Pay per token, not per month — but the cheap-model era has a plot twist.

Side one

What the subscriptions cost

Consumer pricing as of mid-2026, from third-party roundups — sentisight, felloai, perspectiveai. None is a provider page.

Notice the shape of that table. Every vendor sells three or four tiers distinguished mostly by usage caps — $8 gets you a little, $20 more, $200 a lot. That is usage-based pricing wearing a subscription costume: you are already paying for volume; you just can’t see the meter. The API gives you the same meter with the cover off.

Side two

What tokens cost

Staleness is the failure mode of every pricing post, so: teemoon hardcodes no price menu — it fetches each provider’s live catalog and shows current rates in the model picker. The near.ai rows below come from the live public catalog (cloud-api.near.ai/v1/models), the Fireworks rows from its serverless pricing page, both re-verified 2026-08-12.

The same open-weight model is sold by competing hosts at competing prices. GLM 5.2 lands at exactly $1.40/$4.40 on both — two hosts, different hardware, no coordination, one number. That is what a commodity looks like when it has finished converging.

The arithmetic

When does $20 flat actually win?

Do the math honestly or don’t do it. A chat turn resends the whole conversation as input and gets a reply as output, so model a turn as ~2,000 input and 400 output tokens. A genuinely active user — 20 turns a day, every day — sends about 1.2M input + 240K output tokens a month. At the vendors’ own published rates, not a reseller’s:

The Anthropic row costs the argument something real, so out loud: $2/$10 is what Anthropic charges, and it is not a price a teemoon user can reach. The GPT and Gemini rows don’t have that problem — both expose OpenAI-compatible endpoints. A comparison that hides its worst case isn’t worth the one it wins.

Twenty turns a day on the priciest model in the table above still comes in under ChatGPT Plus. On anything mid-tier the subscription is a 10–100x overpayment for this usage — and most people are not 20-turns-a-day users. A $20 subscription used for three questions a week is paying roughly a dollar per question that the API prices at a tenth of a cent.

One thing this table must not be read as saying: that the rows are interchangeable. A $0.24 month and a $13.20 month buy different things. What the table actually shows is that metered billing lets you match the model to the task — cheap tokens for the 80% of conversations that are ordinary, frontier tokens for the ones that aren’t — instead of paying frontier rates for everything because that’s the tier your plan is on.

Now the crossover. Input cost scales with conversation length: a thread grown to 30K tokens costs about $0.15 per additional turn at GPT-5.6 Sol rates, and 130 such turns is $20.

The meter has a second gear too — all three Western frontier vendors charge a long-context surcharge past a documented threshold: OpenAI steps to $10/$45 above 272K prompt tokens, Gemini 3.1 Pro to $4/$18 above 200K, xAI exactly doubles above the same 200K. So metered billing loses to the flat rate for people who live in very long contexts on the priciest models, run agentic workloads burning tens of millions of tokens, or want the bundled features a subscription includes.

For conversational use — which is what a chat app is — metered access costs a small fraction of every mainstream subscription, and the gap widens the less you use it.

The counter-argument

Frontier labs sell tokens below cost in their own apps

Price out what a power tier actually delivers and the result is uncomfortable for this post’s thesis, so here it is plainly. Anthropic’s Max tiers are sold as usage multiples, 5x Pro for $100 and 20x Pro for $200, which already means the top tier buys 20x the usage for 10x the money.

Against API rates the gap becomes a chasm: analyses of saturating Claude Code users put a $200 Max plan at roughly $5,000/month in API-equivalent compute — about 25x, and one eight-month case study reported $15,000 of consumption for $800 of spend. Developer-reported, so treat them as ceilings. Conservative breakpoint analyses land in the same place: past ~50M tokens a month the subscription wins, past 200M it isn’t close.

For heavy users, a subsidized flat-rate subscription is dramatically cheaper per token than anything you can do with your own API key — teemoon included — and no amount of framing changes that.

Two notes before the counterweights. First: that arithmetic is only possible for Anthropic, because its limits are published as multiples and its power users measure their burn. OpenAI publishes caps as messages, Gemini and Grok as “5x more usage”-style language — none of it convertible to tokens. You cannot compute what a subscription’s tokens are worth because the vendors decline to tell you, and that opacity is itself part of the argument. Second: the subsidy exists for a reason. Nobody sells $5,000 of compute for $200 out of kindness.

That last point stopped being hypothetical the day a vendor put a figure on it. From late 2024 xAI priced the data trade explicitly: share your API requests for training, receive $150/month in free credits. xAI ended the program in May 2025; the docs page now 404s. Both halves are instructive. While it ran it was the most clarifying document in this subject — a frontier lab stating on a public price card that a month of your traffic is worth about $150, the same trade every “free” tier runs at an unstated rate. And its termination demonstrates the other lesson in miniature: a subsidy is a business decision, and business decisions end.

The plot twist

The cheap-model story is compressing

A year ago this post would have leaned hard on one fact: Chinese open-weight models cost roughly a tenth of Western frontier models. Below the frontier that is still true. At the frontier, Kimi K3 moved onto the Western price card on purpose.

On 16 July 2026 Moonshot released Kimi K3 at $3.00 / $15.00 per million — a 3–4x repricing from K2.6, landing to the cent on Claude Sonnet’s card, a match observers called out immediately. The tempting reading — that this abandons the cheap-Chinese-model thesis — fails a capability check.

K3 doesn’t benchmark as a Sonnet-class model. It debuted at 57 on Artificial Analysis’s Intelligence Index, within a few points of Claude Fable 5, and on agentic deliverables scored second among all models measured. Its actual peers sit on the $5/$25, $5/$30 and $10/$50 rows above — against which $3/$15 is a 40–70% discount. (Caveats in proportion: several headline scores are vendor-reported on vendor-chosen harnesses, and K3 measures notably worse than Opus on adversarial traps and knowledge reliability.)

What died is the order-of-magnitude discount. What survived is a capability-adjusted undercut of roughly 2–3x. The thesis compressed from “a tenth of the price” to “a third to a half, for now.”

And the mechanism matters more than the number. The K3 weights shipped not under MIT but under a bespoke license: any licensee operating what the license calls a Model as a Service business — third-party inference access where the customer controls the inputs — must sign a separate agreement with Moonshot once aggregate revenue passes $20M over any twelve months (Tom’s Hardware). No public source confirms a revenue share (see open questions).

The market effect is unmistakable: nobody undercuts $3/$15. Fireworks sells K3 at exactly $3.00/$15.00; OpenRouter’s listing routes to Moonshot’s own endpoint. Contrast GLM 5.2 two tables up, where two hosts land on the same number. Open weights made hosting a commodity; a license toll gate switches that mechanism off. K3 is technically open-weight and economically price-controlled.

Below the frontier, nothing moved at all: K2.6, GLM 5.x, Qwen 3.x and DeepSeek V4 remain genuinely cheap and competitive across hosts. The repricing is a frontier phenomenon — so far.