Token-Max/docs/VERIFICATION.md

Provider Verification Log

This document is the audit trail for the curated plan data in data/coding-plans/. Every provider was checked against its official pricing, quota or legal pages during the September 18–19, 2026 audit pass. Facts that could not be confirmed against a primary source are explicitly labelled unknown or low-confidence and carry that label in estimatedTokenBudget.estimateMeta.

• Machine-readable provenance: estimateMeta on every tier (sourceUrl, sourceQuote, sourceType, confidence, verifiedAt, basisModel, cacheAssumption).
• Invariant enforcement: npm run validate-data (runs in CI with lint, test and build).
• Multi-account policy: stackingPolicy + evidence note on every plan, surfaced by Dangerous Dave mode and the TOS page.

Verification status by provider

#	Provider	Official source checked	Status	Notes / open items
1	Aider	aider.chat	Verified	Free OSS, BYOK only; no vendor pricing.
2	Alibaba Cloud (Token Plan)	alibabacloud.com token-plan docs	Verified, derived estimates	Prices/credits official; credit→token coefficients unpublished, derived from the official qwen3.6-plus worked example. Coding Plan (request-based) is a separate, phased-out product.
3	Amazon Q Developer	aws.amazon.com/q/developer/pricing	Verified, low-confidence Pro	Free = 50 agentic requests/mo (official); Pro = $19 with no published cap → scenario band (21/105/210M).
4	Anthropic API	docs.claude.com pricing + commercial terms	Verified	Sonnet 5 $2/$10, cache reads 0.1x, writes +25%, batch −50%, spend tiers Start/Build/Scale; no training on Customer Content.
5	Augment Code	augmentcode.com/pricing	Verified	$20/$100 pools; flat 40% LLM fee; 50 seats; top-ups 12 months.
6	BytePlus ModelArk	docs.byteplus.com ModelArk 1925114/2165245 + AI terms	Verified, quota docs conflict	Lite $10/Pro $50; docs quote ≈1,900 req/5h, 12,000/wk, 24,000/mo for Lite (older FAQ lower); no token conversion; Customer Data not used for training.
7	Claude Code	claude.com/pricing + docs.claude.com costs	Verified, low-confidence estimates	Prices and 5-hour/weekly shared-pool mechanics verified; Anthropic publishes no token numbers → 12/60/240M research estimates. Faros AI & Morph empirical telemetry confirmed Max 20x tier asymmetry (20x in 5h burst, but only ~6x weekly) and shared capacity pool across CLI, Claude.ai, and Cowork.
8	CommandCode	commandcode.ai/pricing + terms	Verified	Full 5-tier matrix verified: Go $1 ($10 credits, 250/5h, 1k/7d), GOAT $10 ($70 credits, 500/5h, 2k/7d), Pro $20 ($80 credits, 750/5h, 3k/7d), Max 10x $100 ($150 credits, 1.5k/5h, 6k/7d), Max 20x $200 ($300 credits, 3k/5h, 12k/7d); dual standard & premium pools; one account per person (stacking prohibited).
9	Cursor	cursor.com/pricing + FAQ	Verified, pool sizes unpublished	Pro+ = 3x and Ultra = 20x Pro limits (official); Pro pool modelled at ~10M tokens. Reseller purchases unauthorized.
10	DeepSeek API	api-docs.deepseek.com pricing	Verified, fixed	Previous USD figures did not match official CNY rates; now quotes ¥9/¥27 with cache-hit ¥0.15–0.30 and 50% off-peak.
11	Fireworks.ai	fireworks.ai/pricing + account quotas docs	Verified	$1 credits; 6,000 RPM account-wide cap with card. Exact free RPM unpublished.
12	GitHub Copilot	github.com/features/copilot/plans	Verified	1 credit = $0.01; Pro $15, Pro+ $70, Max $200 total credits; Flex allotment is variable → conservative estimates use base credits only.
13	Google AI Studio	ai.google.dev pricing + rate limits	Verified	Flash $0.75/$3.75 promotional through 2026-12-31; tier caps $250/$2,000/$20,000+; free data trains, paid does not.
14	Google Antigravity	antigravity.google/pricing + one.google.com plans	Verified, fixed	Ultra no longer $249.99; official Google AI Ultra 5x $99.99 and 20x $199.99; quotas 5-hour refresh until weekly cap. Research estimates rebased on the 250K–900K OSINT agent-task band (Pro 75M floor / 165M / 270M).
15	Groq API	console.groq.com rate limits	Verified, fixed	Console tables are Developer-plan base limits; Free is lower and per-model.
16	Kilo Code	kilo.ai/pricing + teams + kilo-pass	Verified	$15/user platform fee, inference at provider rates with no markup; Kilo Pass $19/$49/$199 with up to 50% bonus.
17	Kimi Code	kimi.ai/help membership pricing	Verified	$19/$39/$99/$199 (annual $15/$31/$79/$159); shared credit pool + separate 5h/weekly Kimi Code limit; credit counts unpublished.
18	Kiro	kiro.dev/pricing + FAQ	Verified	Credit multipliers published; no rollover; third-party automation harnesses not permitted.
19	Lovable	lovable.dev/pricing + subscription docs	Verified	$25/$50 (annual $21/$42); credits expire after 2 months; unlimited workspace members.
20	Meta Model API	dev.meta.ai pricing/rate limits	Verified	Now serves Muse Spark 1.3/1.2; Standard $1.25/$4.25 no-training, Contributor $0.10–$0.20 with training rights; hosted Llama tiers superseded.
21	Meta Muse Code	developer.meta.com Muse Code	Verified	$5/$15/$50; High = 5x, Power = 20x Everyday (official); keys limited to Muse Code.
22	MiniMax Token Plan	platform.minimax.io token-plan docs	Verified, token quota unpublished	Prices, windows, agents and 1,000 credits = $1 verified; monthly token quotas not published → research estimates.
23	Mistral API	mistral.ai/pricing	Verified, fixed	Studio includes $10/mo credits; training is opt-out (previous "No" was wrong).
24	Ollama Cloud	ollama.com/pricing	Verified, fixed	Team = $1,000 credits and 10 streams (not $1,600/50); one account per person; peak Mon–Fri 12–18 UTC.
25	OpenAI API	platform.openai.com docs + enterprise privacy	Verified, fixed	Sol $2/$10 and Luna $0.10/$0.60; ZDR is not default (30-day retention, ZDR on request); spend caps $100–$200k.
26	OpenAI Codex	developers.openai.com/codex/pricing	Verified, low-confidence estimates	Free/Go $8/Plus $20; Pro from $100 (5x/20x); no published token quotas → scenario estimates rebased on the 250K–900K OSINT agent-task band (Plus 20M floor / 44M / 72M). DevForth telemetry confirms observed window limits: Plus ($15/5h, $70/7d), Pro 5x ($75/5h, $350/7d), Pro 20x ($400/5h, $2,560/7d).
27	OpenCode	opencode.ai/zen + docs.opencode.ai/go + terms	Verified	Zen $20 minimum + $1.23 fee, zero markup; Go $10/mo open-model subscription covering 28 models across $15, $30, $60 monthly allowance pools with 20% 5h and 50% 7d rolling window exhaustion caps. Multiple-account circumvention prohibited.
28	OpenRouter	openrouter.ai/pricing + limits + terms	Verified	5.5% platform fee; free-model limits 20 RPM / 50–1,000 RPD; multiple accounts to bypass limits prohibited.
29	Replit	replit.com/pricing + terms	Verified	Core $20 ($18 annual), Pro $100 ($90); registering multiple accounts prohibited.
30	Tabnine	tabnine.com/pricing + terms	Verified	$39/$59 annual per user; BYO LLM unlimited; provider LLM +5% handling; no training on customer code.
31	Together.ai	docs.together.ai rate limits	Verified, fixed	Dynamic rate limits replaced the stale fixed 60 RPM/60k TPM claim.
32	Windsurf (Cognition)	windsurf.com/pricing + AUP	Verified	Free/Pro $20/Max $200/Teams $80+$40; credential sharing banned. Data-training policy not published → unknown. Estimates rebased on the 250K–900K OSINT agent-task band (Pro 75M floor / 165M / 270M); Max restored to a clean 5x Pro relation.
33	Z.ai GLM Coding Plan	docs.z.ai devpack overview/teamplan/usage-policy	Verified	Lite $18, Pro $80, Max $168 with official weekly allowance ceilings: GLM-5.3 (97M / 582M / 1,358M) and GLM-5.3-Flash (584M / 3,504M / 8,176M); concurrency caps (1 / 1-2 / 2+); 95% prompt cache; off-peak 50% credit rate outside Mon–Fri 14–18 UTC+8.

