- Enforce token-first fit invariant in WorkflowCalculator: require caps.basis >= requiredTokens and poolDrain.overageCost === 0 to qualify as fits
- Re-anchor Google Antigravity Ultra 20x Gemini Flash to empirical 3.0B tokens/mo (down from 18.7B virtual dollar division) and Claude Sonnet to 180M tokens/mo matching DevForth ~00/7d telemetry
- Cap Ollama Cloud Team tier at 4,500M tokens/mo (10-stream physical socket bandwidth ceiling) and Kimi Code Vivace K2.7 Code at 1,200M tokens/mo
- Unclamp negative savings in WorkflowCalculator to transparently show when plans cost more than direct API
- Integrate rate-limit lockout wage exposure into Breakeven Verdict headline to penalize severe lockout risks
- Add parsePlanCacheAssumption and symmetric cache-rate scaling to computeApplesToApples
- Fix internal credit burn rate in calculatePoolDrain when models have dedicated perModelTokenBudgets
- Add unit tests verifying 115 passing tests across 8 suites
- ADV-06: Standardize leaderboard monthlyRequests and costPer1kRequests on 21K normalized agent requests; preserve vendor chat quotas in vendorQuotaRequests and display informative subtitle badge
- ADV-07: Enforce discrete single-seat clamping in Standard Mode, eliminating fractional plan scaling and surfacing unspentBudget
- ADV-08: Re-align Claude Code Max 20x monthly budget to 72M tokens/month steady-state (6x Pro weekly ceiling), with 144M midpoint and 240M peak burst
- ADV-09: Upgrade CommandCode Max partitioned pool premium classification to include OpenAI flagships/reasoners, Gemini Pro, Grok, and >= /M blended rate models
- Update AGENTS.md rules K & L and expand test suite to 112 passing tests
- Add scripts/audit-external-sources.mjs to cross-check llmprice.com and devforth.io
- Add claude-code-max20x, openai-codex-pro5x, and openai-codex-pro20x throttle profiles
- Document Claude Code 5h burst vs ~6x weekly ceiling asymmetry in plans and gotchas
- Expand OSINT_USAGE_STATISTICS.md and DATA_SOURCES.md with Faros, DX, and DevForth telemetry
- Harmonize plan generation prompt cache rate to shared 75% default from estimate-constants.json
- Update TokenTranslator to resolve plan token capacity model-specifically via resolveTierModelBudget
- Fix calculatePoolDrain to drain model-specific allowances and eliminate Math.max deflation on low-cost models
- Make buildStackCandidates model-aware for accurate Dangerous Dave and Mix & Match stacking
- Update WorkflowCalculator to resolve plan capacity model-specifically for active workloads
- Reconcile Cursor Pro evidence quotes in throttle-profiles and reasoning token policies
- Add unit tests for model-aware stack candidates and pool drainage (102/102 passing)
- Canonicalize model name resolution in resolveTierModelBudget to avoid mismatching OpenRouter model names (e.g. Z.ai GLM 5.3 Flash)
- Scale subscription request yields from per-model token capacity rather than tier pool fallback
- Wire perModelTokenBudgets into Google Antigravity, BytePlus, Alibaba Cloud, MiniMax, and Meta Muse Code in generate_plans.py
- Add Dedicated Drain vs Shared Pool badges in Decision Engine and per-model native capacity in Token Budget Translator
- Fix LaTeX KaTeX/MathJax parsing errors in docs/DATA_SOURCES.md by replacing raw dollar delimiters with USD/M
Evaluated quota calculation mechanics against DeepFrugal and AI-10-USD:
- Added time-of-day off-peak blend shares (80/20) and agent cache presets to estimate constants.
- Implemented calculateTimeBlendedCost and calculatePoolDrain in pricing.ts with full unit tests.
- Extended _schema.json and generate_plans.py with modelAllowances for router plans (CommandCode, OpenCode Go).
- Enhanced WorkflowCalculator.tsx with a Hybrid Pipeline toggle (75/25 workhorse/frontier) and subscription pool drain metrics.
- Regenerated all 33 plans and consolidated static data.
Incorporates findings from multi-batch web crawl across all 33 providers:
- Added annual pricing discounts for Cursor Pro+/Ultra, Claude Code Pro, and Replit Core/Pro.
- Reconciled DeepSeek API endpoints and models against official docs (Flash vs V4-Pro, 1M context, 2,500 concurrency).
