Value Index audit contract · audit checked 2026-07-20

Value Index methodology audit contract

Coding subscriptions are modeled as metered headroom, not flat unlimited plans. This is the receipt trail for how a ranked $/M-token number moves from native vendor meter to normalized Value Index row.

The Value Index does not treat “unlimited,” credits, effort, requests, five-hour windows, qualitative quotas, or Auto/API pools as interchangeable. It keeps the native meter visible, accepts only current official sources for plan facts, then applies a dated token-normalization policy with confidence labels and caveats where vendors do not publish exact quotas.

Source-freshness proof · overdue.

Status: stale / weekly audit overdue

Source audit run 2026-07-01; oldest scoped verification 2026-07-01; 11 scoped records; build checked 2026-07-20; canonical artifact src/data/cron/pricing-refresh.json; artifact present / validated / overdue. Value Index snapshot/recompute date remains 2026-07-20.

Status

Stale proof. The validated source audit ran 19 days before this build, beyond the 7-day weekly tolerance. This page refuses a current source-freshness claim.

Artifact

Canonical path: src/data/cron/pricing-refresh.json. Run 2026-07-01T09:15:45.205Z; oldest scoped verification date 2026-07-01.

Coverage

11 scoped records validated from the artifact. Artifact summary: “Verified 11 scoped provider/plan prices and headline limits from official sources on 2026-07-01.”

Watch areas

Official-source watch areas that must stay source-bound before any ranking claim is promoted.

Refusal boundary

This artifact is a bounded source-fact audit only. It does not prove the Value Index snapshot was recomputed; the Value Index snapshot/recompute date remains 2026-07-20. Because the audit is overdue, no current-ranking freshness claim is made.

Canonical normalization contract · audit checked 2026-07-20

Each contested meter keeps its official receipt, native unit, refused inference, Value Index treatment, and confidence label together. Plan facts stay official; token/value outputs are derived only where the contract says they can be.

GitHub Copilot · AI Credits

official + derived
Meter / vendor
GitHub Copilot · AI Credits
Meter class
Official credit meter
Native unit
Monthly AI Credits: Pro 1,500, Pro+ 7,000, Max 20,000; 1 AI Credit = $0.01. Paid completions and next-edit suggestions remain unlimited while chat, CLI, cloud agent, Spaces, Spark, and third-party coding agents consume credits.
Normalization rule
Keep the credit allowance as a plan fact, then convert documented credit-priced model usage through the official Copilot model-pricing reference when estimating token-equivalent headroom.
Refused inference
Do not turn an AI Credit allowance into a vendor-promised token bucket, and do not treat unlimited completions as unlimited premium chat, CLI, or agent work.
Value Index treatment
Core agentic coding evidence. Credit allowances are visible receipt facts; token value is derived and carries the credit-meter caveat.
Confidence / source-health
official + derivedOfficial docs plus June 2026 changelog; audit checked 2026-06-24.

Cursor · Auto/Composer + API pool

bounded estimate
Meter / vendor
Cursor · Auto/Composer + API pool
Meter class
Split pool / routing meter
Native unit
Individual plans expose Auto/Composer mechanics separately from an API usage pool.
Normalization rule
Model API-priced usage through documented pricing; keep Auto/Composer as bounded or opaque headroom when the routing mix is not published.
Refused inference
Do not collapse Auto routing into a fixed flagship-token pool, and do not assume overages are automatic.
Value Index treatment
Core agentic coding evidence. API pool math can be derived; Auto/Composer value remains separately labeled rather than silently blended.
Confidence / source-health
bounded estimateCursor canonical docs/pricing; audit checked 2026-06-24; older help URLs demoted if they resolve as generic shells.

Claude Code · Pro/Max unpublished caps

dashboard-visible / unpublished / unknown
Meter / vendor
Claude Code · Pro/Max unpublished caps
Meter class
Subscription headroom / dynamic limits
Native unit
Pro/Max subscription headroom with exact public prompt/token caps unpublished as stable plan facts; local /usage dollars are approximate and not authoritative subscription billing.
Normalization rule
Normalize Claude Code headroom as relative/dashboard-bounded usage, using public Anthropic/API pricing only for reference comparisons where the source supports it.
Refused inference
Do not infer exact public token entitlements from plan names, dashboard behavior, or /usage output.
Value Index treatment
Core agentic coding evidence with an explicit unpublished-cap caveat; no fixed public token bucket is claimed.
Confidence / source-health
dashboard-visible / unpublished / unknownAnthropic pricing and Claude Code costs docs; audit checked 2026-06-24.

Google Gemini Code Assist → Antigravity

bounded estimate
Meter / vendor
Google Gemini Code Assist → Antigravity
Meter class
Deprecated individual tiers / qualitative quotas
Native unit
Gemini Code Assist individual IDE/CLI tiers stopped June 18, 2026 and migrate to Antigravity; Antigravity quota language is qualitative, five-hour, weekly, and settings-visible rather than exact public prompt entitlements.
Normalization rule
Use Code Assist deprecation and quota pages only as dated historical receipts; treat Antigravity quota headroom as bounded/opaque unless current Antigravity sources publish exact entitlements.
Refused inference
Do not carry old exact Code Assist quotas forward as exact Antigravity prompt quotas.
Value Index treatment
Core Google coding evidence is dated to the migration boundary; Antigravity rows stay bounded or unpublished until official exact quotas exist.
Confidence / source-health
bounded estimateGoogle deprecation, quota, Antigravity plans/pricing/change sources; audit checked 2026-06-24.