Discrepancy log (September 2026 audit)

Fixed against official sources:

• Z.ai estimates exceeded the officially published allowance ceilings and are now floor/midpoint/ceiling derived from the docs table.
• Alibaba Cloud used invented "1K credits = $1" and 1,000 tokens/credit arithmetic; replaced with official credit quotas and a documented qwen3.6-plus deduction basis, plus the missing Essential tier.
• Cursor Pro+/Ultra now match the official 3x/20x multipliers.
• Google Antigravity Ultra price split into the official 5x ($99.99) and 20x ($199.99) plans.
• Ollama Cloud Team credits/concurrency corrected ($1,000 / 10 streams) and model list refreshed.
• GitHub Copilot conservative budgets no longer assume the variable Flex allotment.
• Anthropic/OpenAI API token estimates now match their own pricing formulas and the 20k/1k cached-agent workload; OpenAI ZDR claim corrected.
• Amazon Q Free tier now uses the official 50-request cap; Pro is an explicitly low-confidence scenario band instead of 200M.
• DeepSeek CNY pricing, Mistral training opt-out, Groq Developer-plan limits, Together dynamic limits, Fireworks trial math, BytePlus internal anchor, Augment 40% fee, and four zero-budget paid tiers (Kilo Teams, OpenCode Zen, Tabnine ×2) were corrected.

September 19, 2026 — OSINT task-band recalibration

• Antigravity / Windsurf / OpenAI Codex opaque-quota estimates replaced the 150K-tokens-per-task basis (below the OSINT autonomous-task band and inconsistent with the WorkflowCalculator's 40-request task model) with the shared band in data/estimate-constants.json: 250K conservative / 550K midpoint / 900K optimistic tokens per autonomous task (interactive CLI sessions measure 60–240K; autonomous tasks 200–800K in + 30–100K out). Tier relations (5x/20x multipliers, per-seat mirrors) are preserved and every changed tier carries floor/midpoint/ceiling values plus estimateMeta provenance. Evidence: docs/TOKEN_ESTIMATE_VALIDATION.md.

Open items

• Claude Code / OpenAI Codex / Amazon Q Pro / Google Antigravity capacity is intentionally unpublished by the vendor; figures are research estimates with confidence: low (Antigravity/Codex now use the 250K–900K OSINT task band).
• MiniMax publishes prices and credit packages but not token quotas; third-party token figures conflict (~1.6B/mo claims vs our estimate) — verify in the console before trusting the estimate.
• Alibaba credit coefficients are only partially documented (one worked example); frontier-model deductions may differ from the qwen3.6-plus basis.
• BytePlus quota docs conflict across pages (≈1,900 vs 1,200 req/5h for Lite).
• Kimi / Kiro / Replit / Meta Model API rate-limit details are not fully published.
• Windsurf, Kiro, Replit, Kimi, MiniMax, BytePlus training policies are unpublished → TOS matrix shows Unknown.

