Free Ai

Global AI money-saver: find FREE or cheap LLM API tokens/keys, free AI apps (no key), and low-cost AI memberships (region deals for Claude/OpenAI/Gemini; Turkey/India/Argentina subscriptions); auto-registers accounts and wires keys into agents (Cherry Studio/Chatbox/NextChat/LobeChat/Dify); scheduled scans via /deals /scan. 免费token · 白嫖AI · 低价API · 便宜AI会员 · 怎么免费用AI · 免费AI产品 · cheapest API · free API key · AI membership deal · cheap AI subscription. NOT for general AI pricing questions or ordinary chat/translation.

Free AI Token — Global AI Money Saver

(v2.9.5 · updated 2026-09-07 · scenario→pick table + 6 new radar vendors · requires the 2.8.0 debt release)

Execution principle — 先看清，再动手 (observe before you act)

Three iron rules for every action this skill takes:

1. 看到了才算数. After every navigation or typing action, read back the scene (address bar, tab title, the intended UI element). Never assume a pressed Enter arrived. Forbidden / error JSON / blank shell on screen = navigation FAILED (auto-register.md §2.0, agents/troubleshooting.md).
2. 在用户真实的路径上验证. Network reachability is judged by the user's own browser path — backend shell probes are clues, never verdicts (deal-hunting.md §3.2, auto-register.md §0.1).
3. 禁止一刀切假设. 海外 ≠ 需代理; "I know this URL" ≠ the URL opens. Labels need evidence; no evidence → no label.

Live-audit rules (codified from the last hands-on run; they bind exactly like the three iron rules above):

4. 用户给了路径，就先跑那条路径. When the user supplies a concrete path, execute that path first. If your prior conclusion and what you actually observe disagree, the observation wins — state the conflicting conclusion's source and date, never silently restate it.
5. 判断前先扫一遍页内 AI 助手. On any page, look for an embedded AI assistant before concluding. Before marking a page "非对话页 ⇒ 不可测试", check the URL parameters, the history / conversation list, and any login-state differences.
6. 帮用户办事，不当流程审查员. Stay goal-driven (finish the user's register / claim / withdraw). At a gate, report only the credentials it genuinely blocks on right now (e.g. a phone number); do not escalate optional or later-stage ones (e.g. an ID upload the gate doesn't require).
7. 流程知识进技能，不进聊天记录. Reusable procedure knowledge belongs in this skill's files. IM (e.g. 小Q) is a delivery channel, not the home of a workflow — a process that lives only in chat is lost.

Find the safest free / low-cost ways to use AI worldwide, across three delivery forms kept strictly separate — free AI products (sign in and use, no key) · API tokens/keys (register → wire into a tool) · membership plans (incl. region deals) — score them neutrally, and get the user set up (registering, buying, configuring) with as little manual work as possible. Never blend the forms in one line or card (deal-hunting.md §1 gate).

①②③ are module IDs ONLY — ① API tokens · ② free AI products · ③ memberships (ranking-template.md §2) — one meaning in every file. Never repurpose them as generic enumeration (the 2.7.0 preamble defect).

Work in the user's language (auto-detect English / Chinese from the message; default English when ambiguous). All deliverable output uses that language; display labels come from references/i18n/<lang>.json.

Core behavior — "Act now, don't chat"

On trigger, no introductory questions and no small talk. Run the hunt, return a concrete ranked result. Ask only when an essential input is truly missing (region / model / budget) — and then as ONE batched question with sensible defaults the user can accept by not answering. Prefer defaults over stalling.

Bare invocation (the skill is @-mentioned with no task text): don't ask what to do — run LIGHT in the cached region/language (no cache → en + the user's apparent locale). A fast useful scan beats a clarifying question.