Replit · Agent effort

secondary evidence / bounded estimate
Meter / vendor
Replit · Agent effort
Meter class
Secondary app-builder effort meter
Native unit
Replit Agent uses effort-based pricing; Agent interactions can be billable, including Plan Mode where sourced; exact per-task cost is runtime/dashboard-only.
Normalization rule
Preserve effort as an opaque native meter and translate it only as secondary app-builder evidence with spend-control caveats.
Refused inference
Do not treat effort as tokens, do not infer a stable per-task cost, and do not blend app-builder effort into professional agentic coding rankings.
Value Index treatment
Secondary app-builder evidence only; segmented from the core agentic ranking and not blended into default Value Index verdicts.
Confidence / source-health
secondary evidence / bounded estimateReplit pricing, billing, spend-control, and effort-pricing sources; audit checked 2026-06-24.

Source hierarchy: current official sources outrank everything else

For plan facts, current vendor pricing pages, product docs and changelogs outrank launch blogs, stale comparison pages, generic marketing copy, and community reports. Community anecdotes can explain sentiment or buyer pain; they cannot establish a quota, price, meter, or included entitlement.

  • Prefer current official docs, pricing pages, admin/billing docs, and changelogs.
  • Use launch blogs only when they are the current vendor announcement for a billing change, and demote them if docs/pricing later conflict.
  • Do not use community reports as plan-fact authority; they are sentiment evidence only.

Copilot usage-based billing ↗Cursor pricing ↗Claude Code costs ↗

Meter normalization: preserve the native unit before converting

Every row starts from the vendor's native meter — credits, API usage, requests, qualitative quotas, effort, local/cloud tasks, or rolling windows — and only then becomes estimated tokens. A normalized token number is a comparison instrument, not a vendor promise that the plan includes a fixed public token bucket.

  • Native meter stays visible in the row and provider detail before token conversion.
  • Direct API or vendor overage pricing is used as the reference when the vendor documents token-priced or pass-through billing.
  • If the vendor publishes ranges or policy language rather than exact counts, the row remains bounded-estimate or unpublished/opaque.

Copilot model pricing ↗Cursor models and pricing ↗Replit AI billing ↗

Routing assumptions: model menus and Auto pools are not one model

When a subscription can route across models, the Value Index separates documented API-agent usage from Auto/Composer-style pools and labels any unknown routing mix. The Flagship/Efficient toggle tests documented multipliers or model swaps, but it does not claim to know an unpublished Auto routing distribution.

  • Cursor API usage is separated from Auto/Composer pool mechanics.
  • Copilot premium surfaces are modeled through AI-credit/token-based billing mechanics when docs say those surfaces consume that meter.
  • Unknown routing mix stays a caveat instead of being silently collapsed into the flagship model.

Cursor pricing ↗Cursor models and pricing ↗Copilot billing docs ↗

Cache, compaction, and window treatment: hidden cost drivers stay in the caveat

Agentic coding cost is not just output tokens. Cache reads/writes, long-context compaction, tool results, failed or exploratory turns, and rolling-window throttles can decide whether a subscription has enough usable headroom. The model includes these only where a specific source supports the mechanic.

  • Claude Code /usage dollar figures are local estimates and are not authoritative Pro/Max subscription billing.
  • Replit Agent is effort-based secondary app-builder evidence; interactions can be billable, including Plan Mode where sourced, with spend controls available for guardrails.
  • Qualitative or settings-only quota language stays bounded/opaque instead of becoming a fixed public token entitlement.

Claude Code costs ↗Replit AI billing ↗Replit spend controls ↗

Confidence bands: precision is earned, not implied

The confidence label is part of the result. “Official” means the plan fact itself is published by the vendor. “Derived” means the site computes from official inputs. “Bounded estimate” means the vendor publishes a range, multiplier, qualitative policy, or dashboard-visible class but not a precise public count. “Dashboard-visible / unpublished / unknown” means no exact public quota should be inferred.

  • Color is never the only confidence carrier; labels and caveats carry the meaning.
  • A precise-looking $/M-token figure can still sit on an estimated quota; the caveat must travel with it.
  • Unpublished exact quotas remain unknown even when a plan has a visible dashboard for the signed-in buyer.

Claude Code costs ↗Antigravity plans ↗Replit AI billing ↗

Unknown quotas: no false precision for Claude Code, Antigravity, or effort meters

Claude Code, Antigravity, and Replit effort meters are explicitly labeled bounded or dashboard-visible / unpublished / unknown where exact public subscription quotas are not derivable from current docs. A row can still be compared with bounded assumptions, but the Value Index must not imply a vendor-published fixed token entitlement that does not exist.

  • Claude Code Pro/Max: exact public subscription caps are not published; local /usage dollars are not authoritative Pro/Max billing.
  • Antigravity: post-migration quotas are qualitative or settings-visible, so old Code Assist individual quotas are not carried forward as exact prompt entitlements.
  • Replit: effort-based app-builder evidence stays segmented from the core professional agentic ranking.

Claude Code costs ↗Antigravity plans ↗Replit effort-based pricing ↗

How to audit a ranked number

Start at the row, open the source links, then check this contract in order: source hierarchy, native meter, routing assumption, hidden cache/compaction/window cost, confidence label, and caveat. If any step is unpublished, the row must stay labeled as an estimate instead of gaining false precision.

Return to the ranked Value Index →