Pricing is positioning with a number attached, and for a developer audience it is read before the second paragraph of the homepage. “Contact sales” reads as something to hide; a legible price is a trust signal. This section is about the decision - what is free, what is metered, what a seat buys, where the enterprise door is - and about the page that decision lands on. The website section covers the page’s place in the layout; positioning covers the claim the number has to be consistent with.

The free tier is the motion

For most developer products the free tier or the open-source core is the top of the funnel, and self-serve is the default motion: a developer goes from evaluating to building to paying without talking to anyone, and the moment they should talk to sales is obvious and well-timed rather than forced early. Three things follow.

  • Free is a product decision, not a discount. The tier has to reach real value - a deployed app, a production-shaped workload at small scale - or it is a demo with a signup form. Railway’s accountless deploys (August 2026: build first, claim the project within 60 minutes) are the logical end state: value before identity, let alone before payment - and Railway widened it on 2026-09-25 to the machine itself, a free 2 vCPU VM from ssh railway.new with 60 minutes to build and 24 hours to claim before it is deleted. Neon extended it to the agent in September 2026 - Claimable Neon provisions a Postgres project for a coding agent with no account, capped at 100 MB, 1 GB of transfer and 72 hours, for a human to claim later - so the unclaimed tier is now a pricing line written in three numbers, and the caps are the abuse control as much as the offer. The same reader is now resizing the claimed tier too: Neon doubled free-plan storage to 1 GB per project on 2026-10-02 with the 100-project cap unchanged (and a 20 GB account total enforced independently, so the two limits are read together), and said why - developers are “building more than ever with coding agents, a lot of it on free plans”, wanting “lots of projects, lots of experiments, most of them idle most of the time”. A free tier sized for one human prototype is the wrong shape for an agent that creates dozens; size it for the reader that actually arrives.
  • Free is a contract. Developers build on it literally, and changing it is a deprecation event with a date, an email and a migration path - developer experience has the benchmark windows and the brownout pattern. A free tier you cannot afford to keep is cheaper to shrink loudly than to keep silently. The same holds for any term a paid plan included. PostHog cut default event retention on new paid plans from seven years to two on 2026-11-05, kept seven for existing customers “while it stays on its current plan and platform package”, sold the longer windows back through its packages (seven only with Enterprise, $2,000 a month) and gave cost as the reason. The grandfathering is fair and has one sharp edge: a downgrade now carries a deletion (“we can’t restore deleted events”), so the customer cutting its bill can lose five years of history. Grandfather the old term, and put what a downgrade costs in the downgrade flow itself, not only in the docs; PostHog’s public marketing tracker shows it chose a targeted email on downgrade and another 14 days before deletion.
  • Free is the audience you are sizing. The hobbyist segment that made free tiers “free marketing” is contracting - the audience section has the population numbers - so cost the tier against the professional developers it converts, not the signups it collects.

Whether the core is open source is the same decision with a licence attached, and it is read as one: the licence, the governance and what stays free for how long are positioning statements with legal force. The repo covers how developers read them.

Seat, meter, or both

The market settled the shape while the guide was being written. Kyle Poyar’s 2026 State of B2B Monetization survey (230 companies, April to May 2026; 43% of them hybrid SaaS-plus-AI products) found hybrid pricing - a base subscription with a metered component on top - the single most common primary model at 37%, ahead of pure per-seat and pure usage. AI credits are the fastest-moving piece of that: 29% of companies already price with them, a further 33% plan to within six to twelve months, and among companies above $50M ARR roughly one in two plan to introduce them this year, after credit models grew 126% in 2025. Developer tools converged earlier and harder. By mid-2026 the AI-assisted feature had one shape from code generation to code review: a per-seat licence plus metered model consumption - Copilot’s seats with draining AI credits, GitHub Code Quality’s $10 per active committer plus usage, CodeRabbit’s seat plus per-file usage billing.

The shape carries three obligations for the page.