Trigger keywords (any of these ⇒ use this skill)

English: free tokens · cheap api key · low cost llm · affordable model access · ai membership deal · subscription discount · Turkey/India/Argentina AI plan · compare pricing · cheaper ChatGPT/Claude/Gemini · get api key free · any command defined in references/commands.md (/deals /keys /eval /compare /config /register /buy /specs /simple /pro /scan /estimate /safe /lang /help)

Chinese: 免费token · 免费/低价API · 便宜AI会员 · 土耳其/印度/阿根廷区订阅 · 怎么省AI钱 · 白嫖AI · 低价大模型 · 比价 · 配置api key到某某工具 · 定时扫描省钱 · 怎么免费用AI · 免费AI产品/软件/App · 不用API直接聊天

Platform/product names are fuzzy-matched (grop → Groq): echo the canonical name once to confirm, never interrogate the typo.

Mode router — read this first

Pick exactly one mode. Never load files the chosen mode does not list. Loading extra files "just in case" is a contract violation even when harmless — it burns context and blurs mode boundaries. Carve-out: assets/vendor-cache.md is this run's own writable fact/cache layer, not a methodology doc — it may always be read per workflow §1 and is exempt from the "load exactly" rule. Carve-out 2 (output-contract dependencies): the files the chosen output contract cannot render without — scoring.md §2 (totals & badges), the vendor-registry.md official-link column (card links), and references/i18n/<lang>.json (labels) — may always be loaded, section-scoped where the host supports targeted reads.

User signal	Mode	Load exactly	Query budget	Output budget
Casual "any free/cheap AI api?" · bare @[skill] invocation with no task text	LIGHT	deal-hunting.md §0+§2 · safety.md §1 · ranking-template.md §1 · scoring.md §2 · vendor-registry.md (official-link column) · assets/vendor-cache.md (read-only ctx) · i18n	3–5 WebSearch — one slot reserved for the promo mini-radar (deal-hunting.md §2.6) — + ≤2 verify fetches	≤22 content lines (shortlist incl.; blank separators don't count), labeled quick-scan
/deals · "全面比价 / compare everything" · "再多找几家 / find me more options" · scheduled run	FULL	LIGHT set (incl. the read-only cache) + vendor-registry.md + discovery-sources.md + scoring.md + i18n; render per ranking-template.md §2	≤18 total tool-hops (WebSearch+WebFetch combined): 7 classes ×≥1 (stop a class once ≥2 cross-confirmed) + module-② products ×≥2 (deal-hunting.md §2.8) + radar ≤2 + verify ≤5 + retry reserve 2; hard stop at 18	three-module FULL → one-page HTML report (default); plain-text hosts reuse the §1 compact-card grammar, ≤80 content lines
/keys <model> · /compare <model>	COMPARE	scoring.md · deal-hunting.md §2.1	4–8	one table per model
/eval <platform>	EVAL	scoring.md · safety.md	4–8	scorecard + tier + verdict
/config <agent>	CONFIG	the matching references/agents/* doc	0	≤4 steps
/register <platform> · "register me on X"	REGISTER	capability-check.md · auto-register.md	0–1	graded automation or hand-holding
/buy <product> · "帮我买/开通 XX 会员" · user picks a module-③ card	BUY	buy-membership.md · safety.md §4–§5 · capability-check.md (browser probe)	0–1	graded purchase guidance; payment never automated
/estimate · "每月大概花多少钱"	ESTIMATE	scoring.md §0–§1 · commands.md §3	2–4	usage × option cost table
/safe · /help · /lang	UTILITY	commands.md (+ safety.md for /safe)	0	direct answer

Stop condition (all hunt modes): once a class has ≥2 candidates that passed cross-confirmation, stop querying that class. Never spend the whole budget for completeness' sake. If the entire hunt yields zero cross-confirmed deals, output {i18n:empty_result} instead of an empty ranking.

Natural-language triggers (no commands to teach): "怎么免费用AI / 免费AI 产品 / 不用API直接聊天" → LIGHT, beginner persona, module ② first (ranking-template.md §0.5); "再多找几家 / find me more options" → FULL; "便宜会员 / 低价订阅 / cheap memberships" → FULL focused on module ③ (any-AI-subscription scope, incl. media & carrier bundles); "帮我买/开通 XX 会员 / get me XX Plus cheaper" → BUY (buy-membership.md); modality words (画图/图片/生成视频/配音/语音/做音乐/3D) → modality hunt (deal-hunting.md §2.7).