Verification passes

2026-09-20 — External Sources & Empirical Usage Telemetry Audit Pass

Comprehensive investigation across 13 external pricing APIs, machine-readable datasets, and real-world engineering telemetry:

• Secondary Pricing Feed Ingestion & Auditor (scripts/audit-external-sources.mjs):
  • Integrated llmprice.com daily JSON feed (/assets/pricing-data.json), cross-validating 481 models and confirming exact first-party cache-write rates and long-context pricing rules for OpenAI, Anthropic, Google, and DeepSeek.
  • Added non-destructive CLI command npm run audit-sources to cross-check live pricing deltas without build-blocking failure points.
• Claude Code Quota Asymmetry Discovery (Faros AI & MorphLLM Telemetry):
  • Reverse-engineered and documented the 5-hour burst vs. weekly ceiling multiplier disparity: Max 20x provides 20x burst headroom in a 5-hour window but scales weekly capacity to only ~6x of the Pro tier.
  • Documented cross-surface capacity pooling across Claude Code CLI, Claude.ai, and Claude Cowork.
  • Added dedicated claude-code-max20x throttle profile to src/lib/throttle-profiles.ts.
• OpenAI Codex Observed Limit Windows (DevForth Telemetry):
  • Incorporated DevForth empirical tracking of OpenAI subscriptions: Plus ($15/5h, $70/7d), Pro 5x ($75/5h, $350/7d), and Pro 20x ($400/5h, $2,560/7d).
  • Added openai-codex-pro5x and openai-codex-pro20x throttle profiles to src/lib/throttle-profiles.ts.
• Enterprise Spend Benchmarking (DX 400+ Organization Study):
  • Verified ground-truth blended developer spend ($200–$600/month per seat) for hybrid IDE and CLI agent workflows.
  • Documented GitHub Copilot's June 1, 2026 token transition and promotional credit offsets ($30–$70/user/mo).
• Chinese Regional Token Plans Baseline (tokenplan.vip):
  • Verified domestic Chinese token subscriptions and off-peak rate multipliers across 41 platforms.

2026-09-19 — Adversarial Codebase Audit & 4-Phase Systemic Hardening

Comprehensive adversarial investigation and remediation across analytical engines, mathematical models, ingestion scripts, schema invariants, security controls, and UI presentation:

• Phase 1 (Critical & High Math, URL & Allocation Fixes - 91b4d6a):

  • Eliminated URL query parameter crash vectors in WorkflowCalculator.tsx with fallback chains for ?ctx= and ?quality=.
  • Replaced || with ?? across 5 core calculation files (pricing.ts, teams.ts, receipt-math.ts, reasoning.ts, WorkflowCalculator.tsx) to eliminate systemic falsy zero-price bugs where $0 cached prompt tokens or $0 free tiers were overridden by fallbacks.
  • Guarded knapsack against NaN budgets in computeDaveStacks and computeMixAndMatch.
  • Fixed free cache multiplier inversion (cachedInput === 0 correctly evaluates to 0).
  • Prevented wildcard model query explosion (matchesPlanModel rejects queries < 3 characters).
  • Implemented Hamilton-Hare largest remainder algorithm in teams.ts for exact seat allocation invariant.

• Phase 2 (Medium Severity Hardening & Input Sanitization - 2911e19):

  • Protected throttle.ts sprint duration, concurrency, and cadence against NaN, negative values, and zero-division.
  • Clamped URL mix token parsing and inputs in MixOptimizer.tsx.
  • Added sanitizeYamlScalar to exporters.ts to strip control characters and newlines, preventing YAML injection attacks.
  • Hardened log-parser.ts with safeTokenNumber against NaN/non-numeric log tokens, added 25MB parser limit, and enforced 15MB file upload limit in SessionReceipt.tsx.

• Phase 3 (Data Ingestion & Pipeline Invariants - c052c9b):

  • Made fetch-benchmarks.mjs and fetch-models.mjs fail fast with explicit errors on API failure, refusing to clobber existing datasets with empty outputs.
  • Added .sort() to fs.readdir(PLANS_DIR) in build-data.mjs for 100% deterministic plan builds across OS environments.
  • Synchronized _schema.json with strict validation rules (required non-empty limits, models, and non-null estimatedTokenBudget).
  • Added contextual multi-tier TOS classification in tos.ts for mixed policies (free trains vs paid shielded).

• Phase 4 (UI Display Polish - e4e66aa):

  • Replaced misleading $0+ and Free+ labels in PlanGrid.tsx with "Free tier available", "Free", or "Enterprise / Custom".
  • Handled $0 and null pricing in TokenTranslator.tsx with dedicated empty states rather than falling back to $20/mo.

Commands run from repository root, all green:

1. npm run validate-data — 33/33 plans valid against strict schema.
2. npm test — 97/97 unit tests pass across 8 test suites.
3. npm run lint — 0 errors, 0 warnings (oxlint).
4. npm run build — TypeScript and Vite production build succeeds.

2026-09-19 — Multi-Model Mix Optimizer & Shareable URL State

Delivered the Multi-Model Mix Optimizer tool and workflow URL permalinks:

• Shareable Scenario Permalinks: Integrated useSearchParams into src/components/budget/WorkflowCalculator.tsx to encode daily/session parameters, cache rate, and pipeline mode into the URL hash, with a one-click clipboard share button.
• Mix Optimizer Page: Created src/pages/MixOptimizer.tsx (#/optimizer) enabling developers to construct custom model pipelines across all 440+ models, with Autonomous Agent, Daily Driver, and Lean Open presets.
• Multi-Model Pool Drain Engine Integration: Evaluated all 33 subscription plans against custom model mixes, computing pool utilization percentages, pay-per-use overages, and direct API savings.
• Navigation & Routing: Added /optimizer route in src/App.tsx and header navigation item in src/components/layout/Header.tsx.

Commands run from repository root, all green:

1. npm run validate-data — 33/33 plans valid.
2. npm test — 46/46 unit tests pass.
3. npm run lint — 0 errors, 0 warnings (oxlint).
4. npm run build — TypeScript and Vite production build succeeds.

2026-09-19 — Quota calculation methodology evaluation & pool drain engine

Forensic comparative audit of competitor and peer methodologies (DeepFrugal, AI-10-USD) versus Token-Max's pricing logic. Implemented architectural enhancements:

• Shared Constants: Added timeOfDayBlend (80% off-peak, 20% peak) and agentCachePresets to data/estimate-constants.json and src/lib/estimate-constants.ts.
• Time-of-Day Blending: Added calculateTimeBlendedCost() in src/lib/pricing.ts to model off-peak discounts (DeepSeek, Z.ai).
• Multi-Model Pool Drain Engine: Implemented calculatePoolDrain() in src/lib/pricing.ts modeling sequential subscription quota exhaustion across mixed model pipelines with pay-per-use overage billing.
• Model-Specific Tier Allowances: Extended _schema.json and populated modelAllowances in data/generate_plans.py for router plans (CommandCode, OpenCode Go).
• Hybrid Pipeline Workflow Mode: Enhanced src/components/budget/WorkflowCalculator.tsx with a Single vs. Hybrid Pipeline toggle (75% Workhorse / 25% Frontier) displaying subscription pool usage and overages.

