---
title: How AI features get priced — seat, meter, data, and now the outcome
canonical: https://thebeat.dev/threads/how-ai-features-get-priced/
updated: 2026-09-21
status: open
momentum: rising
opened: 2026-07-22
tags:
  - pricing
  - positioning
  - ai-coding-agents
collection: threads
site: https://thebeat.dev/
license: https://creativecommons.org/licenses/by/4.0/
---

# How AI features get priced — seat, meter, data, and now the outcome

> Per-seat licence plus metered consumption settled in as the devtools default for AI features. Meta priced a discount as a data trade and no incumbent answered it by the August 31 clock, so the data axis reads as one vendor's beta mechanism. The meter widened in September (Atlassian metering automation steps, cap shipped a quarter before the bill) and then its unit moved: Sourcegraph bills Agentic Batch Changes per merged changeset and nothing for a rejected diff, while the token it replaces fell 23.2% in a month on Vercel's gateway and moves 71% by harness alone in UC Berkeley's HarnessTax study. The catch is that Sourcegraph's rate is not public. Two live tests: whether that number appears on the pricing page, and whether Copilot's prepaid seats from October 1 produce the first bill-shock story.

**The question**: Is seat-plus-meter still the devtools default, is training data becoming a third thing developers pay with — and is the meter's unit moving from the token to the outcome?

- **Status**: open · momentum rising
- **Opened**: 2026-07-22 · last worked 2026-09-21
- **Guide**: https://thebeat.dev/guide/pricing-and-packaging/, https://thebeat.dev/guide/developer-experience/

The answer we had in July was tidy. AI features in developer tools price as a
per-seat licence plus a metered consumption line — the seat pays for access,
the meter pays for tokens. It is the shape the category converged on without
much argument, and the guide carries it in §01.

August broke the tidiness in one move. Meta launched Muse Code with standard
pay-as-you-go pricing at $1.25/$4.25 per million input/output tokens, and
alongside it a "contributor" tier at $0.10/$0.20 — roughly 12x cheaper on
input and 21x on output — in exchange for Meta training on the user's prompts
and completions. That is not a volume discount and it is not a seat. It is a
third axis: the customer pays with data, and the discount puts a public market
rate on what that data is worth.

The immediate consequence is concrete enough to act on, and it is already a
claim: a training-data policy is now a pricing decision, so it belongs on the
pricing page rather than buried in a DPA. Nobody has to guess what the trade is
worth any more — there is a number.

## The tension

Meta is running both directions at once. The same month it priced Muse Code's
contributor tier as a data trade, it open-sourced Muse Glimmer — 30B
parameters, Apache 2.0, on Hugging Face, running on a single consumer GPU with
no network call at all. One product charges you less for your data; the other
gives away a model that never sees it.

That is genuinely awkward for the thread. The simple reading — "data is the new
currency" — does not survive a free, local, Apache-2.0 release from the same
vendor in the same month. The more careful reading is that the data trade is a
*beta acquisition* mechanism rather than a settled pricing axis, and that
Glimmer is the hedge. If that is right, the contributor tier quietly disappears
at GA and this thread cools.

The incumbent clock supports the careful reading. The loop set in early August
gave incumbents until the end of the month to ship — or explicitly rule out —
a data-for-discount tier of their own, with silence to read as "the entrants
are noise." The clock ran out on 2026-08-31 after three weeks of daily sweeps
found nothing dated either way. Called as specified: no incumbent treats data
as a pricing axis today, so the contributor tier stands as one vendor's beta
mechanism until either a rival ships one or Meta's tier survives GA. What
stays live is the meter itself — GitHub moved every Copilot seat to prepaid
billing from October 1 in the same week its promotional credits head toward
their lapse, which is the bill-shock test the remaining loop watches.

Early September firmed up both halves. The contributor tier survived a model
revision — Muse Spark 1.3 rolled out on September 2 with the same spread —
which weakens the "quietly disappears at GA" reading without settling it. And
GitHub's changelog confirmed the other date directly: Copilot Business and
Enterprise signups are reopening while every seat moves to prepaid billing on
October 1, so the bill-shock test now rests on a primary source rather than a
forecast.

