Every plan gets one number: how many coding turns of mixed work its allowance covers in a month. The allowance is taken at API list prices, and each kind of work is done on the plan’s cheapest model that is good enough.
One turn is an average agentic coding request: 2.9M input tokens, 98% read from cache, and 9.8K output tokens. These are the averages Slopalytics publishes from T3 Code usage, observed Nov 10, 2026. Each model’s turn is priced from its own token prices, so cheap cache reads count for a lot.
Context is the same at every reasoning setting. Output scales with the setting, by the cost ratio Artificial Analysis measured against the model’s high setting. A model listed without a setting is treated as high and labelled “default reasoning”. A model with no cache price pays full input price.
The alternative unit is a benchmark task: one Artificial Analysis Intelligence Index task at the cost they measured. It is light on tokens and caching, so it favours models with expensive cache reads.
Nobody runs one model all month. Work is split into four kinds, each needing a model of a certain score. Every score is a fixed distance below the best model on the API market (today 64.6), so the standards rise as models do: what only the best model could do a year ago is everyday work now. The band is tight at the top and loosens as the work gets more ordinary.
The Work mix setting says how your turns divide between them:
A plan’s value is one number: the turns of that mix its allowance covers in a month. For each kind, the plan uses whichever of its models and reasoning settings stretches the allowance furthest: the cheapest setting that is good enough, not the maximum. The plan always chooses, so an extra model never lowers its value. Claude Max can do everything five Claude Pros can, and comes out at five times Pro or more.
A model that falls short of the score a kind of work needs still counts, for less: every 2 points short halves what a turn counts. A model 2 points under counts half a turn, 4 points under a quarter. A kind of work is done only on models within 6 points of the best the plan has for it; models further down are not used, however cheap: retrying a weak model does not add up to a strong one. A plan whose best model is far short still does the work on it, at the small credit the halving gives, so on a mix with hard work a plan of small models shows how little it gets you. One number needs some exchange rate between quality and quantity, and no data source gives one; this is ours. It decides how plans without strong models fare. Turns of balanced work per dollar if it were different:
| Halves every | 1 point | 2 points (used) | 3 points | 5 points |
|---|---|---|---|---|
| Claude Pro | 23 | 54 | 63 | 69 |
| Claude Max | 69 | 69 | 72 | 77 |
| ChatGPT Pro 100 | 1.4 | 21 | 39 | 55 |
| Google AI Ultra 20× | 33 | 56 | 64 | 71 |
| Copilot Pro+ | 5.1 | 5.4 | 5.9 | 6.3 |
| Muse Code Power | 0.017 | 1.4 | 4.2 | 7.9 |
| Kimi Max | < 0.01 | 0.034 | 0.17 | 0.42 |
| Mistral Vibe Student | < 0.01 | 0.16 | 4.2 | 42 |
The order among plans with strong models holds across these; plans whose best model is far below a kind of work move by orders of magnitude, so read their figures there as “very little”, not as a precise amount.
The second judgment call is the share of frontier work: 5% of turns on Balanced, 10% on Mostly hard, none on the light mixes. No data gives this share. It matters a great deal for plans without the best model, which do that work at a fraction of a turn, and little for plans that include it. It stands for the tasks you would not hand to anything but the best model. The third is how wide each band is: hard work was within 3 points until November 2026, which let only the single best model count in full for a fifth of all work and put every other provider’s plans at a steep discount.
ChatGPT Pro 100 has a model that clears the hard-work score, GPT-6 Astra at 56.9. But a turn on it costs $4.37 at API prices, because its cache reads are expensive and an agentic turn is mostly cache reads. GPT-6.1 Sol at 54.6 is 2.0 points short, so a turn on it counts 50% of a hard turn, but costs only $0.611: it covers more hard work from the same allowance, so hard work goes to it. Frontier work needs 63.6, which no ChatGPT model reaches. It goes to GPT-6.1 Sol at 54.6, 9.0 points short, where a turn counts 4.4% of a frontier turn: those few turns use up a large part of the allowance. On Balanced, ChatGPT gets 21 turns per dollar; on Mostly light, with no hard or frontier work, 138.
The plan always takes the cheapest option that is good enough: on today’s scores Claude Opus 5.5 clears the hard-work score for less than Fable, so Claude’s hard work is counted on Opus and only its frontier work on Fable at its top setting. A model the plan caps, such as Fable at half of the weekly limit on Claude Max, may take only that share of the allowance. Anthropic’s help page says Fable draws from the regular weekly limit, so its meter is treated as a cap inside that limit, not as extra usage. Min. intelligence hides plans whose strongest model scores below the threshold.
