fix(catalog): re-audit free-tier quotas against official pages

[diegosouzapw]

Summary

Re-audits five free-tier entries of the per-model catalog (open-sse/config/freeModelCatalog.data.ts) against the providers' own pages. On top of the current base (release/v3.8.51 @ 2265ce7, which reclassified Cerebras as a one-time $5 signup credit in #12591, −30M) the public headline lands at ~1.47B documented free tokens/month: steady 1,473,225,000, first month with signup credits 2,099,725,000 (~2.10B).

Isolated effect of this PR on the base: −60M gemini, −20M ollama-cloud, +15M groq, +60M nara → base 1,438,225,000 → 1,473,225,000. The move from the ~1.50B quoted in the first revision of this PR to ~1.47B is entirely the upstream Cerebras reclassification, not a change of this audit.

• gemini −60M → uncapped (Google removed the per-model free-tier table on 2025-12-23; nothing left to sum)
• ollama-cloud −20M → uncapped (Free plan is "starter usage credits", no token figure)
• groq +15M → five per-model 200K-TPD caps (6M each, independent — not a shared pool); three IDs retired from the free tier dropped
• nara +60M → one 7M/day account bucket = 210M, the 8 models of the Free plan
• mistral 1B kept — the figure only exists inside the account console; console-verified on 2026-09-02 by the owner (Limits → Free mode → "Tokens per month" = 1,000,000,000) and recorded next to the entry

Every re-audited block now carries an // evidence: comment with its source; the docs, README, free-tiers guide, budget SVG and the CI-gated counts (442 entries / 34 recurring pool keys / 16 quantified pools + 5 Groq caps) follow the catalog.

Provider table

Provider	Before	After	Evidence
gemini	60M (pooled gemini-free, 4 models)	uncapped, never summed (4 models)	https://ai.google.dev/gemini-api/docs/rate-limits — per-model table removed 2025-12-23; limits "can be viewed in Google AI Studio"
ollama-cloud	20M (pooled, 8 models)	uncapped, never summed (8 models)	https://ollama.com/pricing — Free: "Starter usage credits included"; https://docs.ollama.com/cloud — "usage resets monthly"; no token figure
groq	15M (one groq pool, 5 models incl. llama-3.3-70b-versatile, llama-4-scout, qwen3-32b)	5 × 6M per-model caps (gpt-oss-120b, gpt-oss-20b, gpt-oss-safeguard-20b, qwen3.6-27b, qwen3.8-27b), hardStopGuaranteed	https://console.groq.com/docs/rate-limits — Free Plan Limits: 200K TPD per model, 429 on limit; https://console.groq.com/docs/billing-faqs — no payment method on file; https://console.groq.com/docs/deprecations — retirements 2026-07-17 / 2026-08-16
nara	150M (3 models)	210M (7M/day × 30, one bucket, 8 models)	https://router.bynara.id/api/plans — plan free: token_cap_daily=7000000, rpm_limit=15, model list (public, unauthenticated)
mistral	1B (pooled, 5 models)	1B kept (pooled, 5 models)	console-verified 2026-09-02 (console.mistral.ai → Limits, Free mode, "Tokens per month" = 1,000,000,000); public pages only confirm the cap exists: https://docs.mistral.ai/admin/billing-usage/usage-limits, https://help.mistral.ai/en/articles/698531

Headline: ~1.47B steady on the current base · pools 37 → 34 (gemini/ollama-cloud left the steady set; Groq's caps are not a shared pool) · entries 437 → 442 vs the base (nara +5, groq retired IDs replaced).

Evidence classes

Every number added or changed by this PR cites its source in a comment next to the entry (FREE_TIERS.md → Methodology → Evidence classes):

• public-page — a provider page anyone can read (gemini, groq, ollama-cloud);
• api-public — an unauthenticated endpoint of the provider (nara /api/plans);
• console-verified <date> por <who> — the figure is only visible inside an account console; the comment records who saw it and when, plus the public page that says the cap exists. Today only mistral is console-verified. The regression test tests/unit/free-tier-reaudit-2026-09.test.ts pins this: the 1B pool is accepted only while a dated console-verified line is present in the source; otherwise every Mistral row must be recurring-uncapped (0).

Entries that predate the 2026-09-02 re-audit inherit the earlier (2026-06-17) research until they are touched; the catalog header now says the file is hand-curated (there never was a generator) and states the rule.