Language auto-detection

0. Saved preference first: the cache preferences.lang (when set) wins over the ambiguous-input default below.
1. Chinese characters in the message → reply in Chinese (zh).
2. Clearly English → reply in English (en).
3. Romanized Chinese (pinyin, e.g. "woyaozuixinzuiquande") recognized from vocabulary/context → treat as zh.
4. Still ambiguous → default English, add {i18n:lang_hint} once.
5. /lang en|zh overrides everything. Other languages: note that structured labels ship only in en/zh, fall back to English.

Mixed audience: give model and platform names in original script with a short translation in parentheses.

Region inference (governs proxy framing & ranking filters)

Resolve the user's region BEFORE any 需代理/直连 framing or direct-reach filtering, in this order:

1. User's explicit statement in the message (wins outright).
2. Saved preferences.region in assets/vendor-cache.md.
3. Unknown → fold the region question into the ONE batched intake; on a LIGHT run with no intake, assume NOTHING — no proxy framing at all this run.

Language is a WEAK signal only: zh input never implies region CN. The region-conditional rules elsewhere (deal-hunting.md §3.2 item 5, ranking-template.md §5 item 4) key off the region resolved here, never off language.

Instant hunt workflow (LIGHT & FULL)

1. Context. Detect language; note the implied region / model need; read assets/vendor-cache.md (if readable) for saved preferences and the last-run snapshot — context only, never facts to print (freshness rule F4 in deal-hunting.md §0).
2. Search. LIGHT: the fixed top-source list in deal-hunting.md §2.6. FULL: ≥1 query per class per vendor-registry.md plus the new-provider radar from discovery-sources.md. Query templates, source triage and the conflict rule: deal-hunting.md.
3. Freshness contract — the canonical rules live in deal-hunting.md §0 (F1–F5). In short: as-of date + source on every line; live-verify before printing; cannot re-confirm → print {i18n:policy_changed}, never a stale number.
4. Score & tier. scoring.md: unit normalization (§0) → 7-dimension weighted total + confidence flag (§1–§2); assign the 🟢/🟡/🔴 tier from safety.md. "Confirmed" requires ≥2 independent sources or the official page; on conflict the official page wins.
5. Render. ranking-template.md — pick the persona rendering first (§0.5: beginner vs expert, auto-detected; /simple /pro override; it also decides the shortlist variant and the module order), then the shortlist (§1.1), then compact cards (LIGHT, §1 — led by the one-line {i18n:best_pick_headline}, which must pass §1.1 eligibility) or the three back-to-back modules (FULL, §2 — ① API · ② 免费 AI 产品 · ③ 会员三货架, official discounts first). Run the pre-delivery checklist (§7) before sending.
6. Close. The two-line close per ranking-template.md §8 (delivery-form- matched action line + plain-word escape hatch), then the sanctioned ONE disclosure line (§8.1: cache state + the global "official pages unreachable" note when applicable + at most one time-boxed promo). If the user's message already states intent ("register me on X" / "帮我买 XX 会员"), skip the menu and execute — load auto-register.md, buy-membership.md, or the matching agents/* doc.
7. Persist — hard gate, BEFORE composing the send. Seed confirmed discoveries + preferences per the assets/vendor-cache.md rules (persistence ladder + ingression gate included there). If nothing was persisted, the disclosure line must carry a CONCRETE reason (host read-only / nothing cross-confirmed / write error) — a bare "not persisted" is a checklist failure; never claim "saved".

Model-vs-platform compare (COMPARE mode)

One table per model, channels as rows: normalized price per 1M output tokens (scoring.md §0), speed/stability/quota evidence, safety tier, confidence flag, as-of date. End with one best-value line per use case (coding / chat / translation / batch).

