5dive Pulse · Token Maxxing
Which AI subscription gives you the most tokens per dollar?
We take what each plan lets you use in a month, price it at what those tokens would cost you through the API, and divide by what the plan costs. Higher means more for the same money.
Most providers do not publish a usage limit, so most of these numbers are estimates. Every row shows how it was worked out, what it rests on, and when we last checked it.
Value per subscription dollar
Individual plans, from the model providers themselves. This ranks value, not model quality — that is a separate question, on the intelligence board.
Each row is what a plan gives an agent that runs until the plan stops it. Open one for its range, sources and check date; rows we metered ourselves say so.
- 1
Claude Pro
$20/mo · Claude Opus 5
~20× - 2
ChatGPT Plus
$20/mo · GPT-5.6 Sol
~17× - 3
Claude Max (5×)
$100/mo · Claude Opus 5
~17× - 4
ChatGPT Pro (5×)
$100/mo · GPT-5.6 Sol
~16× - 5
ChatGPT Pro (20×)
$200/mo · GPT-5.6 Sol
~16× - 6
Claude Max (20×)
$200/mo · Claude Opus 5
~15× - 7
Google AI Pro
$19.99/mo · Gemini 3 Pro
~9.5× - 8
Google AI Ultra (5×)
$100/mo · Gemini 3 Pro
~9.5× - 9
Google AI Ultra (20×)
$200/mo · Gemini 3 Pro
~9.5× - 10
GLM Coding Plan (Max)
$168/mo · GLM-5.3
~6.7× - 11
GLM Coding Plan (Pro)
$80/mo · GLM-5.3
~6.0× - 12
SuperGrok
$30/mo · Grok 4.6
~4.6× - 13
SuperGrok Heavy
$300/mo · Grok 4.6
~4.6× - 14
GLM Coding Plan (Lite)
$18/mo · GLM-5.3
~4.5× - 15
SuperGrok Plus
$100/mo · Grok 4.6
~4.5× - 16
Qwen Token Plan Pro
$80/mo · Qwen3.8-Max + more
~3.9× - 17
Qwen Token Plan Standard
$25/mo · Qwen3.8-Max + more
~3.1× - 18
Kimi Moderato
$19/mo · Kimi K3
~2.6× - 19
Kimi Vivace
$199/mo · Kimi K3
~2.6× - 20
Qwen Token Plan Lite
$8/mo · Qwen3.8-Max + more
~2.4× - 21
Kimi Allegro
$99/mo · Kimi K3
~2.2× - 22
Kimi Allegretto
$39/mo · Kimi K3
~2.1×
Glass-box methodology
One formula. Every number sourced or flagged.
Value multiple = the API-equivalent retail value of a month's modelled usage, divided by the plan's monthly list price.
Why subsidised plans win it
Value per dollar is really subsidy per dollar. Plans riding an expensive API top the list.
Every ceiling is modelled
No plan publishes a token quota, so usage is modelled from one stated workload: a coding agent running continuously until the plan's own rate windows stop it.
What API-equivalent means
The public retail API price of the underlying model, not the provider's own inference cost — retail subsidy, not margin. No quality reweighting: a weaker model's tokens count the same as a flagship's.
Which is why the same month of use is worth far more at an expensive-API flagship (Claude Opus, ChatGPT Sol) than at a cheap one (GLM, Qwen), and why plans on pricey models — plus Kimi, a flagship-priced API sold cheap — top the list.
Why most of these are estimates
Almost no provider publishes a token quota, so we model a month of an agent running continuously until the plan's own limits stop it. That model is an upper bound, not a measurement.
Where we have metered a plan end to end, the row publishes measured dollars and is marked as measured. A metered row can therefore read LOWER than an estimated one: measurement is a correction, not a bonus.
How the floor and ceiling are modelled
No consumer plan publishes a token quota, so every plan's monthly usage is MODELLED from one stated, reproducible workload: SUSTAINED AGENT LOAD — a coding agent running continuously at ~40,000 tokens/turn (35K input + 5K output), ~30 turns/hour, until the plan's own rate windows stop it ≈ 105.6M tokens/mo on an entry plan. Each row scales that by its own sustained ceiling relative to its provider's entry tier: a published request or credit ceiling where one exists, otherwise proportional to PRICE — never to an advertised “N× usage” session-burst label. Input is split at the provider's PUBLISHED cache-read rate, because a measured 91.9% of an agentic workload's input is re-sent context billing as a cache read, typically a tenth of fresh input. What is left between floor and ceiling is uncertainty in that plan's own ceiling, not a change of workload. Agents are bounded by the QUOTA, not by the clock: on the one plan we could check, a day of round-the-clock work drew 22% of a SEVEN-day allowance, so saturating the quota is worth ~1.61x this modelled volume, not six-fold. That lift is applied only where a published quota or our own meter anchors it. Each row LEADS with the QUOTA-SATURATED end of its own band, because that is where an agent that simply runs ends up; the full range, the confidence and the check date sit behind that row’s (i). Every row declares where its throughput input came from — observed on our own meter, vendor-derived, third-party logged, or modelled — plus its sample count and, where the source carries a date, when it was last observed. Weekly quotas convert at 4.348 weeks/mo (365.2425/12/7), not 4. Where a provider meters chat, its coding agent and its work product on SEPARATE allowances, the row names the ONE surface it ranks — the provider's coding surface — and never blends two allowances. Where two sibling tiers print the same multiple only because the model divided price by price, the row says “unmeasured; tied by assumption”, not “same rate”.