Collateral changes

• open-sse/config/freeTierCatalog.ts (legacy provider-level map): gemini/ollama-cloud removed, nara 210M, groq 30M (map kept in descending order; cerebras stays removed as fix(providers): reclassify Cerebras as a one-time $5 signup credit #12591 left it — 17 legacy keys, 1,475,025,000).
• open-sse/config/providers/registry/{groq,nara}: Groq registry gains openai/gpt-oss-safeguard-20b; NaraRouter registry lists the 8 Free-plan models.
• src/app/(dashboard)/dashboard/combos/page.tsx: FREE_STACK_PRESET_MODELS swaps the retired groq/llama-3.3-70b-versatile for groq/openai/gpt-oss-120b.
• src/shared/constants/providers/apikey/{frontier-labs,gateways}.ts: Groq freeNote ("30 RPM / 14.4K RPD" was stale) and NaraRouter onboarding note re-written; docs/reference/PROVIDER_REFERENCE.md regenerated (npm run gen:provider-reference).
• i18n (src/i18n/messages/*.json): providers.onboardingProviderDescriptions.groq and .gemini rewritten in en and pt-BR; the other locales carry the documented __MISSING__:<en> sentinel (runtime falls back to English, queued for the translation pipeline) — except vi, which must not hold placeholders (tests/unit/i18n-vi-completeness.test.ts) and received a real translation that needs a native-speaker review.
• Docs: README.md (headline, counts, hero/budget alt texts), docs/reference/FREE_TIERS.md (TL;DR, dated re-audit note, contributors, per-provider table, Evidence classes, resolved Groq finding), docs/diagrams/free-tier-budget.svg (bar/grid regenerated for 16 pools + Groq caps; Gemini/Ollama Cloud moved to the "un-countable" row), docs/diagrams/readme-hero.svg, docs/getting-started/FREE-TIERS-GUIDE.md, docs/diagrams/README.md.
• Tests: tests/unit/free-tier-reaudit-2026-09.test.ts (new, pins every re-audited fact incl. retired Groq IDs and registry routability), free-tier-catalog.test.ts, free-providers-batch-2026-07.test.ts, autoCombo/strict-zero-cost-filter.test.ts (fixture re-pointed to groq/openai/gpt-oss-120b).
• Base merge: origin/release/v3.8.51 merged in (no rebase); the duplicate in locale removed upstream by feat(i18n): locale-expansion tooling — add-locale orchestrator, aliases, translation-ratio gate; retire duplicate 'in' locale #12496 stays removed.

Validation

Base merged: origin/release/v3.8.51 @ 2265ce7 (8 conflicts resolved — freeTierCatalog.ts, free-tier-catalog.test.ts, README.md, FREE_TIERS.md, FREE-TIERS-GUIDE.md, docs/diagrams/README.md, free-tier-budget.svg, readme-hero.svg). Rule applied: keep the base's Cerebras reclassification (one-time $5 credit — rows, legacy-map removal, cerebras === undefined assertion, per-provider table row) and every re-audited row/prose from this branch, then let npm run check:docs-counts decide every published number. docs/screenshots/free-tier-budget-card.svg was taken from the base unchanged (this branch owns no generator change for it). Totals on the merged head: steady 1,473,225,000 · first month 2,099,725,000 · 34 pool keys · 442 entries · legacy FREE_TIER_BUDGETS 17 keys / 1,475,025,000.

Local gates run on the merged head (this devbox is saturated and the full unit suite times out here; the full matrix runs in CI on this PR):

Check	Result
npm run check:docs-counts	✓ All checks pass
npm run check:docs-sync	PASS (41 llm.txt mirrors, 41 CHANGELOG translations)
npm run typecheck:core	exit 0
npm run check:dashboard-typecheck	exit 0 (run in Task 1.4 after the combos/page.tsx preset change)
npm run check:provider-consistency	OK (273 REGISTRY entries, 356 canonical providers)
Focused unit tests (free-tier-reaudit-2026-09, free-providers-batch-2026-07, free-tier-catalog, free-model-catalog, free-tier-summary-route, check-docs-counts-sync, removed-providers-blocklist)	83 ok / 0 failed (re-run after the 2265ce7 merge)
Vitest tests/unit/autoCombo/strict-zero-cost-filter.test.ts	15/15
ESLint on the changed .ts/.tsx/.mjs files (project suppressions applied)	exit 0, 0 errors
i18n:check-ui-coverage / i18n:check-value-drift:warn / i18n-vi-completeness	PASS / PASS / 5/5 (Task 1.4 fix round)

⚠️ base-red inherited: #12581 (🔴 Release branch not green: release/v3.8.51, opened 2026-09-03; the failing gates there do not touch this PR's files)

Refs: internal research 2026-09-02 (primary-source quota audit).

CI triage (first run on 09942a5)

Check	Verdict
Docs Gates (fast-path) — check:fabricated-docs flagged /api/plans in FREE_TIERS.md	ours, fixed in 113b608 (the NaraRouter example no longer looks like an OmniRoute route)
Fast Quality Gates — file-size gateways.ts 1462 > 1459	ours, rebaselined in 113b608 (+3: nara authHint wrap + Prettier reflow of two pre-existing over-long lines; same rationale as the seekai/logfare/freebuff entries)
Fast Quality Gates — open-sse-typecheck open-sse/executors/glm.ts TS2554 and API Route Typecheck (same error)	inherited: glm.ts was last changed on the base by #12179; this PR does not touch it
Unit Tests fast-path shards — stream-failure sanitization / error-boundary / redaction tests (#12506 campaign), sanitizeErrorMessage ReDoS timing	inherited: none of those modules are in this diff
Unit Tests fast-path — #7253 release-green: docs contain zero fabricated API/file-ref drift	ours, same /api/plans cause, fixed in 113b608

CI triage (final run on 113b608)

Check	Verdict
Docs Gates (fast-path), Merge integrity, Vitest (fast-path), Change Classification	green
Fast Quality Gates / API Route Typecheck — open-sse/executors/glm.ts TS2554	inherited (#12179 on the base; not in this diff)
Unit Tests fast-path (1–4/4) — 16 failing tests (stream-failure sanitization, credential redaction, HuggingChat/Kiro/codex-image error boundaries, model-sync route)	inherited: none of the 24 failing test files are in this diff, and the identical 16 tests fail on unrelated branches fix/reset-aware-model-family (run 33825690606) and fix/overloaded-not-provider-breaker (run 33825612505)