Methodology

Confidence grades

Every allowance carries a grade. It describes the strength of our evidence, not the size of the number.

official

Stated on the vendor's own pricing page or documentation. Requires a source URL and the vendor's wording.

measured

Measured by us or by the community. Requires a citable source and a description of the method.

estimated

Derived from other known figures. Requires the reasoning to be written down.

A derived figure takes the weakest grade among its inputs. This comparison page shows only official and measured figures; estimated ones appear in the planner, where you can see and change the assumptions.

What the vendor discloses

Separate from confidence, and often confused with it. Anthropic's "5x more usage than Pro" is official-grade evidence — their own words — of a plan whose allowance is still not a number.

published

An absolute figure, or an explicit "unlimited".

relative

Only a ratio to another plan, whose own allowance may be unpublished.

undisclosed

No figure at all.

Capability tiers

Tiers label the models whose vendor published a comparable score. They appear on the model chips and nowhere else: a tier is a coarse bucket, never a score and never a ranking within itself, and nothing on this site filters or ranks by one.

A tier comes from a vendor-published Terminal-bench 2.1 score where one exists: frontier at 80 and above, strong from 70 to 79.9, efficient below 70. Where no score is published but the vendor states an unambiguous ordering about its own model, that statement sets the bucket and is recorded verbatim. Where neither exists, the model stays untiered.

Untiered does not mean weak. It means nobody published a figure we can compare. It used to cost the model its place under a capability floor — that floor is gone, because what it rejected was overwhelmingly vendors who publish no benchmark rather than models that fall short. A value figure names the model it was priced on instead.

Two things we deliberately do not use: scores from different benchmark versions, which move a model a whole bucket on nothing but the version; and distributors' billing tiers, which track price closely enough that ranking value-per-price on them would be circular.

Current coverage: 9 of 19 base models carry a tier.

Assumptions

Converting an allowance into dollars needs a model of how a coding session consumes tokens. These are editorial assumptions, not measurements, and they are adjustable in the planner.

How a value figure is built

A plan's API-equivalent value comes from one allowance, priced on one model. Three steps decide which allowance and which model.

1. 1Within one model, the rows that govern it are constraints. A plan-wide allowance and a row naming that model are not two offers to choose between — the second narrows the first — so the model's ceiling is the tighter of them.
2. 2Across models, the same rows are alternatives. Models sharing a five-hour window spend it on whichever one you pick, so the group is worth the best choice in it. The model that wins is the one the figure is priced on.
3. 3Across groups, every row binds. A five-hour window and a weekly cap are both hit, so the plan is worth the tightest of them.

Value multiple = that figure ÷ the monthly price. Above 1.00 means the allowance buys more, at published API rates, than the subscription costs.

Four cases produce no figure rather than a soft one: a plan priced only on request; a published allowance we cannot convert to dollars anywhere in the chain; a free plan, which has no price to divide by; and allowances that reach none of the models we list.

Which model for which kind of work

The planner routes each category of work to a model. This is the weakest evidence on the site, and the coverage below is the first thing to read.

Coverage

Where a benchmark exists, a model's band is set relative to the best published score in that same category: preferred at 92% of it or above, capable at 70% or above, unfit below. Relative rather than absolute because these benchmarks are not equally hard — the field tops out around 85% on one and 65% on another, and a fixed cut would grade the same model differently on nothing but the difficulty of the test.

Where no benchmark covers a category, a model inherits `capable` from its capability tier and carries no success rate. It can still be routed, but it can never win on a cost comparison it was never measured in, and the planner marks that row as chosen by tier.

Three kinds of evidence, in order of strength: benchmarks a vendor publishes in its own comparison table; aggregate behaviour from third-party leaderboards, which is not yet used because each one's redistribution licence has to be cleared first; and reader feedback on this site, which can raise a proposed change for review and never edits anything by itself.

Architecture and design has no mature public benchmark and is not expected to acquire one, so the planner shows the capability tier to look for and names no model for it.

From token price to cost per task

A stronger model costs more per token and needs fewer attempts, so 'dearer per token' and 'cheaper per task' are routinely true of the same model. The planner compares the expected cost of finishing one task: the category's baseline tokens, adjusted for the model, divided by its solve rate.

Baseline tokens per task

Every result is a range, produced by varying token use by ±30% and solve rate by ±10 percentage points. A single figure built on an editorial baseline times a benchmark solve rate would be the most confident-looking number on this site and the least defensible.

Feedback thresholds

A disagreement about which model suits a category becomes a proposed change once 5 readers have said the same thing and they are at least 70% of the votes on that pair. The proposal goes to a person, who checks it against a published benchmark; feedback never edits the data by itself.

Reports of hitting a plan's limit are grouped by plan, period and usage intensity, never pooled across intensities — heavy users cluster on the larger plans, and pooling would show the expensive plans running out sooner as an artefact of who bought them. A rate is shown only once a group holds at least 20 reports. Below that the cell says so.

How the figures are kept current

A watcher re-reads each vendor's pages every three days and compares what it finds against what we hold. Nothing it finds is published until a person confirms it against the source. When a vendor's page stops being readable, that shows on the plan rather than being absorbed quietly.

Last checked: 2026-07-29