How AI Coding Is Priced Now: Four Pricing Models Compared

AI coding pricing has stopped being a list of numbers and started being a list of shapes. Four of them now compete for the same engineering budget: a seat subscription with a metered usage boost, a prepaid token plan drawn down against rolling windows, a monthly-reset quota ladder, and a contributor tier that trades training rights for a steep discount. Two plans at the same monthly price can behave nothing alike the week a sprint spikes.

That matters more than it used to, because most teams are no longer standardised on one tool. A typical setup now runs a primary assistant on a seat plan, a bulk workhorse on a prepaid or metered endpoint, and at least one experimental agent on whatever tier was cheapest that month. The finance question is not “which is cheapest per token” — it is “what happens to each of these when the work comes in bursts.” The shapes answer that question very differently.

This guide sets out the four shapes, prices each one against the vendor page it came from, recomputes the ratios that vendors leave implicit, and ends with a budgeting framework for teams mixing several at once. For the per-tool seat comparison across Cursor, Copilot and the desktop agents, see the June seat-economics guide; for the running list of per-token cuts and promo cliffs, see this week’s API pricing tracker. This post is the taxonomy, not the price list.

Strip the branding away and every AI coding plan on the market answers three questions: what unit are you buying, when does it come back, and what happens when you run out. Group by those answers and four distinct shapes fall out. They are not tiers of the same thing — they behave differently enough that a finance model built for one will misprice the others by a wide margin.

Anthropic’s consumer and team plans are the clearest example of the shape. Reading claude.com/pricing on August 7, 2026: Claude Pro runs $17–20 per month depending on annual versus monthly billing and is described as including Claude Code, with no separately published Claude Code allowance on the pricing page itself. Claude Max sits on top of Pro as two usage multipliers — 5x and 20x — priced from $100 per month, both described as including Claude Code with higher output limits.

Team billing splits the same idea across seats. Standard seats are listed at $20 per month on annual billing and $25 monthly; Premium seats at $100 annual and $125 monthly, with the page describing Premium as 5x more usage than standard seats. Both include Claude Code and Cowork. Enterprise is the outlier: the pricing page quotes it as a seat price plus usage at API rates — around $20 per seat plus metered consumption — the only tier billed on pure usage rather than a bundled allotment.

The mechanic that defines the shape is pooling. Anthropic’s support documentation is explicit that the chat app and the CLI draw from one shared allotment rather than two independent ones — which means a heavy afternoon in the chat interface reduces what the coding agent can do that evening. Teams that assumed a Claude Code seat and a Claude seat were separate budgets tend to discover this the hard way.

When the allotment runs out, the documented paths are to upgrade the tier, enable usage credits billed at standard API rates — an explicit opt-in, not a silent charge — switch to pay-as-you-go through the Claude Console, or wait for the reset. That opt-in is what makes this a boost rather than an overage: the seat price stays fixed and the metered spend is a separate, deliberate decision. It is also the reason this shape produces the fewest budget surprises of the four, and the most schedule surprises. We covered the mechanics of that opt-in in more depth in the guide to Fable 5’s usage-credit mechanics, and the broader trade-off in subscriptions versus usage credits.

Alibaba’s Qwen Token Plan is the cleanest example of the prepaid shape. On the qwencloud.com Token Plan pricing page as retrieved on July 21, 2026, Individual tiers ran Lite at $6 per month, Standard at $18, and Pro at $68 — promotional rates against list prices of $8, $25 and $80. Annual Individual plans were $65, $195 and $770. Team seats ran $20, $75 and $200 per month, carrying between 25,000 and 250,000 credits per month.

Measured against those promotional monthly rates, the annual commitment saves roughly 9.7% at Lite and Standard ($65 against $72, $195 against $216) and about 5.6% at Pro ($770 against $816) — a noticeably thinner annual discount than the 15–21% Anthropic and Moonshot offer on their own plan pages. Prepaying further out buys less here, which is worth knowing before a finance team defaults to the annual line.

The dual-window design is the structural difference. A single pooled limit answers “how much can I use in total.” Two overlapping rolling windows answer “how hard can I push right now” and “how hard can I push this week” separately, and the answers can diverge sharply. A team that plans around the short window alone will plan a sprint it cannot finish. The offsetting advantage is that the plan speaks OpenAI-SDK-compatible endpoints, so repointing an existing coding agent at it is a configuration change rather than a migration. The full credit mechanics are unpacked in Alibaba’s Token Plan, decoded.

One clause deserves a calendar entry rather than a footnote. Prepaid token plans are sold as a bucket of credits governed by plan terms, and those terms carry an expiry window — unspent credits are not guaranteed to roll forward indefinitely. Read the expiry language in the terms attached to your own plan before committing to a long prepay, and set a reminder well ahead of it. Prepaid capacity that expires unused is the most avoidable line item in this entire category.

Moonshot’s K3 plan ladder is the monthly-reset shape in its purest form. On kimi.com’s K3 pricing page as retrieved August 7, 2026, the four rungs are Moderato at $15 per month on annual billing ($19 monthly), Allegretto at $31 / $39, Allegro at $79 / $99, and Vivace at $159 / $199. The annual discount is close to a flat 20% at every rung — about 21% at Moderato, 20.5% at Allegretto, 20.2% at Allegro and 20.1% at Vivace — which makes the annual decision a clean one: it is the same trade at every level.