The week of September 7 answered the thread's first open loop, once. Atlassian
added two meters beside its Rovo AI credits — automation steps and AI agent
resolutions — announced September 1 with billing from December 3, the admin
usage views and limit controls shipped at once, and the allowance (400 steps
per Jira Standard user a month) and overage ($0.50 per 1,000 steps) published
in its support docs. An automation step is not an AI feature. That is the
first non-generative unit to reach a usage meter since this thread opened,
and it says the meter has become a billing habit rather than a pass-through
of model cost — which is the reading the loop was written to test. One vendor
is not a pattern, so the loop stays open for a second. The sequencing is the
part worth keeping regardless: the cap shipped a quarter before the meter,
which is the honesty test Copilot is about to sit in the other order.

## The unit moved to the outcome

The week of September 14 changed what the meter counts rather than whether
there is one. Sourcegraph took Agentic Batch Changes to general availability
on September 16 billed per changeset merged into the customer's codebase —
not per token, seat or attempt, and nothing for a pull request the team
rejects — and said plainly which default it was rejecting: every other coding
agent charges for effort, per token with a fee on top, whether or not
anything lands. Beta customers merged more than a thousand changesets.

The reason it moved now is that the token stopped being a price a vendor
owns. Vercel's gateway index for September put the average price per token
down 23.2% in August, the third consecutive monthly fall, with open-weight
models at 56% of token volume and 14% of spend (Vercel's own traffic). UC
Berkeley's HarnessTax study, the same week, ran one model through three
coding-agent harnesses at $1.540 and $0.441 per resolved task — a 71% gap —
with no quality difference surviving correction across 42 comparisons. A
token meter passes through someone else's price list, and the harness on top
of it moves the bill by itself. An outcome the buyer already counts — the
merged PR, the resolved ticket (Intercom's Fin at $0.99 per defined outcome;
Atlassian's AI-agent-resolutions meter from September 1) — is the only unit
left that means the same thing on both sides of the invoice.

Two things keep this from being the thread's answer. Outcome billing means
the vendor eats every failed attempt while the customer decides what
"merged" means, so the rate has to carry the merge ratio, which suits an
enterprise contract and not a $20 seat. And Sourcegraph has not published the
rate: its pricing page reads "Enterprise, starting at $16K, includes credits
for AI features". An outcome price the buyer cannot read is a token meter
with better copy, and the new loop above waits for the number or for a
second vendor.

## Why it moved now

The pricing question got sharp because the product question went flat. Three
terminal coding agents launched inside forty-eight hours in early August —
Muse Code, Memcode, Clark Code — and none of them competed on what the agent
can do: Meta on price-as-data-trade, Memcode on repository memory, Clark on an
Apache-2.0 licence. A fourth, Mirafold, sold a UI layer over the incumbents
rather than an agent at all.

When the accessory market arrives and nobody is arguing about capability, the
terms sheet becomes the product. That is why pricing structure is worth
following weekly right now, and it is the thread's actual bet.

The week of August 17 moved the meter without touching the question. OpenAI
cut GPT-5.6 Sol's list price (20% on input, a third on output) and Vercel
stacked its 50% gateway discount on top — a reseller discounting someone
else's model as an acquisition play — while Neon launched its AI Gateway on
explicit no-markup pass-through pricing, betting the value is consolidation
rather than a margin on inference. Both are evidence the meter's *rate* is
now fiercely competitive and increasingly not where anyone's margin lives;
neither is the incumbent data-for-discount answer the open loop is waiting
on. Silence on that clock runs to August 31.

## Open loops

- Does a *non-generative* feature get metered? Answered once on 2026-09-01 — Atlassian meters automation steps at $0.50 per 1,000 above a per-seat allowance, billing from December 3 — so the meter is a billing habit for at least one vendor. A second vendor metering a non-AI unit makes it the pattern; none by the end of the year leaves Atlassian as one company's billing reform. *(by end of 2026)*
- What happens when Copilot's promotional credits lapse (~September 2026) and every seat goes prepaid (2026-10-01)? The first public bill-shock story is the test of whether the meter was priced honestly. *(by around September–October 2026)*
- Does the outcome unit become public pricing — Sourcegraph's per-merge rate on its pricing page, or a second coding agent billing per merged changeset? One vendor with a hidden rate is a sales motion described in outcome language; a published number or a second vendor makes it the fourth axis. *(by end of 2026)*

## On the record

20 dated entries filed onto this thread, newest first.