What maxxed means, and why a metered row can read lower
Every row's band runs from the plan's binding quota saturated by SUSTAINED AGENT LOAD — a coding agent running continuously at 40,000 tokens a turn, 30 turns an hour, hitting the rate windows naturally — up to that same load PLUS what an optimizer adds: schedulable work moved into a RECURRING off-peak window, prompt caches preserved, usage valued at the highest API-priced model the plan includes. The row LEADS with the SATURATED end: an agent reaches it by default, while almost nobody runs the optimizer's tactics, so a mean would flatter every row. Temporary promotions and one-off bonus credits are excluded everywhere — a permanent score cannot rest on terms that expire. Each row says which quota binds it first and how its ceiling was obtained, strongest first: (1) a cycle we metered against the vendor's own gauge; (2) a published credit pool converted through the vendor's deduction coefficients; (3) a published request cap times measured tokens per request; (4) nothing better exists, so the modelled ceiling stands in and the row says so. Class 4 is an upper bound on a modelled token mix, not a saturation measurement. Where a row WAS metered it publishes the measured DOLLARS and never re-values the tokens: a real cycle carries its own mix, and this board's modelled mix assumes far more output than agent traffic produces. On the one row with both readings the gap is 3.15x, almost entirely because the model assumes 12.5% of tokens are output where agent traffic runs under 1%. So read a MODELLED row as an upper bound hot by roughly that factor — correcting the mix on every row is a separate change this release does not make. That is also why a metered row can read LOWER: measurement is a correction, not a bonus.
Where the headline number comes from
The quota-saturated end of that row's range: what an agent that simply runs reaches. The other end adds recurring off-peak scheduling, cache preservation and valuing usage at the priciest model the plan includes — almost nobody does any of that, so a centre between the two would flatter every row. It is still printed behind the row icon as the upper bound.
Where the range is uncertainty about how much a TIER delivers rather than optimizer effort, both ends are already saturated, so that row leads with their geometric mean instead — geometric because these are multipliers. The row says which it did. Either way the headline reads no input the range does not, which is why it prints with a tilde.
Why some tiers tie
Where a provider publishes no ceiling, each of its tiers is modelled proportional to price — and price divides straight back out, so those tiers print the same headline by construction. Not a rounding artefact: the dearer ChatGPT and SuperGrok tiers buy headroom, not a better rate. Tied rows are marked and listed cheapest first. Providers that do publish a ceiling — Kimi, GLM, Qwen — separate on their own.
Which price the value is divided by
The list monthly price, for every provider, so no two rows are ever compared at different discount states. Where a provider is currently advertising a lower rate, that price and the headline it would produce both appear in the row's details. They do not move the rank.
How to read a wide band, and what moves the order
The band is throughput uncertainty in that plan's own sustained ceiling — both ends describe the same reader, a continuously running agent. A wide band is not a weak rank: a row whose ends move apart (Claude Max, Google AI Ultra) is one where the ceiling itself is contested.
What does move the order is each row's scale factor. Where a provider publishes no ceiling we scale proportional to price, never to an advertised N× usage label: those are five-hour-session bursts that weekly caps stop you sustaining across a month, which is why an entry plan can out-value its own higher tier.
What these figures are not
Quotas are dynamic. Every figure here is a SNAPSHOT of what a plan's published terms implied on the date in its Last-checked column, not an entitlement you can hold a provider to. These plans re-price and re-cap without notice, sometimes weekly, and none of them publish a token quota at all.
This board prices TOKENS, not experience. Latency, queueing, 429/rate-limit behaviour, uptime and model quality are all excluded from the multiple — a plan can rank high here and still be unpleasant to work on, and model strength is a separate board.
Every row says where its throughput input came from. Observed means we metered it on our own subscription against the provider's quota meter; where that subscription has ended the row says sample ended and carries the date, so a frozen sample never reads as a live one. Vendor-derived means the vendor publishes a credit or request ceiling, or a usage ratio, and we applied it. Community-observed rests on third-party logs, which widens the band and caps the confidence. Modelled is the fallback: no telemetry, no published ceiling, so the tier is scaled proportional to price.
An observation is printed beside the modelled headline, never in place of it — the ranking stays on the one formula every row shares. Each row also carries its sample count, the date of the last observation where its source is dated, how fast that plan's terms move, and grades plan-terms confidence separately from throughput confidence: a price you can read off a vendor page tells you nothing about the quota behind it.
Kimi is priced here from Moonshot's INTERNATIONAL catalog (kimi.ai, USD), because every other plan on this board is priced from its internationally marketed page. Moonshot also sells a separate China catalog (kimi.com, CNY) with a different tier ladder — Andante CNY 49, Moderato CNY 99, Allegretto CNY 199, Allegro CNY 699 — which at today's rate is materially cheaper and would rank Kimi differently. Which catalog you can buy from depends on your region.