Commands run from repository root, all green:

1. python3 data/generate_plans.py — regenerated all 33 plan files with modelAllowances.
2. node scripts/build-data.mjs — rebuilt public/data/plans.json.
3. npm run validate-data — 33/33 plans conform to schema invariants.
4. npm test — 46/46 unit tests passing (+4 new tests for pool drain & time blending).
5. npm run lint — 0 errors, 0 warnings (oxlint).
6. npm run build — TypeScript and Vite production build succeeds.

2026-09-19 — Comprehensive live web crawl & quota validation pass

Systematic multi-batch web crawl of all 33 coding plans using Firecrawl MCP (firecrawl-firecrawl_scrape) and live search verification against official documentation and pricing portals. Detailed audit log stored in docs/CRAWL_AUDIT_2026.md.

Key updates and validations:

• Annual Pricing Ingestion: Added official annual discounted rates: Cursor Pro+ ($576/yr), Cursor Ultra ($1,920/yr), Claude Code Pro ($200/yr upfront), Replit Core ($216/yr), Replit Pro ($1,080/yr).
• DeepSeek API Specifications: Reconciled official DeepSeek docs (https://api-docs.deepseek.com/quick_start/pricing), validating DeepSeek-V4.1-Flash ($0.15/$0.60 off-peak) and DeepSeek-V4-Pro ($0.66/$1.98 off-peak) with 1M context windows, 384K output max, and 2,500/500 concurrency ceilings.
• Provider Pricing URLs: Updated explicit canonical pricing links for Together AI (https://www.together.ai/pricing) and Fireworks AI (https://fireworks.ai/pricing).
• Quota & Tier Sanity: Confirmed all 33 plans and tiers maintain valid non-zero token budgets, models lists, and limits definitions per schema.

Commands run from the repository root, all green:

1. python3 data/generate_plans.py — regenerated all 33 plan files with updated annual prices and DeepSeek specs.
2. node scripts/build-data.mjs — consolidated public/data/plans.json and updated freshness metadata.
3. npm run validate-data — all 33 plan files pass 100% schema invariant checks.
4. npm test — 42/42 tests pass across pricing, cache blend, stacking, and TOS auditing.
5. npm run lint — 0 errors, 0 warnings (oxlint).
6. npm run build — TypeScript check and Vite static production build succeed.

2026-09-19 — token-estimate recalibration (Phases 0–4)

Commands run from the repository root, all green:

1. python3 data/generate_plans.py — regenerated all 33 plan files; re-running produced no diff (idempotent).
2. node scripts/build-data.mjs — consolidated public/data/plans.json; the refactored models.json agent blend was byte-identical to the pre-refactor artifact.
3. npm run validate-data — all 33 plan files pass schema invariants (midpoint ≥ conservative, optimistic ≥ midpoint on every rebased tier).
4. npm test — 42 tests pass (pricing blend modes, cache-write pricing, stacking, TOS).
5. npm run lint — 0 errors, 0 warnings.
6. npm run build — TypeScript + Vite production build succeeds.

Evidence and OSINT sources: docs/TOKEN_ESTIMATE_VALIDATION.md.

Re-verification checklist

1. Visit each url in data/coding-plans/*.json; confirm prices, tier names and quota units.
2. Update estimatedTokenBudget figures only from published numbers; otherwise keep the research label and update estimateMeta.verifiedAt.
3. Re-check dataTraining, ipIndemnity, stackingPolicy and stackingPolicyNote; a stacking claim that is prohibited requires a quote.
4. Run npm run validate-data && npm test && npm run lint && npm run build.
5. Record the pass here and bump LAST_VERIFIED in data/generate_plans.py.

2026-09-20 — External OSINT Comparative Audit & Runtime Telemetry Integration

Findings from comparative analysis across 4 primary external sources:

1. sites.diy (Bilal Bakht Ahmad, May 1, 2026): Live MITM proxy logging of Claude Code, Codex, Kimi, GLM sessions measured 92.4% cache read / 2.4% fresh input / 5.2% output (97.47% cache hit rate on context). Validated Token-Max's deepAgent: 0.92 cache preset. Documented 5-hour rolling limits vs weekly hard ceilings (Claude Pro $4.75/5h, $38/wk = 8 max sessions/wk, 7.6× subsidy, 26.9M tok/mo ceiling on Opus 4.7; Codex Plus $22/5h, $134/wk = 26.8× subsidy, 250M tok/mo ceiling on GPT-5.5).
2. tokenplans.dev (Krzysztof, September 2026): Hand-verified pricing ledger of 89 plans across 31 providers tracking Req / $1, confidence levels (Exact, Proxy ~, Coarse ≈), and borrowed-cap flags.
3. sessionwatcher.com (Soren Starck, 2026): Turn-level (~4,800 tokens) vs agent run (50K–100K+) sizing; prompt caching discounts (90% Anthropic, 50% OpenAI); median developer uses <40% of plan allowance; rate-limit lockout wage economics ($75/hr × 2 hrs/wk lost = $600/mo productivity loss).

Enhancements implemented:

1. src/components/plans/TokenTranslator.tsx — added EMPIRICAL_SATURATION_DATA card comparing conservative usable floors (40h human workweek) against theoretical 24/7 saturation ceilings (sites.diy empirical proxy logs) with 5h cap, weekly cap, sessions/week, and subsidy multipliers.
2. src/components/budget/WorkflowCalculator.tsx — added interactive Rate-Limit Lockout Economics & Wage Risk card ($75/hr baseline) calculating weekly lockout exposure and monthly developer wage loss when daily demand exceeds 5-hour rolling capacity.
3. scripts/audit-external-sources.mjs — integrated https://api.tokenplans.dev/plans.json public teaser endpoint to automatically cross-reference tracked coding providers and flagship models.
4. docs/TOKEN_ESTIMATE_VALIDATION.md & docs/OSINT_USAGE_STATISTICS.md — added detailed citations and methodology alignment documentation.