- **2026-09-21** — [Price the merge, not the token](https://thebeat.dev/issues/2026-W38/): 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.
- **2026-09-16** — Sourcegraph: [Sourcegraph prices coding-agent PRs by merge, not tokens](https://thebeat.dev/signals/#2026-09-16-sourcegraph-outcome-based-agent-pricing): Sourcegraph shipped general availability for Agentic Batch Changes with outcome-based pricing: it bills only for changesets that actually merge, not for the tokens an agent spent producing them. Sourcegraph frames it as fixing an industry default where a coding agent gets paid for effort whether or not its output ever lands in the codebase.
- **2026-09-14** — [Your free tier now has a claim step](https://thebeat.dev/issues/2026-W37/): 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.
- **2026-09-08** — Atlassian: [Atlassian expands usage-based pricing to automation and AI agents](https://thebeat.dev/signals/#2026-09-08-atlassian-usage-based-pricing-expansion): Atlassian is adding two new usage meters on top of its existing Rovo AI credits: individual automation steps (replacing whole "flow runs" as the billing unit) and autonomous AI agent resolutions. Announced September 1, with new Atlassian Administration tools giving org and billing admins real-time visibility and the ability to set limits ahead of the change; allowance limits and billing for the two new meters take effect December 3.
- **2026-09-07** — [Your docs are production now](https://thebeat.dev/issues/2026-W36/): 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.
- **2026-09-04** — GitHub: [GitHub reopens Copilot Business and Enterprise signups](https://thebeat.dev/signals/#2026-09-04-github-copilot-signups-reopen): GitHub is gradually reopening Copilot Business and Enterprise signups for customers who pay by credit card or PayPal, alongside stronger account vetting. The same notice confirms the billing change flagged for October 1: on those accounts, new seat assignments require payment before access, and every already-assigned seat starts incurring an upfront charge at the next billing cycle.
- **2026-08-31** — [The agent's front door is a tool, not a file](https://thebeat.dev/issues/2026-W35/): 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.
- **2026-08-22** — OpenAI: [Cuts GPT-5.6 Sol pricing, and Vercel stacks its own discount](https://thebeat.dev/signals/#2026-08-22-openai-gpt-5-6-sol-price-cut): GPT-5.6 Sol list pricing dropped 20% on input and a third on output, to $4.00 and $20.00 per million tokens on the standard tier. Vercel's existing 50% AI Gateway discount stacks on the new price for the next month or so, taking the Default tier from $2.50 to $2.00 per million input and from $15 to $10 per million output. Existing integrations bill at the new rate automatically because the model ID is unchanged.
- **2026-08-18** — Neon: [Puts an LLM gateway in the Postgres backend, pass-through](https://thebeat.dev/signals/#2026-08-18-neon-ai-gateway): Neon shipped AI Gateway, a beta backend primitive that lets an app call frontier and open-weight models through one endpoint, alongside its existing Postgres, storage, functions and auth. Usage lands on the same Neon invoice.
- **2026-08-17** — [The rail is the product now](https://thebeat.dev/issues/2026-W33/): 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.
- **2026-08-14** — Vercel: [Gives GLM 5.2 away free on AI Gateway to seed default share](https://thebeat.dev/signals/#2026-08-14-vercel-glm-free-promo): Vercel is giving eve agents free use of Z.ai's GLM 5.2, a coding model with a 1M-token context, for about two weeks, served by Blackbox AI over its AI Gateway. New eve agents default to GLM 5.2 and existing ones switch with one config change. The faster glm-5.2-fast variant is excluded, and standard AI Gateway metering resumes when the promo ends.
- **2026-08-12** — Vercel: [AI Gateway data shows DeepSeek overtaking Google on tokens](https://thebeat.dev/signals/#2026-08-12-vercel-ai-gateway-deepseek-report): Vercel's AI Gateway index shows DeepSeek at 25% of token volume, past Google's 11%, with Anthropic holding 30% of volume on 65% of spend. Open-weight models grew from 11% to 36% in three months, and median cost per token fell 13.6% in the latest month after a near-20% spike. Three-quarters of teams changed their model mix, yet the median team's cost dropped only 2.9%.