  • Publish both parts. The seat is the floor, not the price. A meter the buyer discovers on the invoice is a trust debt you chose.
  • Publish the cap. The runaway-bill fear is what kills usage-based signups, and it has to be answered in the same eyeline as the meter, not in a FAQ. Supabase’s pricing page is the reference implementation: fixed and variable cost separated, the meter linked to a dashboard, and a spend cap on by default. Atlassian’s September 2026 meter expansion is the sequencing to copy: two new meters (automation steps and AI agent resolutions) announced 2026-09-01, admin usage views and limit controls shipped at once, the allowance (400 steps per Jira Standard user a month) and the overage ($0.50 per 1,000 steps) published in the support docs, and the first bill dated 2026-12-03 - the cap a quarter ahead of the meter. Automation steps are not an AI feature, which says the meter has become a billing habit rather than a pass-through of model cost.
  • Publish the number at volume. The buyer’s first question is what this costs at scale; a page that only shows the entry price makes the champion do the arithmetic in a meeting you are not in.

Pure per-seat is the wrong metric when one engineer uses the product all day and another once a month, and pure usage is the wrong metric when the buyer needs a number they can budget. The hybrid won because it answers both, and the honest version of it is the one where the reader can predict the invoice.

The meter’s unit is moving from the token to the outcome. By September 2026 the token had stopped being a price a vendor owns: Vercel’s gateway index put the average price per token down 23.2% in August, the third monthly fall in a row (Vercel’s own traffic, not a panel), and UC Berkeley and Arena’s HarnessTax study (September 2026) ran seven models through three coding-agent harnesses and found the harness alone moved what the same task cost by multiples - up to five times at the widest gap - while success rates moved by a few points (the study’s page is script-rendered and every secondary this desk has read quotes its per-task dollar figures differently, so take the ratio and not the decimals) - so a token meter passes through someone else’s price list, and the harness on top of it can move the bill more than the model does. The answer arriving is to bill the unit the buyer already counts. Intercom’s Fin charges $0.99 per outcome, with the outcome defined on the pricing page; Atlassian’s September 2026 meters include AI agent resolutions; and Sourcegraph’s Agentic Batch Changes reached general availability on 2026-09-14 (on Sourcegraph Cloud; self-hosted followed with 8.0 on 09-16) billed per changeset merged into the codebase, with a rejected diff costing nothing. Two obligations come with the unit. Its definition belongs in the same sentence as the price, because the customer controls it - a team that merges three of ten proposals pays for three while the vendor paid for ten, and the rate has to carry that ratio, which is why the model suits an enterprise contract better than a $20 seat. And the rate has to be public to count as pricing: Sourcegraph’s is not - its pricing page lists Enterprise “Starting at $16K” on a “Minimum annual contract” that “Includes credits for AI features” and names no per-changeset rate - which leaves outcome pricing a sales motion described in outcome language until the number appears.

The number quoted moved with the unit. When the rate falls faster than a pricing page is reread, the vendors stop leading with it. Anthropic’s Opus 5.5 (2026-09-22) cut list 20% to $4/$20 per million tokens, but the sentence it wrote was “at default settings it will cost 40% less than Opus 5 on typical workloads”, with cache reads down 60% and every benchmark table stated as cost per task; OpenAI’s GPT-6 Sol and Luna, the same day, halved the GPT-5.6 rates ($2/$10 and $0.10/$0.50, confirmed permanent rather than promotional) and pitched $0.27 per task and “roughly 80% lower cost per task” on its own benchmark runs; OpenRouter added a batch tier at about half price for a 24-hour window. The bill for a defined job is the claim now, because it is the number the buyer budgets and remembers - and it is also the number the vendor gets to define, since “typical workloads” appears on no pricing page. The honest form publishes the workload with the figure: one worked example of what the quickstart’s first hour costs at list, beside the rate. The product-side version of the same move is to show the bill per unit of work rather than as an aggregate: Vercel now lists billable duration and CPU minutes on every deployment (dashboard, vc inspect, REST API; build plus post-build time rounded up to the minute, multiplied by vCPUs), and DigitalOcean’s Managed Agents bill per second of active CPU with the idle waits stated as costing $0. A rate cut is the news; the bill is the claim, and the two should ship together. A cut with no end date is not a term at all: Google announced Gemini 4 Argon (2026-09-30) at an introductory $2/$10 per million tokens rising to $4/$20 “after the introductory period expires”, and for two weeks named no date for that and listed the model on no pricing page - so a buyer could not budget it, and a reader comparing rates was comparing one vendor’s list to another’s promotion. The date arrived with the product: on general availability (2026-10-13) the Gemini API pricing page took the model on, reading “$2.00 through November 30, 2026. $4.00 starting December 1, 2026” for input and the same shape for output and cache, and excluding it from the free tier and the Google AI Pro plan. A promotion becomes a term when it carries both halves; a dated discount with no price after it - Vercel’s AI Gateway listing Fish Audio’s Drama 3 at “80% off for 30 days (through Nov 12)” with no list rate stated, 2026-10-13 - is the same gap from the other side.