Commands run, all green:

• npm run validate-data (422 entries across 34 providers pass invariants)
• npm test (109/109 tests pass across 8 suites)
• npm run lint (0 errors, 0 warnings across 71 files)
• npm run build (TypeScript + Vite build succeeds in 637ms)

2026-09-19 — Leaderboard quality-filter leak fix + Z.ai V3 price/ceiling update

Findings: with Quality Baseline = Top Frontier (≥75), Z.ai GLM tiers surfaced as "Multi-Model" rows because (a) matchesPlanModel matched "GLM-5.3-FlashX" via greedy substring before the base z-ai/glm-5.3 (AI-index 74.8), and (b) the unmatchedTiers fallback path skipped the quality gate entirely. Normalized-at-budget columns were additionally displayed without the tier's actual native allowance, misleading on tier cost vs yield.

Fixes:

1. src/lib/pricing.ts — exact-match priority + variant guard in matchesPlanModel (base names no longer match flash/mini/nano/micro/lite/small variants); added rawMonthlyTokens/rawMonthlyRequests to subscription rows; unmatched-tier fallback rows are now gated to minCodingIndex === 0 so unbenchmarked suites cannot leak into quality-filtered leaderboards.
2. src/lib/types.ts — ApplesToApplesOption.rawMonthlyTokens/.rawMonthlyRequests plus tier estimateMeta typing (matches _schema.json).
3. src/components/budget/LabDecisionEngine.tsx — leaderboard shows subscriptions' native token allowance under the normalized yield and "+$X unspent" under tier cost; explanatory tooltips updated.
4. data/coding-plans/z-ai.json + data/generate_plans.py + rebuilding public/data/plans.json — Pro $72 → $80, Max $160 → $168 (official V3), official weekly-ceiling limits added as limits + estimateMeta.sourceQuote evidence (up to 97M/582M/1,358M GLM-5.3 tokens per week on Lite/Pro/Max).

Commands run from the repository root, all green:

1. npm run validate-data — 33/33 plan files pass schema invariants.
2. npm test — 98/98 tests pass (8 suites), incl. new matchesPlanModel variant tests.
3. npm run lint — 0 errors, 0 warnings.
4. npm run build — TypeScript + Vite production build succeeds.
5. Simulation (six tiers @ $200, minCodingIndex=75): Z.ai rows = 0; top ranks are Gemini 3.8 Flash (CI 76.3), Kimi K3 (76.2), Claude Opus 5 (78). Unfiltered run: Z.ai Max resolves to Z.ai: GLM 5.3 (CI 74.8, native 2,927M, $168).

2026-09-19 — UX Phase 0: design-token AA hardening + shared primitives

Contrast audit (WCAG 4.5:1 for text) of semantic tokens on surface/steel chrome found two failures: --color-primary (was blood-500, 3.67:1) and --color-primary-light on the header steel gradient (4.19:1).

Fixes:

1. src/index.css — --color-primary: hsl(0 85% 55%) (4.71:1 on surface) and --color-primary-light: hsl(0 80% 64%) (4.94:1 on header steel). No component changes needed; all chrome reads the semantic tokens.
2. New shared primitives under src/components/shared/ (token-only styling): DataBadge.tsx, Card.tsx, TableShell.tsx (incl. Th/Td with sticky header), StatBlock.tsx, Toolbar.tsx, Tooltip.tsx — Phase 1–3 pages migrate onto these incrementally.
3. src/pages/TosAudit.tsx — migrated hero pill to DataBadge and all four metric cards to StatBlock as the wired-up reference usage.

Verification: lint 0/0, tests 98/98, production build clean; built CSS confirmed to carry the updated primary/primary-light values (#ee2b2b, #ed5a5a).

2026-09-19 — UX Phases 1–5 (grouped IA, shareable state, power-tool pass, Quiet theme, perf/PWA)

1. Phase 1 — Navigation & mobile: Header rebuilt with grouped IA (Calculate / Forecast / Data / Tools) as keyboard-navigable hover dropdowns; full-screen hamburger drawer below lg; footer quick-route chips; no setState-in-effect (links close menus on click; Escape closes).
2. Phase 2 — First-touch & state: useQueryState hook (shareable hash query + localStorage restore) wired into budget (?b=), quality baseline (?q=), lab (?lab=), unit (?unit=), cache (?cache=), engine mode (?mode=); "Copy Link" button reproduces the exact ranking; leaderboard shows top 3 with "See full ranking" expander; hero CTA smooth-scrolls to the Decision Engine.
3. Phase 3 — Power-tool pass: sticky table headers via new thead-sticky utility applied to ModelTable, LeaderboardTable, TrainingMatrix and the Decision Engine leaderboard; "Copy CSV" exports the visible ranking.
4. Phase 4 — Quiet theme: [data-theme='quiet'] token overrides (semantic tokens + chrome gradients) with ThemeToggle (persisted, respects prefers-color-scheme); Blood & Steel remains the default.
5. Phase 5 — Perf/PWA: all 12 routes lazy-loaded with Suspense fallback; header data-freshness pill (replaces the "Machine Spirit" place holder); production-only service worker (public/sw.js: cache-first hashed assets, network-first data JSON + index).

Verification: npm run lint 0 errors/0 warnings; npm test 98/98; npm run build clean with per-route chunks (core 90 KB gz, pages 5–16 KB gz, shared chart chunk 99 KB gz); dist/sw.js emitted. Known limits: Quiet theme is best-effort on third-party-styled surfaces; deeper per-page component migrations (card tables < md, wizard splits) scheduled for the next pass.