The detail that separates this shape from a simple volume ladder is that the context ceiling itself is plan-gated. Only the upper two rungs unlock extra-long chat capacity up to the model’s full 1M-token window; the lower two are capped below it. Concurrency scales alongside — on the August 7, 2026 retrieval, the lower two rungs listed two concurrent agent tasks against four for the upper two. That means the plan decision is partly an architecture decision. If your workflow depends on holding a large repository in context, the entry rungs cannot run it at any volume. A hands-on walkthrough of the setup lives in our Kimi Code field guide.

One more surface belongs in the picture and is easy to conflate with the plan: the K3 API is sold separately as pure pay-as-you-go, at $3.00 per million input tokens on a cache miss, $0.30 per million on a cache hit, and $15.00 per million output — flat across the full 1M context window with no distance-based tiering. That is a different shape entirely from the plan ladder, and mixing the two figures in a single forecast is the most common error we see in this category.

Meta’s Muse Spark 1.2 ships the same model under two model IDs at two prices. On developer.meta.com’s Model API pricing tables, muse-spark-1.2 lists at $1.25 input / $0.15 cached / $4.25 output per million tokens, while muse-spark-1.2-contributor lists at $0.10 / $0.002 / $0.20. Recomputing those against each other: input is 12.5× cheaper, output 21× cheaper, and cached input 75× cheaper. The consideration is a terms clause stating that prompts may be used to improve Meta’s products.

The discount is not free of engineering consequences either. Meta’s live pricing and rate-limit documentation, read raw on August 7, 2026, gives the standard tier 3,000 requests per minute and 4,000,000 tokens per minute, and the contributor tier 60 requests per minute and 2,100,000 tokens per minute. Recomputed, that is a request ceiling 50× tighter on the contributor tier but a token ceiling only about 47% lower. The shape of the throttle matters more than its depth: a chatty agent loop making many small calls will hit the 60-per-minute wall long before it approaches the token budget, while a workload that sends fewer, larger requests barely notices.

Two caveats belong on the record. Launch-day materials described availability as restricted to select countries, but Meta’s current Model API page describes expanded global access for Muse Spark 1.2 as a whole — treat the restriction as an at-launch condition that may have lifted and verify eligibility on the live page rather than on launch coverage. And note that a $0.30 per-million contributor figure circulating in secondary coverage does not match the primary pricing table; the listed rates are the ones above. The full eligibility and terms analysis sits in our breakdown of Muse’s contributor-tier trade-off.

Headline prices are easy to compare and tell you almost nothing about how a plan behaves under a spike. The mechanic that actually determines that is the window: what it measures, how long it is, and whether unused capacity accumulates. The table below normalises all four shapes onto that axis, with the metered surfaces underneath for contrast.

Read the last column top to bottom and the pattern is stark: none of these shapes lets you save up. Every one of them is a rate limit wearing a different costume, and rate limits are hostile to how software work actually arrives — in bursts, around releases, incidents and deadlines. The practical consequence is that plan sizing should be driven by your peak week rather than your average month, and the gap between those two numbers is the real cost of choosing a plan-shaped product over a metered one.

Every pricing shape works well for the workload it was designed around and fails for the ones it was not. The matrix below is the one we actually use when scoping a stack: what you are buying, the entry price with its surface, what happens the moment you hit the ceiling, and the team profile it quietly punishes.

The column that surprises people is the third one. Two plans at the same monthly price can put the same team in completely different positions on a Thursday afternoon: one asks for an explicit upgrade decision, one silently slows down until a window slides, one simply stops until the cycle turns, and one throttles on a dimension — requests per minute — that has nothing to do with how much you spent. Budgeting on price alone hides all four of those outcomes behind a single number.

Most teams now run several of these at once, which means the finance model needs a row per shape rather than a row per vendor. The matrix below is how we route the decision when scoping an engineering stack.

Three habits do most of the work. First, size on the peak week — every shape in this category is a rate limit, and rate limits are sized by peaks, not averages. Second, instrument burn per tool before signing an annual commitment; the annual discounts vary enough between vendors (roughly 15–21% on the Anthropic and Moonshot plan pages against 5.6–9.7% on the Qwen annual Individual lines, measured against their promotional monthly rates) that a year of prepay is a materially different bet depending on where you make it. Third, keep one metered fallback warm at all times. The single cheapest insurance against a plan ceiling is a pay-as-you-go endpoint already configured, tested and approved, so hitting a limit costs a routing change rather than a day. If you want that instrumented properly, it is exactly the sort of thing our analytics and measurement work is built for.

It is also worth reading where the shapes are heading, because the consumer end of the market is moving first. OpenAI’s release notes of August 6, 2026 state that GPT-5.6 Luna becomes the default model for Free and Go users this week, and that unlimited text chats plus a new Think button are due to start for those tiers next week — the week of August 10 — subject to abuse guardrails, with limits still applying to file uploads, images and other tools. Plus and Pro are described as getting an updated GPT-5.6 Sol and a new slider controlling how much thought goes into a response, and the notes state that Work and Codex are not changing as part of that release. The unlimited chats and the Think button are scheduled for next week rather than shipped, and that “unlimited” applies to text chats on those tiers rather than to every model or every surface.

The direction it implies is the one worth planning around. As the cheap, high-volume end of consumer usage drifts toward unmetered text, the metering does not disappear — it migrates to the expensive surfaces: tool calls, agent concurrency, long context and file handling. That is already visible in the developer plans, where the interesting boundaries are no longer volume ceilings but capability ceilings, and where the tier you buy determines what the agent can attempt rather than only how often. Expect the next round of plan design to gate concurrency and context more aggressively than tokens, and expect the shapes described here to outlive every price in this article.