- Corrected canonical pricing URLs for Together AI and Fireworks AI.
- Regenerated all 33 plans via generate_plans.py and consolidated public/data/plans.json.
- Documented live audit trails in docs/CRAWL_AUDIT_2026.md and docs/VERIFICATION.md.
Antigravity, Windsurf and OpenAI Codex published no token quotas; their
estimates used a 150K-tokens-per-task basis that sat below the measured
autonomous-agent band (200-800K in + 30-100K out) and disagreed with the
WorkflowCalculator's 40-request task model. The band now lives in
data/estimate-constants.json (250K conservative / 550K midpoint / 900K
optimistic), generate_plans.py derives every affected tier from it via
task_estimate()/scale_estimate(), official multipliers (5x/20x) and
per-seat mirrors are preserved, and each tier carries explicit
midpoint/optimistic values plus estimateMeta provenance. Windsurf Max's
'10x' note is corrected to the 5x relation its number actually used.
Verified: validate-data, 42 tests, lint and production build all pass.
- docs/VERIFICATION.md: all 33 providers with official sources, status,
discrepancy log, open items and a re-verification checklist
- DATA_SOURCES.md: corrected directory (prices/quotas/models), unified value
score formula, new estimateMeta/stackingPolicy schema docs, TOS matrix now
matches the structured classifier
- README: corrected category membership and counts (14/5/14)
- AGENTS.md: validator/test commands, estimateMeta + stackingPolicy invariants
- remove unused budget-precomputed.json artifact and dead useBenchmarks/useBudget
hooks; build-data no longer generates the unused table
- BytePlus/Kilo/Tabnine/Meta corrections from final verification pass
- Anthropic/OpenAI API: fix spend tiers, model lists, cache-aware token bands
and correct ZDR/retention claims; Augment drops its own 40% fee correctly
- BytePlus anchored to our Claude Pro estimate; Cursor Pro+/Ultra now match
the official 3x/20x multipliers; Amazon Q Free uses the official 50-request
cap and Pro becomes an explicitly low-confidence scenario band
- Antigravity Ultra split into official 5x ($99.99) and 20x ($199.99) tiers
- Ollama Team credits $1,000 (not $1,600), 10 streams, current model list
- Copilot conservative estimates use base credits only, Flex in midpoint
- DeepSeek V4-Pro CNY rates quoted; Mistral training opt-out; Together dynamic
limits; Fireworks free-trial math; Google AI Studio Flash price basis
- Z.ai: conservative/midpoint/optimistic now derived from the official
95%-cache token allowance table (Lite 208/314/420M, Pro 1.26/1.88/2.51B,
Max 2.93/4.39/5.85B, Team 1.38/2.07/2.76B per seat) with peak/off-peak
floor/ceiling; models corrected to GLM-5.3/5.3-Flash with routing note;
campaign, account-sharing and no-training facts added
- Alibaba: added missing Essential tier; replaced invented credit math with
the official credit quotas and a documented qwen3.6-plus deduction basis;
corrected quota windows (7-day personal, monthly team), extra bundles,
Coding Plan phase-out, data-training split and seat rules
Replace nonexistent/obsolete model references so TokenTranslator Part 2
matches cleanly: Gemini 3.8 Pro -> Gemini 3.1 Pro (google-ai-studio),
Llama 4 Code 70B -> Llama 4 Maverick (meta/ollama blocks), Mixtral 8x7B and
FireFunction v2 removed (obsolete) in favor of modern open-weight equivalents,
GPT-5.2 Mini -> GPT-5.2, Kimi K3 variants normalized
- Replace fabricated Jules task-based tiers with official tiers: Individual ($0),
Google AI Pro ($19.99), Google AI Ultra ($249.99), and consumption-based
Organization tier via Google Cloud
- Correct model list (Gemini 3.8/3.7/3.6 Flash, 3.1 Pro, Claude Sonnet/Opus 4.6,
gpt-oss-120b; Gemini-only for Enterprise)
- Swap unsupported 'tasks/day x 7K tokens' methodology for labeled research-based
agent-token estimates (~150K/task) with explicit low-confidence assumptions
- Document quota model (5-hour + weekly refresh, work-correlated limits), AI-credit
overage at GEAP pricing, OpenClaw ToS breach, no-BYOK, Claude -> Anthropic terms
- Fix pre-existing unclosed write_json block for ollama-cloud that broke
python3 data/generate_plans.py
John Smith