covers.

  • 2026-W39
    Quote the bill, not the rate

    Anthropic and OpenAI both cut list prices on Tuesday, and neither led with the rate. The headline was the bill on a typical workload and the cost per task, because a rate that halves twice a quarter is no longer a number a buyer can hold you to.

    21–27 Sept 2026
  • 2026-W38
    Price the merge, not the token

    Sourcegraph now bills its coding agent per pull request that merges, and charges nothing for a diff the team rejects. The token meter it replaces prices a unit whose cost fell 23% last month, so the outcome is the only number left that means the same thing to you and the buyer.

    14–20 Sept 2026
  • 2026-W37
    Your free tier now has a claim step

    An agent can now build on your product before anyone signs up. Neon lets one create a Postgres database with no account, and the signup moves to the end, where a person claims what the agent built.

    7–13 Sept 2026
  • 2026-W36
    Your docs are production now

    Developers read your docs before your marketing. Now agents do too, and they run what they read. Three things that happened this week, and what to do about each.

    31 Aug – 6 Sept 2026
  • 2026-W35
    The agent's front door is a tool, not a file

    A satirical cats.txt passed every proof the industry cites for llms.txt, the same week Neon and Postman shipped free, callable entry points for agents. Spend your effort, and your evidence standard, on the side you can count.

    24–30 Aug 2026
  • 2026-W34
    The trust surface is the campaign now

    LeadDev's buyer research put numbers on developer skepticism the same week Docker ran a four-post trust campaign, Vercel paid to be hacked in public, and GitHub printed its own outage numbers. The vendors who reach skeptical buyers are marketing proof, not promises, and the human gates are coming down behind them.

    17–23 Aug 2026
  • 2026-W33
    The rail is the product now

    SpaceX closed the largest devtools acquisition on record, GitHub and MongoDB shipped agent rails, Meta started indexing the web for itself, and Vercel's own data showed model loyalty is a one-line change. The durable asset in the agent era is the rail, not the model.

    10–16 Aug 2026
  • 2026-W32
    The coding agent became a commodity, now the terms are the pitch

    Three terminal coding agents launched in forty-eight hours and none competed on capability. Meta competed on a price that is really a data trade. When features converge, the terms sheet becomes the positioning surface, and this week priced it.

    3–9 Aug 2026
  • 2026-W31
    Twelve months or six weeks, the deprecation window is the positioning now

    MCP's biggest spec revision started a twelve-month migration clock under every MCP server the same week GitHub retired Models on a six-week runway with no like-for-like replacement. How you end things is becoming as much of a marketing surface as how you launch them.

    27 Jul – 2 Aug 2026
  • 2026-W30
    The AI seat gets a price, and the empty ones get counted

    GitHub priced AI code review as a seat plus a meter on Monday, then shipped the dashboard that counts unused Copilot seats on Wednesday. The vendor selling AI seats is now arming buyers to find the shelfware, because the ROI-skeptical renewal conversation demands it.

    20–26 Jul 2026
  • 2026-W29
    Incumbents start shipping for the agent-as-buyer

    GoDaddy and Atlassian both shipped product for the agent-as-user in the same week, so the thread graduated from tiny Show HNs to platform roadmaps, while Cerebras showed how quietly sunsetting a free tier burns trust it took years to earn.

    13–19 Jul 2026
  • 2026-W28
    The agent reading your docs is starting to shop

    Machine-mediated discovery moved a step this week, from AI assistants reading your docs to agents comparing and picking vendors, while a zero-budget open-source launch topped Hacker News on reputation alone.

    6–12 Jul 2026

The patterns of the quarter

Pattern · 6