Per-Model Subscription Token Yields (2026-09-19)

Mechanism

• PlanTier.perModelTokenBudgets: per-model yields assuming the entire quota drains exclusively on that model; keys are lowercase substrings of real tier.models entries; optional basis (official-table, official-multiplier, list-price-credit, equal-rate) and confidence.
• scripts/validate-data.mjs enforces keys ⊆ tier.models, positive finite token values, monotonic mid/optimistic values, and the basis/confidence enums.
• computeApplesToApples renders one subscription row per (tier, matched model), resolving per-model budgets via resolveTierModelBudget (longest-key-first) and falling back to the tier pool otherwise. Stacking (Mix & Match / Dangerous Dave) stays tier-pool-based; no double-counting.
• Tier matching fixed to forward matching + per-tier per-model claims so "GLM-5.3-Flash" can no longer be stolen by the shorter "GLM-5.3".

Included plans (verified official sources via Firecrawl OSINT)

Plan	Basis	Source
z-ai	official-table: GLM-5.3-Flash = exact 3.03x GLM-5.3 (Lite 146-292, Max 2,047-4,095 M tok/wk)	docs.z.ai/devpack/overview
kiro	official-multiplier table (Auto 1.0x; Opus 2.2x, Sonnet 1.3x, Haiku 0.4x, Qwen3 Coder Next 0.05x, DeepSeek 0.25x, MiniMax 0.15/0.25x)	kiro.dev/docs/models
commandcode	list-price-credit via officially published at-cost token rates; GOAT/Pro per-model allowances (MiniMax M3 $47/$57, MiMo deals, DeepSeek V4.1 Flash $60/$70)	commandcode.ai/docs/resources/pricing-limits
cursor	list-price-credit on officially published per-model rates for both usage pools ($20/$60/$200 modeled pools)	cursor.com/docs/account/pricing
claude-code	list-price-credit with Fable drawn at official 50%-of-weekly-limits cap	claude.com/pricing
github-copilot	AI credits (1 credit = $0.01) drained at official per-token tables; 300/3,900/10,000 base credits	docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
kimi-code	list-price-credit (K3 3/15/0.30 vs K2.7 Code 0.95/4.00/0.19)	kimi.com/code/docs + CommandCode at-cost mirror
openai-codex	list-price-credit at OpenAI GPT-5.6/6 rates	OpenAI published pricing (mirrors)
meta-model-api	equal-rate: Muse Spark 1.1/1.2/1.3 share identical Standard pricing	dev.meta.ai/docs/pricing-rate-limits

Fallback plans (tier-pool yield, no official per-model differential retrievable)

alibaba-cloud, augment-code, byteplus, google-ai-studio, google-antigravity, groq-api, kimi-code: (covered), meta-muse-code, minimax, ollama-cloud, opencode Go (already had modelAllowances $ semantics), replit, together-ai, windsurf (Devin quota mechanism unpublished), plus any pseudo-model slot ("Auto mode", "API models", "frontier pool", "SWE-2", hetero pools).

Verification

npm run lint 0/0; npm run validate-data OK (33 plans); npm test 100/100 (added per-model + fallback cases); npm run build clean; data regenerated via python3 data/generate_plans.py and node scripts/build-data.mjs.

Adversarial Review Recalibration (2026-09-19)

Key Audit Findings & Mathematical Fixes

1. Cache Rate Harmonization: Eliminated _PER_MODEL_CACHE_RATE = 0.95 in data/generate_plans.py which was artificially doubling plan token capacity. Plan generation now enforces the shared defaultCacheRate: 0.75 from data/estimate-constants.json across all credit pools, reserving 95% caching only for providers with documented guarantees (Z.ai).
2. Model-Aware Token Resolution in TokenTranslator: TokenTranslator.tsx now calls resolveTierModelBudget(currentTier, comparisonModel.name, comparisonModel.id) rather than comparing direct API yields against monolithic tier scalars. Relabeled linear request counts to "Normalized 21K Turns".
3. Multi-Model Pool Drain Engine Fix: Fixed calculatePoolDrain in src/lib/pricing.ts to support perModelTokenBudgets alongside modelAllowances. Removed the Math.max(monthlyPrice, ...) flaw that was artificially inflating pool capacity for low-cost models.
4. Stack Candidates Model-Aware Ranking: buildStackCandidates now scores and scales candidates using the matched model's resolved token capacity (resolveTierModelBudget), preventing Dangerous Dave and Mix & Match from assigning inaccurate token yields to models.
5. Workflow Calculator Model Alignment: WorkflowCalculator.tsx now resolves plan capacity using resolveTierModelBudget for the active primary model in the workflow and verifies fit against poolDrain.overageCost === 0.
6. Subsystem Policy Reconciliation: Updated Cursor Pro evidence quotes and reasoning token descriptions across src/lib/throttle-profiles.ts and src/lib/reasoning.ts to reflect the dual-pool architecture (Cursor Models pool vs Other Models at API rates).

Verification

• npm run lint — 0 errors, 0 warnings (72 files).
• npm run validate-data — 33 plans and 33 models pass all schema invariants.
• npm test — 102/102 tests pass (8 suites, including new tests for model-aware stack candidates and pool drainage).
• npm run build — Clean production build with Vite + TypeScript.

Recalibration & Machine-Readable Usage Limits Export (2026-09-19 / 2026-09-20)