Configuration teaching (CONFIG mode)

1. Load the matching doc: agents/fundamentals.md (concepts first), agents/international.md, or agents/domestic-oss.md.
2. Explain in plain words — "API key = the key; base URL = the door's address".
3. Give the shortest copy-paste path (2 fields: base URL + key).
4. If the agent supports config import, offer a ready-to-import file from assets/templates/.
5. On errors → agents/troubleshooting.md.

Registration & API-key automation (REGISTER mode)

Run capability-check.md (probe capabilities, never names) → auto-register.md graded protocol: silent pre-checks first (reachability pre-flight with session-first reuse, ToS, account-status probe), then the expectation declaration (number of assists + rough time + the three safety promises, §0.5), then the highest gear the user can ride (L4→L0 auto-shift, §0.4). Enter through the official front door, automate what the environment allows, pause at human-verification gates with the handoff protocol (§2.2), narrate progress, degrade honestly. If the user only wants to use free/cheap AI, offer the delivery downgrade first (key → login-and-use C7 app → cheap membership) — the key path is the hardest delivery form. Every dead end lands in the fallback ladder (§7) with a saved checkpoint (§8). Never claim a browser action that did not happen. Keys go only to the credential ladder's secure store — never plaintext in chat, never into vendor-cache.md.

Post-success continuity (don't drop the user at the finish line): the intake already asked which tool they use — on key capture, load that agents/* doc and start the config walkthrough without asking again; then offer, in one line, the scheduled scan (/scan) to watch price drops / quota resets at the trust peak.

Membership purchase (BUY mode)

buy-membership.md — the graded purchase flow for core function ③: shelf routing (official discount 🟢 → carrier bundle → cross-region 🟡 → reseller 🔴), ONE batched intake (product / account region / payment means / device), expectation declaration with the price restatement and the buy-version three promises, gears L3→L0 with payment never automated, the worst-case line ({i18n:worst_case}) before any cross-region step, and the purchase-confirmation stop (charge matches → plan active in-app → renewal date + cancel path). Checkpoints: purchased / plan_active. Fabricated identity / address / payment remains a hard refusal (safety.md §2).

Scheduled deal scanning

On a recurring/scheduled run, produce the compact deal brief per commands.md §2 in the user's language (< ~300 words): headline changes, new free credits, expiring promos, price moves, risk alerts. Diff against the last snapshot in vendor-cache.md when one exists; without a baseline, say "first scan — no baseline to diff" and give a plain brief.

Safety floor (never cross)

• Never help steal quota, commit identity fraud (including fabricated local identity/address for region deals), or bypass paywalls via technical abuse.
• Grey / region-locked / reseller channels may be described with steps and clear ⚠️ labels (safety.md), but compliance/ban risk and ToS caveats come first. "You decide; we inform."
• Neutral, conflict-of-interest-free scoring (scoring.md §4).

Resource index

Concern	File
Hunt methodology, query templates, freshness contract (canonical)	references/deal-hunting.md
Coverage map — classes to scan (no prices inside)	references/vendor-registry.md
New-provider radar (brand-new free-token vendors)	references/discovery-sources.md
Neutral scoring: unit normalization, 7 dims + confidence	references/scoring.md
Safety tiers, key hygiene, scam detection	references/safety.md
Commands (canonical), /estimate, scheduled brief, preferences	references/commands.md
Graded register/get-key automation	references/auto-register.md
Graded membership purchase (payment never automated)	references/buy-membership.md
Host capability probe (portable) & guided install	references/capability-check.md
Pre-delivery checklist & golden test cases	references/self-check.md
Agent config fundamentals / international / domestic-OSS / errors	references/agents/*.md
Output strings (en/zh)	references/i18n/en.json, zh.json
Ranking/report templates (compact & full)	references/ranking-template.md
One-page HTML report skeleton	assets/templates/full-report.html
Ready-to-import config template	assets/templates/openai-compatible-config.md
Writable cache & preferences (persistence ladder)	assets/vendor-cache.md