IQnomics

How we measure

IQnomics compares model intelligence and cost, then asks what work done on a subscription would cost billed by usage. Each figure has a source and a time period.

Intelligence

The intelligence score is the Epoch Capabilities Index (ECI), published by Epoch AI under a CC BY 4.0 license. It combines many benchmarks into one scale, on which Claude 3.5 Sonnet sits at 130 and GPT-5 at 150. Our snapshot includes models released in its 270-day window that have an API or downloadable weights and a public price we could verify. It does not include every model Epoch has scored.

We don't change Epoch's numbers. Where Epoch gives a 90% range, the model's page shows it.

Price per token

Prices come from each lab's own pricing page, never from a reseller or aggregator. Every model's page links the page we read and the date we read it. We use the standard price: not batch, not priority, and not the surcharge some labs add for very long prompts.

The blended price weights input three times as heavily as output, because most real work sends more text in than it gets back.

When a price is a limited-time promotion, we recheck it before the promotion ends.

Cost per task

Price per token leaves out how many tokens a model uses. A model that thinks twice as long costs twice as much for the same answer, so we run models on real work and count what each task takes.

The work is six small Python projects, each with a failing test file: a cache with expiry, a CSV parser and a date parser among them. The model has to make the tests pass without editing them. We aim for three runs of each project, one at a time, in a fresh copy. The six released Claude cohorts have 18 to 19 included attempts each; sample counts differ. It runs through its lab's own command-line coding tool on a paid subscription, at high reasoning effort.

We count tokens from the coding tool's own report, split into fresh input, cached input, cache writes and output, and price them at the lab's list prices above. After each run we run the tests ourselves to record a pass or a fail. Cost and output tokens per task are means across included runs; task time is the median.

Sometimes a coding tool stops a run because the model tried a command outside the few we allow, such as checking for a Python environment. We leave these interrupted attempts out of cost and pass rate and show their count separately on each measured model page. The reported pass rate covers included runs only; exclusions affect comparisons between models.

These are estimates from one setup. A different coding tool, task set or reasoning effort would give different numbers, so compare models within this site rather than against other sites.

Subscriptions and API-equivalent use

Allowance under review. We withhold all weekly allowance figures, task counts per week and fractions of a weekly pool until evidence, plan applicability and publication approval are established. A subscription price or usage multiplier does not establish absolute weekly capacity.

Historical meter observations belong to their original plan, product pool and date. They are not transferred to today's plan names. Conditional diagnostic calculations remain private and do not become public chart values.

The receipt shows each checked tier's advertised monthly price, which may differ from an account invoice or regional offer. For a verified usage sample, API-equivalent dollars mean the approximate cost of that same work at the dated usage-based prices. Usage may be observed while its API-dollar valuation is estimated; the receipt labels both and gives the observed dates, source and any supported range. It is not cash credit or guaranteed savings.

A weekly observation is never divided directly by a monthly subscription price. Only a validated full week with an explicit repeated-workload/full-use assumption may be converted to a monthly equivalent, using weekly API dollars × 52 ÷ 12. A partial test window gets no monthly ratio. Price alone never determines allowance, and historical plans are not mapped to current tiers by price.

Model charts use Epoch's model-level capability score. Our high-effort coding runs do not establish an effort-specific ECI score; we draw no effort trajectory without distinct sourced capability observations. Current time figures are median wall time per included attempt, not elapsed time to working code across retries.

This release includes six Claude task cohorts with 109 included passing attempts. Grok usage-derived results are withheld pending source clearance. Historical calibration experiments are a separate dataset and do not establish current subscription capacity.

What we leave out

We use no data from other comparison sites, show no lab logos, and publish totals and this method rather than the raw transcripts of our runs. Nothing here is endorsed by Epoch AI or any lab.

OpenAI measurement surface

Current OpenAI subscription measurement targets Codex Desktop only. No current Desktop capacity has been measured. The October 20 historical Plus scenario used Codex CLI coding work under a historical Plus fee assumption; it does not measure ChatGPT web chat or current Desktop capacity.

Google public limits

PUBLIC SOURCE / NOT TESTED. These facts describe personal-account Gemini Apps on web/mobile, not Gemini API, CLI/Code Assist, Antigravity, Workspace or Flow credits. No Google subscription, account probe or model call was used.

Google describes a five-hour refresh subject to a weekly cap. Prompt complexity, chat length, selected model and features affect use; no fixed prompts/day or universal reset timestamp was established. These multipliers are not equal API dollars or tokens. The Ultra limit variant belonging to a catalogue price is unverified, so no new priced offer is inferred.

Thinking labels are Low, Medium and High, default Low. App/model labels can change or route automatically; they are not pinned API snapshots or effort-matched intelligence observations. Context windows are not weekly token allowances. Google currently describes Deep Think as being replaced by Gemini 4 Argon; no old fixed daily Deep Think limit is carried over. API-dollar capacity, tasks/week and task share remain unavailable.

Sources checked 2026-11-04: usage and context, model and thinking labels, US plan comparison, Deep Think status, separate credit rules.