Key Achievements & Recalibrations

1. Machine-Readable Usage Limits Dataset (public/data/usage-limits.json):
   • Compiled and published 394 normalized limit records across all 33 services as part of the GitHub Pages build/deployment pipeline.
   • Includes empirical agent task counts (250K conservative, 550K midpoint, 900K complex turns), mathematical basis formulas, and source links.
   • Enforced schema validation in scripts/validate-data.mjs (CI fails if entries are missing, empty, or non-finite).
2. CommandCode Full 5-Tier Matrix Audit:
   • Expanded from single-tier stub to complete matrix: Go ($1, $10 credits, 250/5h, 1k/7d), GOAT ($10, $70 credits, 500/5h, 2k/7d), Pro ($20, $80 credits, 750/5h, 3k/7d), Max 10x ($100, $150 credits, 1.5k/5h, 6k/7d), and Max 20x ($200, $300 credits, 3k/5h, 12k/7d).
   • Modeled dual standard and premium credit pools, rolling 5-hour pools, and 7-day sliding window exhaustion rules.
3. OpenCode Go 28-Model Matrix Recalibration:
   • Audited the full 28-model Go catalog across $15, $30, and $60 monthly allowance pools with 20% 5h and 50% 7d rolling window caps.
   • Integrated model-specific request limits and token yields into data/generate_plans.py and data/coding-plans/opencode.json.
4. Z.ai GLM Coding Plan Formal Quota Alignment:
   • Recalibrated Lite ($18), Pro ($80), and Max ($168) against official weekly allowance ceilings: GLM-5.3 (97M, 582M, 1,358M tok/wk) and GLM-5.3-Flash (584M, 3,504M, 8,176M tok/wk).
   • Documented project concurrency limits (1, 1–2, 2+), 95% prompt cache, and off-peak 50% credit pricing.
5. Primary Documentation Index (SOURCE.md):
   • Established SOURCE.md as the unified primary evidence archive with direct links, documentation quotes, and mathematical formulas for all 33 services.

Verification

• npm run lint — 0 errors, 0 warnings (72 files).
• npm run validate-data — 33 plans and 394 usage limits entries pass all schema invariants.
• npm test — 102/102 tests pass across 8 test suites.
• npm run build — Clean production build with Vite + TypeScript.

Adversarial Math Audit & Telemetry Remediation Pass (September 19, 2026)

Key Audit Findings & Remediations (ADV-01 through ADV-05)

1. ADV-01: Non-Stackable Plans in Dangerous Dave Mode:
   • Resolved user requirement allowing non-stackable plans into Dangerous Dave mode while enforcing mathematical and contractual reality.
   • Plans with stackingPolicy === 'prohibited' are strictly clamped to copies: 1 in computeDaveStacks (src/lib/pricing.ts).
   • Lab Decision Engine displays an amber warning banner highlighting ToS single-account constraints rather than artificially projecting 4x multipliers.
2. ADV-02: Robust String Request Quota Parsing (parseTierRequestLimit):
   • Replaced brittle Number(tier.limits.fastRequests) with regex-powered parseTierRequestLimit capable of parsing human-readable quotas like "500 fast requests/mo", "1k/7d", "250/5h", and hyphenated ranges ("200-300").
   • Fixed pool drain calculations across OpenCode, CommandCode, and Z.ai tiers where numeric conversion previously yielded NaN or undefined.
3. ADV-03: Partitioned Sub-Pool Architecture:
   • Extended poolDrain in src/lib/pricing.ts to support dual standard and premium model allowances.
   • Workflows mixing standard and premium models now deplete their respective partitioned allowances rather than overflowing into overages prematurely.
4. ADV-04: Expanded Per-Model Yield Catalog for Credit Pool Providers:
   • Added explicit, multi-model perModelTokenBudgets across the full model fleets of credit-pool IDEs and API providers:
     • Augment Code (augment-code): Standard ($14.28 net credit after 40% LLM service fee) and Business ($71.43 net credit) across Claude Opus 5, GPT-5.6 Sol, DeepSeek V4-Pro, Gemini 3.8 Flash, and Cosmos.
     • Replit (replit): Core ($20 credit) and Pro ($100 credit) across Claude Opus 5, GPT-5.6 Sol, DeepSeek V4-Pro, Gemini 3.8 Flash, and Replit Agent.
     • Ollama Cloud (ollama-cloud): Pro ($60 credit), Max ($300 credit), and Team ($1,000 credit) across DeepSeek V4.1 Flash, DeepSeek V4-Pro, MiniMax M3, GLM-5, and Kimi K3.
5. ADV-05: Real-World OSINT Telemetry & Obfuscation Auditing:
   • Added isEstimatedCeiling, disclosedByVendor, and osintSource fields to public/data/usage-limits.json, tracking 414 total normalized entries across 33 providers.
   • Cross-linked reverse-engineered client telemetry, community telemetry traces, and proxy analytics to docs/OSINT_USAGE_STATISTICS.md for opaque services (Claude Code, Cursor, Windsurf, Google Antigravity, OpenAI Codex, Amazon Q).

Verification

• npm run lint — 0 errors, 0 warnings (72 files).
• npm run validate-data — 33 plans and 414 usage limits entries pass all schema invariants.
• npm test — 107/107 tests pass across 8 test suites (expanded by 5 tests for non-stackable clamping, regex parsing, partitioned sub-pools).
• npm run build — Clean production build with Vite + TypeScript.

---

Pass 10: Xiaomi MiMo Token Plan Integration (2026-09-20)

Scope: Added Xiaomi MiMo as the 34th curated coding plan. MiMo Token Plans use a credit-based pricing system with model-specific credit consumption rates that vary by cache hit, cache miss, and output tokens.

Changes

1. New Plan File (data/coding-plans/xiaomi-mimo.json):
   • 4 tiers: Lite ($6/mo, 4.1B credits), Standard ($16/mo, 11B), Pro ($50/mo, 38B), Max ($100/mo, 82B).
   • Per-model token budgets for MiMo-V2.5 and MiMo-V2.5-Pro derived from official credit-to-token conversion tables.
   • Off-peak 0.8× multiplier (UTC 16:00–24:00) noted in gotchas.
   • MiMo-V2.5-Pro-UltraSpeed (PAYG-only, MXFP4 quantized, >1000 TPS) documented as excluded from Token Plans.
2. Plan Generator (data/generate_plans.py):
   • Added xiaomi-mimo to STACKING_POLICY (prohibited).
   • Full plan generation with per-model credit-to-token math.
3. Provider Color (src/lib/pricing.ts): Added xiaomi: '#ff6900' brand orange.
4. Series Taxonomy: Added MiMo to KNOWN_SERIES in both scripts/series-taxonomy.mjs and scripts/validate-data.mjs.
5. Tests (src/lib/pricing.test.ts): Added 2 tests for Xiaomi provider color and MiMo tier budget resolution (109 total).
6. Build Data: Regenerated public/data/plans.json (34 plans) and public/data/usage-limits.json (422 entries across 34 providers).

