Methodology
Every number in TokenLens is derived from a documented calculation. This page explains each one.
Token cost
The base formula for API cost given a fixed input/output token count and a model with published per-token pricing:
cost = (input_tokens / 1,000,000 × input_price_per_1M)
+ (output_tokens / 1,000,000 × output_price_per_1M)Cost per 1M tokens
To normalise across models, we compute an effective blended cost per 1M tokens using the default output ratio (35% output, 65% input):
cost_per_1M = (1 - output_ratio) × input_price_per_1M
+ output_ratio × output_price_per_1MThis allows apples-to-apples comparison across models with different input/output price splits. The default ratio of 0.35 reflects typical coding workloads.
Your cost per 1M tokens
For a subscription, the rate that matters is not the model’s list price — it is what the monthly fee works out to across the tokens the plan actually allows:
your_cost_per_1M = monthly_price / estimated_monthly_cap × 1,000,000
Plans with no published cap have no derivable rate, so the comparison grid shows a dash rather than a number. Pay-per-token plans are billed at the model rate directly.
Token cap estimation
A plan only gets a cost per 1M token if its vendor publishes a limit we can convert into tokens. The first available source wins:
1. monthly_tokens_cap (stated directly) 2. monthly_credit_usd / rate (published $ credit pool) 3. weekly_tokens_cap × 4.33 (weeks per month) 4. daily_tokens_cap × 30.4 (days per month) 5. usage_multiplier × cap(referenced_plan) 6. nothing published → no cost per 1M
Multiplier resolution is capped at depth 3 to prevent circular references. All derived caps are labelled with their derivation method.
Step 6 is the common case. At the last refresh, GitHub Copilot was the only vendor publishing a convertible limit — a dollar credit pool of $15, $70 or $200 a month, which divides by the blended rate of the cheapest included model to give a token budget. Every other paid plan sells an allowance described as “extended”, “generous”, or a multiple of another unquantified plan.
Those plans are listed with their price and their models, and the cap column says what the vendor actually said. They are not given an invented cap so they can be ranked. A multiplier whose base plan publishes nothing resolves to nothing, not to a guess.
Dev-Hour Equivalent (DHE)
A DHE represents the token cost of approximately one hour of AI-assisted developer work, anchored to a specific task archetype. Each benchmark defines:
token_budget = { input_tokens, output_tokens }
assumed_human_work = 60 minutesThe four benchmark archetypes are:
- Light debug — small bug fix with diagnosis, patch, and unit test (25K in / 15K out)
- Medium feature — multi-file feature with validation and tests (80K in / 60K out)
- Large refactor — module restructure with regression tests (180K in / 120K out)
- Long-context review — large codebase review with patch notes (220K in / 70K out)
To account for real-world overhead, token budgets are multiplied by efficiency factors:
adjusted_tokens = budget × verbosity_multiplier × retry_multiplier dhe_cost = token_cost(adjusted_input, adjusted_output, model)
A verbosity_multiplier of 1.2 means the model produces 20% more tokens than the theoretical minimum. A retry_multiplier of 1.1 means 10% of requests are retried.
Scenario modeling
The scenario calculator projects monthly token usage and cost from usage parameters:
total_tokens = hours_per_day × days_per_month × tokens_per_hour output_tokens = total_tokens × output_ratio input_tokens = total_tokens × (1 - output_ratio) api_cost = token_cost(input_tokens, output_tokens, model)
Subscription viability is assessed by comparing api_cost against the plan's monthly price, and total_tokens against the estimated token cap.
Confidence tiers
- High — sourced directly from official pricing pages with explicit token counts.
- Medium — derived from documented limits or community reporting with cross-verification.
- Estimated — inferred from marketing language, multipliers, or historical patterns. Treat with caution.
Data freshness
All data is manually maintained. Each plan record includes last_verified and effective_from dates. AI pricing changes frequently — always verify against the source URL before making purchasing decisions.