With Engineering index on, scores come from Artificial Analysis’s Engineering Index instead: engineering knowledge, reasoning, agentic coding and terminal use, broader than coding alone. It runs a few points higher, so the normal and light bars are carried over by a straight-line fit between the two indexes across every model; the frontier and hard-work bars follow the best model on that index.
One line per product, running from its cheapest tier to its dearest. In the per-month view, hover a plan to see its equal-deal line: plans above it give more per dollar, plans below give less. The Pareto line, off by default, marks the most value available at each price; the plans on it are circled with a dotted ring.
For each product, the line gives the price of its cheapest tier that covers the number of turns you need a month, on the chosen work mix,, stepping up where you outgrow a tier. The lowest line at your usage is the cheapest plan; the dotted Cheapest overall line traces it across all products, with a dotted ring at each plan where the cheapest option changes. Unlike Price vs value, this does not assume you use your whole allowance.
Pick a monthly budget. Every combination of plans that fits in it is checked, up to the number you allow (two by default), and the ones covering the most turns of the chosen work mix are listed, with the best single plan for comparison. Each plan counts as in Price vs value, on the chosen work mix. Plans in a setup simply add up; giving each kind of work to the plan best at it would do better. A plan is only added if it earns its price: it must bring at least a tenth of the setup’s own turns per dollar, otherwise the setup without it is the better buy. Only plans one person can buy monthly are used: no team seats or yearly billing, and student plans once. Several accounts at one provider are allowed by default and can be turned off; providers’ terms may not allow them. The figures assume you use every plan’s whole allowance.
Moving to another account loses the prompt cache, so the agent reads its working context again at full input price. Each switch is charged as 150K tokens of context at the difference between the input and cache price, about half a turn on most models, taken from the account switched to. The same re-read costs a small plan as much as a big one, so it takes a bigger share of a small plan. You switch to an account each time the one before it runs out: about once a day (30 times a month) for plans with a five-hour limit, about four times a month for weekly pools, and once a month for monthly allowances. The account doing the most work is used first and is not charged. Cache-write surcharges, re-exploring a codebase in a new session, and the friction of a second agent are not counted, so the real loss is higher.
Each plan’s largest allowance at API prices divided by its price. 1× costs the same as paying the API for the same model; above 1× the provider is subsidising you. Allowances assumed rather than measured (Gemini 4 Hydrogen on Google AI) are left out.
Each dot is a model at one reasoning setting; left is cheaper per turn, up is smarter. By default there is one line per provider across its standard monthly plans, and each model sits on the cheapest of those plans that includes it, so a model only the bigger tiers include (Fable, Gemini 4 Carbon, Muse Wonder Full) still appears. A dot is kept only if it scores higher than everything cheaper on its line; settings that cost more and score less are hidden. The dashed line is the API: the cheapest model at each score bought by the token. All tiers draws every plan as its own line.
Every plan is graded by the weakest of its inputs: published when the provider states the allowance, on its site or in the subscriber’s account; measured when it was read off real accounts; derived when it is a published or measured figure scaled by a multiple the provider has stated; estimated when it rests on an assumption of ours; anecdotal when only unverified user reports exist. Estimated and anecdotal plans are drawn hollow and tagged “est.”.
The provider states the allowance. GitHub Copilot gives monthly AI credits worth $0.01 each. OpenCode Go gives a dollar allowance per model. Z.ai gives weekly credits and a token-to-credit formula; the range runs from the worst token mix at peak hours to the best off-peak. Mistral states the monthly allowances of Vibe Pro and its Education version in the subscriber’s account: €127.50 of Vibe Code usage and €12.75 of API usage (read 12 November 2026).
Not published, but independently measured on real accounts with a described method.
The measurements priced usage with 92–97% of tokens as cache reads; the turn profile used here has 98%. OpenAI moves to a shared token allowance in December 2026; its figures need re-measuring after that.
The figure depends on unverified user reports. Treat these as a rough guide.
MiniMax and ChatGPT Go publish no absolute allowance, and no measurement or credible report exists for the plans sold outside China (Real API Pricing has one for MiniMax’s Chinese M Plan, which is priced differently). They sit on the price-only row below Price vs value, show as “?” in Subsidy vs API, and are listed in Model access. Meta One includes no coding tool; Meta’s coding plans are Muse Code.
Model scores and prices come from the Artificial Analysis Data API, synced Nov 13, 2026 (739 model entries, Intelligence Index v5). Measured allowances from Real API Pricing (MIT licence) are synced from its published data, last Nov 13, 2026, and reviewed before they go live. Other plan prices and limits are maintained by hand, last checked Nov 12, 2026; every plan links its sources on the Plans page. The History page values every date since 24 July 2026 with this same method: before Nov 13, 2026 on plans reconstructed from model release dates and dated plan changes, after it on each day’s recorded inputs. Logos are from the lobe-icons set and belong to their owners.