Subscription or API?

Standardized workload

What can €20 actually buy?

At the 12 Aug 2026 snapshot, standard consumer plans were close to $20 per month and some premium tiers reached $200. Prices and conditions can change; first-party plan pages are linked below.

Using the costs observed across the 60 exercises common to all configurations, we estimate how much rubric-score-equivalent workload the same budget would purchase through an API. Prices include input and billable output tokens, including reported reasoning tokens. The conversion €1 = $1.1392 is the ECB reference rate for 23 Jun 2026 ↗.

Data & methodDownload the score-and-ranking CSV ↗ and the token-and-cost CSV ↗. Workload is calculated as budget divided by observed mean euro cost per exercise.

Gemma 4 31B IT is excluded because its benchmark campaign rate was recorded as zero, which is not a universal market price. The scale difference is striking: roughly 49 Sol exercises versus 3,555 with DeepSeek V4 Flash for 1.48 fewer mean score points.

Standardized unit. These counts do not imply a user submits only full university examinations. They provide a common unit for explanations, corrections, alternate derivations and follow-up questions.

Annual perspective

A €20 monthly plan corresponds to €240 per year

Data & methodAnnual capacity applies the same reproducible formula with a €240 budget. Inputs: historical cost CSV ↗ and the dated ECB exchange-rate source ↗.

These values describe API capacity under benchmark-like workloads. They do not imply that subscriptions permit the same volume: message limits are not expressed as equivalent API tokens and may vary by model, context length and demand.

Real workflows also include uploaded files, revisions, long conversations and verification requests. Each increases processed context and may materially raise API expenditure. The table is therefore a baseline before accounting for the wider workflow.

Reference threshold

Where is the subscription break-even point?

At €20, API expenditure reaches the reference price after approximately 49 Sol exercises, 59 Fable exercises, 1,501 DeepSeek V4 Pro exercises or 3,555 DeepSeek V4 Flash exercises.

These thresholds show when theoretical API expenditure exceeds €20—not whether a subscription can sustain the same workload within its message, session or weekly limits.

The break-even workload differs by orders of magnitude across models. Subscription attractiveness therefore depends strongly on the API price of the underlying model and on the actual number of substantial requests made each month.

Marginal value

How much is the final percentage point worth?

GPT-5.6 Sol reaches 98.72%. DeepSeek V4 Flash reaches 97.24% while costing approximately 72 times less per exercise.

Sol has the higher observed benchmark mean. Whether the additional 1.48 points justify the price depends on the cost of error, not only the cost of the request.

Is DeepSeek V4 Flash the economic optimum?

Within the common panel, Flash presents a strong observed quality–cost combination: 97.24% at approximately €0.0056 per exercise. V4 Pro costs around 2.4 times more while scoring 96.71%, so it is economically dominated on this workload. Product labels such as Pro, Max or Ultra do not guarantee a superior quality–cost relationship for every domain.

Reasoning allocation

Is thinking worth its additional cost?

DeepSeek V4 Flash without thinking costs about €0.0017 per exercise but falls from 97.24% to 86.31%. Thinking costs roughly 3.2 times more, yet the absolute price difference remains small.

This differs sharply from the Sol comparison, where a 72-fold increase buys 1.48 points. The useful question is not whether thinking is more expensive, but how much additional accuracy each euro purchases.

Why cheaper does not always mean better value

Haiku, GPT-4.1 and Sonnet cost less than frontier models but score roughly 65–72%. DeepSeek Flash is both cheaper and substantially more accurate. A low nominal price is therefore insufficient: correction, regeneration and external verification can make an apparently inexpensive model the costlier system.

Product distinction

Subscriptions and APIs sell different things

As documented on 12 Aug 2026, ChatGPT Plus ↗ was separate from API billing and subject to product limits. Claude Pro ↗ likewise separated subscription and API access, while Google AI Pro ↗ bundled Gemini with storage and other services. Current terms should be checked on those first-party pages.

Their prices can be compared, but their value cannot be reduced to token costs alone.