Data Sources

• Official Token Plan pricing: https://mimo.mi.com/docs/en-US/price/token-plan
• Pay-As-You-Go overseas pricing: https://mimo.mi.com/docs/en-US/price/pay-as-you-go
• Rate limits: https://mimo.mi.com/docs/en-US/api/guidance/rate-limit
• Model cards: https://mimo.mi.com/models/en-US/mimo-v2.5 · https://mimo.mi.com/models/en-US/mimo-v2.5-pro
• OpenRouter listings: https://openrouter.ai/xiaomi

Verification

• npm run lint — 0 errors, 0 warnings.
• npm run validate-data — 34 plans and 422 usage-limit entries pass all schema invariants.
• npm test — 109/109 tests pass across 8 test suites.
• npm run build — Clean production build.

---

Adversarial Audit Remediation Pass (September 20, 2026)

Full hostile audit (15 findings, VULN-01 through VULN-15) remediated in one pass.

Summary of Fixes

1. VULN-01 (Critical) — Data pipeline supply chain: update-data.yml no longer auto-pushes third-party-derived data to main; refreshes open a PR (branch data-refresh/<date>) for human review. All GitHub Actions pinned by commit SHA. validate-data.mjs gained price sanity bounds (non-free models must have positive finite in/out prices; ceilings $1,000/$2,000 per M; cachedInput must be null or finite ≥ 0).
2. VULN-02 (High) — Cache scaling on metered quotas: computeApplesToApples now applies symmetric cache scaling only to dollar-credit pools (isDollarPoolTier: limits containing a credit key or perModelTokenBudgets on a credit/equal-rate basis). Usage-metered quotas deliver fixed tokens regardless of the user's cache rate and are never scaled.
3. VULN-03 (High) — payg-overage never throttled: simulateSprintThrottle now emits a distinct overage status for payg-overage profiles (Google AI Pro, OpenCode Zen, DeepSeek PAYG) once a window/monthly ceiling is crossed; SimulationResult.overageTurns reports billed excess; overall status resolves by severity (smooth < queued < overage < blocked). BurstSimulator gains an Overage filter, badge, chart color, and stat column.
4. VULN-04 (High) — Hardware feasibility & payback basis: computeHardwareEconomics derives requiredInferenceHours from the preset's flagship decode throughput and reports infeasibleWorkload / feasibleMonthlyVolumeM. Payback is only quoted when the amortized monthly alpha is positive, so headline savings and payback cannot contradict. HardwareBreakeven renders the infeasibility warning with the required-vs-available decode hours.
5. VULN-05 (High) — Synthetic receipts: ParsedAgentSession.measuredTokens distinguishes measured from parser-fabricated token counts; SessionReceiptReport.tokensMeasured propagates it; the receipt page and copied receipt both flag estimated sessions; the headline prefix "ESTIMATED …" prevents synthetic numbers passing as a bill.
6. VULN-06 (Medium) — Impossible reasoning combos: calculateReasoningCost checks input+reasoning+output against contextWindow and reasoning+output against maxOutput (unknown caps treated as feasible); ReasoningExploder disables infeasible effort pills and renders the impossibility reason.
7. VULN-07 (Medium) — Cache-assumption regex hijack: parsePlanCacheAssumption anchors the percent to cache context (NN% cache|hit or cache … NN%), so discount percentages cannot be misparsed as cache rates; results clamp to ≤ 0.95.
8. VULN-08 (Medium) — Divergent cache heuristics: the ad-hoc (1 − assumed×0.8) formula for unmatched tiers was replaced by the same model-priced ratio used for matched tiers (via a lab-representative basis model), gated identically on dollar pools.
9. VULN-09 (Medium) — Arbitrage spend asymmetry: the callout now discloses when the API leg outspends a capped single-seat subscription by >1.25x and points to Dave Mode / Mix & Match.
10. VULN-10 (Medium) — ToS classifier liability: 'no guarantee' removed from TRAINS evidence; ipIndemnity: true renders "Indemnity Offered (warning)" with a carve-outs caveat instead of "Full Indemnity"; string indemnities get a verify-scope label; classifyGotcha only assigns critical data-privacy severity on strong training/retention signals, weak mentions ("training" inside a caching sentence) downgrade to warning.
11. VULN-11 (Medium) — Benchmark starvation: build-data.mjs gained an effort-suffix-stripped fuzzy matching fallback (recovered 11 of 35 models; match rate 29% → 60%) and a hard floor: builds fail below a 25% match rate unless ALLOW_LOW_BENCH_MATCH=1.
12. VULN-12 (Medium) — Quota/time realism: weekly quotas convert at 52/12 weeks (not ×4); the throttle simulator carries a deferred-turn backlog so slow-queue delay actually reduces completed turns instead of being cosmetic.
13. VULN-13 (Medium) — Zero-hostile URL params: shared numParam helper (src/lib/params.ts) replaces Number(x) || default in HardwareBreakeven and BurstSimulator; ?salvage=0, ?prior=0, ?idleHours=0 now render as configured.
14. VULN-14 (Medium) — Cache-write premium + clamp unification: cacheWriteShare (0.06) added to data/estimate-constants.json and wired into agentBlendedCost in both fetch-models.mjs and build-data.mjs (amortized write cost at the provider write rate); salvage clamp unified at 50 via SALVAGE_CLAMP_MAX; costPer1kRequests || 1 → ?? 1 (invariant I).
15. VULN-15 (Low) — Literal extraction: session cache-bust residual (0.15) and provider cache-read fallback ratios moved into data/estimate-constants.json; teams.ts no longer fabricates a 10% cache discount for unknown cache prices (conservative full-input fallback). Residual: classifyModelTier's hardcoded model-name table and teams.ts chat workload shapes remain flagged for a future data-driven refactor.

Verification

• npm run lint — 0 errors, 0 warnings.
• npm run validate-data — 34 plans and 422 usage-limit entries pass all schema invariants (now including price sanity bounds).
• npm test — 127/127 tests pass across 8 suites (expanded by regression tests for VULN-02/03/04/05/06/07/09 and updated expectations for VULN-10/12).
• npm run build — Clean production build.