The meter and the plan are two pages for two readers. DigitalOcean published both within two days in October 2026: its pricing page lists the raw agent meters - Harness Runtime from $0.0535 an hour, an Action Gateway from $0.10 per 1,000 calls, inference from $0.05 per million tokens, pitched as one bill instead of a harness, a sandbox, a model and a search API stitched together - and Agent Droplets (2026-10-01) wraps the same meters in a subscription: $5 of credit with no card, $50 a month at 15% off usage, $200 at 20% off, with SSO, MFA, role-based access and audit logs at every tier and “no seat charges”. Metronome’s CEO named the progression the same week on Stripe’s blog: token billing belongs in the cost model rather than on the invoice, and pricing moves from tokens to unified credits to a price on the output, with roughly one in six Stripe users past key revenue milestones rolling out hybrid pricing as of August 2026 (Stripe’s own book of business; Metronome sells the billing layer the argument favours). The builder - and the agent - reads the meter and tests you on it; the team budgets the plan. Keep the meter on the page, state the plan’s discount as a percentage off that meter so the two reconcile, and if the security features cost nothing per seat, say so in the plan’s first line. The trade-off is that the discount is the product, and every point of it is margin.

The margin can be a number, and the rate can be a dial. Warp’s Factories pricing (public from 2026-10-07, replacing a closed beta that offered qualifying organisations up to $10,000 of free usage) states its markup outright: Pay as you go at $0 a month with “Usage at 20% above API rates”, Build at $20 a month and Business at $50 a user a month with “Usage at API rates”, Enterprise custom, and factory usage “billed separately based on inference, compute, and agent usage” (whether the 20% covers compute is not said). The subscription is the price of not paying the markup - at list, the $20 tier pays for itself past about $100 a month of model usage - and it is the first time a vendor in this guide’s record has put its resale margin on the page as a figure. The rate underneath is itself becoming a setting: Claude Haiku 5.5 (2026-10-07) lists $0.10/$0.50 per million tokens up to a 100,000-token context, “around 75% less to run” than Haiku 4.5, with an adjustable effort setting from Low to Max so the buyer chooses between cost and intelligence; the page states no per-level price, so the dial moves the tokens spent rather than the rate. A markup on a rate the buyer can turn is a bill nobody can quote in advance, which is one more reason the worked task above belongs beside the rate.

The discount names the model. Four plans in three weeks of October 2026 drew the same line, and it runs through the vendor’s cost base rather than the buyer’s needs. DigitalOcean’s Agent Droplets (2026-10-01) discount only “inference on DigitalOcean-hosted open models” and leave Claude and GPT pay-as-you-go at standard rates. Anthropic’s Max and Team plans (platform release notes 2026-10-07) include $100 to $500 a month of credit on Anthropic’s own API, not Claude Code, no rollover. xAI’s X Subscription (2026-10-12) closed Cursor’s $20 Pro to new signups behind a $30 Premium whose weekly pool covers Grok models only, stated as “4x more usage compared to Plus” with Plus’s allowance given in no unit, while Claude and GPT are “billed at API list prices from the first use” out of prepaid credit - so a Pro user who spent the old plan’s $20 of frontier usage on Claude pays about $50 for the same month, a sum this desk made from Cursor’s own pages because Cursor publishes no worked example. And Vercel’s Token Pass (2026-10-15) sells $80 of monthly credit for $30 that only open-weight models served by a third party can spend, in early access, revocable, and failing on a 403 when the credit is gone. Two obligations follow for the page. Name the models the discount covers and say what happens at the first token of one it does not: the most-liked complaint on Cursor’s forum was not the price but the discovery that “model neutrality” had ended. And state the pool in a unit, because a multiple of an unstated number is what Elena Verna, who leads growth at Lovable and says Lovable prices the same way, calls “arcade economics”. Warp is the one vendor on this record pointing the other way - included usage at API rates with no model restriction, and a markup rather than a discount on its free tier - so hold this as a pattern to watch rather than a rule; it would break on a vendor that only buys frontier models and discounts one. OpenCode Go ($10 a month) shows what stating the pool in a unit looks like when the cost varies by model: it prints a dollar allowance per model, $100 on DeepSeek V4.1 Flash down to $15 on Claude Haiku 5.5 and GPT 6 Luna, and says why - bulk discounts and reserved capacity widen it, and a model it has no deal on gets “a little more than if you paid the model providers directly”.