- **2026-08-11** — Meta: [Meta open-sources Muse Glimmer, a 30B model built for local agents](https://thebeat.dev/signals/#2026-08-11-meta-muse-glimmer): Meta released Muse Glimmer, a 30-billion-parameter agent model distilled from Muse Spark, under Apache 2.0 on Hugging Face, with support across Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM and SGLang. Compressed to roughly 4-bit with block-level speculative decoding, it runs on a single consumer GPU or a Mac with no network call. It's a fully open release from the company that just priced Muse Code's beta around a paid, data-trade contributor tier.
- **2026-08-10** — [The coding agent became a commodity, now the terms are the pitch](https://thebeat.dev/issues/2026-W32/): 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.
- **2026-08-07** — Coding agents: [Three new terminal coding agents in forty-eight hours](https://thebeat.dev/signals/#2026-08-07-coding-agent-entrants-pattern): Three terminal coding agents launched within two days of each other. Meta's Muse Code arrived in beta with a contributor tier priced roughly 12–21x below list in exchange for training data, Memcode entered public beta betting on persistent repository memory, and Clark Code shipped as an Apache-2.0 open-source client over a freemium runtime. A fourth arrival in the same window, Mirafold, sells a UI layer over the incumbents rather than an agent of its own.
- **2026-08-06** — Mirafold: [An MIT-licensed browser UI wraps three terminal coding agents](https://thebeat.dev/signals/#2026-08-06-mirafold-generative-ui-coding-agents): Mirafold launched on Show HN, a locally run daemon that puts a browser front-end on Claude Code, Codex and Gemini CLI. The agent can render cards, tables, charts and pinned live widgets alongside its text, and a real terminal passthrough keeps `sudo` and `ssh` working. The whole product, relay included, is MIT-licensed and free to self-host, with a $12/month Pro tier that only buys QR device pairing and an end-to-end encrypted relay for remote access.
- **2026-08-05** — Vercel: [AI Gateway lists on AWS Marketplace to unblock procurement](https://thebeat.dev/signals/#2026-08-05-vercel-ai-gateway-aws-marketplace): Vercel put AI Gateway on AWS Marketplace. Buyers get private offers with annual contract terms plus usage-based pricing beyond the contract, billed through their existing AWS account with no markup over provider rates. Teams can put inference spend on the AWS bill they already have instead of opening a separate vendor agreement.
- **2026-08-05** — Meta: [Muse Code prices a data-sharing tier 12x cheaper than standard](https://thebeat.dev/signals/#2026-08-05-meta-muse-code-launch): Muse Code is a terminal coding agent in beta, powered by Muse Spark 1.2, aimed straight at Claude Code and OpenAI Codex on macOS and Linux. Standard pricing is $1.25/$4.25 per million input/output tokens. A "contributor" tier drops that to $0.10/$0.20 in exchange for letting Meta train on your prompts and completions — a discount priced as a data trade.
- **2026-08-05** — Memcode: [A terminal coding agent bets everything on persistent memory](https://thebeat.dev/signals/#2026-08-05-memcode-terminal-agent): Memcode is a terminal coding agent in public beta, launched on Show HN. Its bet is persistent memory: it remembers a repository's history, architecture and decisions, so every session starts already oriented instead of re-reading everything. It ships as a single static binary with a one-line install, and runs on Memcode credits or your own API keys.
- **2026-07-27** — [The AI seat gets a price, and the empty ones get counted](https://thebeat.dev/issues/2026-W30/): 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.

## Sources

- [TechCrunch — Meta launches Muse Code, an AI agent for large code bases](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/)
- [Meta — Muse Glimmer](https://developer.meta.com/ai/models/muse-glimmer/)
- [Sourcegraph — Coding agents usually can't price on outcomes. Ours can.](https://sourcegraph.com/blog/agentic-batch-changes-pricing)
- [Vercel — AI Gateway Production Index, September 2026](https://vercel.com/blog/ai-gateway-production-index-september-2026)

## Related

- [Guide — Positioning for developers](https://thebeat.dev/guide/positioning/)
- [Archive — The seat-plus-meter settlement](https://thebeat.dev/articles/2026-07-22-seat-plus-meter-pricing/)

---

Part of [The Beat — developer marketing, on the record](https://thebeat.dev/). Content license: CC BY 4.0 — quote it, link the canonical page.