“Included” has to say whether it means access or usage. When GitHub put Copilot into the $21 Enterprise seat (2026-10-28; at launch only for enterprises working with a GitHub or Microsoft account team, self-serve customers to follow in a staged rollout), it included the feature and none of the usage: “There is no included usage allowance to track or top up”, every chat, code review and cloud-agent task is metered at a cent a credit from the first request, and an empty budget stops all of those while autocomplete keeps running. Atlassian, folding Rovo into every paid cloud plan a year earlier, sold it as coming “at no extra cost” with a small pooled allowance (25 credits a user a month on Jira Standard) at the same cent a credit past it, extra-usage billing from 2026-12-03. Same shape, different word: GitHub’s changelog and pricing page both say “access”. The prepaid version was not a discount either - Copilot Business’s $19 seat was exactly its 1,900 credits - so the change moves the money from the seat to the meter rather than removing it. If an AI feature comes with the seat but its usage does not, say “access”, and print on the same page what stops when the budget does.

The terms that are now priced

Two things joined the price list in 2026 that used to live in a trust centre.

The training-data policy now has a market rate. Meta’s Muse Code launched in August 2026 with a “contributor” tier roughly 12–21x below list in exchange for training on the user’s prompts and completions - as far as this desk can find, the first time a training-data policy carried a public price. Whatever your policy, it belongs on the pricing page rather than in a FAQ: “we don’t train on your code” is a positioning claim with a number attached, and its absence is read as an answer.

Self-serve enterprise is the other. The measure of “enterprise-ready” is shifting from what your support team covers to how much of the setup a customer can complete alone. Auth0’s August 2026 SafetyCulture case study - a self-serve SSO configuration wizard cut SSO support tickets 60% at a 50% self-completion rate (the second figure stated on a podcast, single-sourced) - puts a public number on what removing the human gate is worth at the bottom of the funnel, not just the top. The enterprise tier is still where SSO, compliance and a contract live; the change is that the developer expects to reach the door without a form.

Where self-serve hands off

The motion has three shapes and most companies run two. Product-led: the free tier converts to a card, and sales exists to expand accounts that are already paying. Sales-assisted: the developer adopts, the buyer signs, and a human appears at the moment the account crosses a threshold - seats, spend, a compliance question - that the product itself detects. Open-source-led: the core is free and the paid product is hosting, support or the features a company needs and a hobbyist does not. The choice decides what the rest of this guide weighs: product-led lives on activation, sales-assisted on the buyer’s door, open-source-led on the repo and community.

The handoff is the part most teams get wrong in one of two directions. Too early, and a developer evaluating on a weekend gets a calendar invite; too late, and an account running you in production has nobody to call. The trigger should be a fact the product already knows - the third seat, the first production deploy, the spend cap being raised - and the first message should offer something the developer wants (a higher limit, an architecture review) rather than a discovery call. First-party usage is the signal nobody else can buy, and measurement covers scoring on it.

The page, briefly

Put a real number on it. Show the seat and the meter, the cap beside the meter, the price at volume, one worked task with its workload defined, the training-data policy, and one door for the buyer. Date it. Everything else on the page - the comparison table, the FAQ, the logos - is worth less than those six lines being true. Two September 2026 moves show where the number goes next: PlanetScale publishes its price list as a read-only Postgres database an agent can query in SQL, and PostHog put “97% of users pay us $0” under its signup button with no method given - a free tier stated as a statistic works only if the denominator is published, or the skeptic supplies one.