AI Free-Tier
Operating Manual

Before writing this, I specifically verified the "classic free-tier paths" circulating online, and found that a significant portion have already expired, or look free but are no longer what they seem. If you're about to follow an old tutorial, check this table first:

Below, channels that are still alive are listed by category. Every entry follows the format: "open link → do what → get how much → what's the limit."

OpenRouter—open openrouter.ai ↗, filter for models with the :free suffix (currently ~17 on offer, mainly Gemma, GLM, MiniMax, Nemotron, etc.; list scrolls with updates ↗, refer to the real-time list). Free tier rate-limited to 20 req/min; the key operation is to top up your account with at least $10 (even just once), and the daily cap permanently jumps from 50 to 1,000—this is a lifetime unlock, not a balance threshold (official rate-limit docs ↗, rules explanation ↗).

Occasionally, time-limited free top-tier "stealth models" pop up—the latest is Ox Alpha ↗, launched August 20, 2026: 1M-token context, purpose-built for coding and agentic tasks, OpenCode gave it nearly unlimited free usage for a week ↗. Historically, Quasar Alpha (→GPT-4.1) ↗, Horizon Alpha (→GPT-5) ↗, and Sonoma Alpha (→Grok 4 Fast) ↗ were all unmasked within weeks and either went paid or delisted—use them if you spot them, but don't treat them as long-term channels. This example played out especially fast: on August 26, Zhipu (Z.ai) confirmed to media that Ox Alpha was actually their new flagship GLM-5.3-Flash in disguise ↗; the free stealth route was promptly delisted from OpenRouter and replaced with a paid official API.

NVIDIA NIM—register at build.nvidia.com ↗, no credit card required; 1,000 credits auto-credited, enterprise email verification can request 5,000; rate ~40 RPM, covering 80+ models including Llama 4, Qwen, Hermes 3 (developer forum credit discussion ↗).

Groq Cloud—register at console.groq.com ↗; free tier runs open-weight models like gpt-oss-120b/gpt-oss-20b/qwen3.6-27b, rate-limited to 30 RPM / 1K RPD / 8K TPM / 200K TPD; adding a credit card (no charge) boosts limits 10×.

Google AI Studio (Gemini API)—ai.google.dev ↗ register to get free quota for Gemini Flash/Flash-Lite series; however, since April 2026 the Pro series has been removed from the free tier, and rate limits no longer publish fixed numbers—check the console for actual quotas.

Mistral La Plateforme (Experiment tier)—mistral.ai/pricing ↗ free tier covers all models (including Mistral Large, Codestral), limited to non-production evaluation; check Admin Console for specific rate limits.

These providers share a common trait: none require enterprise credentials; phone number / personal identity verification gets you basic free quota, and enterprise verification only raises the quota ceiling.

Don't confuse them: Zhipu's GLM-5.3-Flash, released August 26, 2026, is the new flagship and is not free—320B-parameter native multimodal model, MIT-licensed open weights but API billed per token (50% off promo through Sept 9); the one you can free-ride in the table is GLM-4.7-Flash ↗—they are not the same thing.

MiniMax: No official first-party announcement found for "register and get a fixed token pack"; phone/email/WeChat registration ↗ works, developer门槛 is low, but quota info should be based on actual console display. ModelBest MiniCPM: Focused on open-weight model local/edge deployment ↗, not a stable commercial API free-tier channel; suited for those with local compute.

Hermes Agent (Nous Research, MIT-licensed open source)—hermes-agent.nousresearch.com ↗ deployed on your own server, the software itself is permanently free. It natively supports multiple model providers; the practical approach is to point the model backend at the free channels from sections 1 and 2—OpenRouter's :free pool, NVIDIA NIM's registration quota, GLM-4.7-Flash's permanently free key—so neither the software nor the model calls cost anything (free provider list ↗). The same logic applies to "bring-your-own-key" coding agents like Cline and Continue.dev: the tool itself is free, the model backend it connects to is free, so the entire agent is free.

GitHub Copilot Free—activate directly with a GitHub account; official quota ↗ is 2,000 code completions + 50 chat requests/month; model auto-assigned, cannot manually select.

Cursor Free (Hobby)—register and use directly; official pricing page only says "limited Agent requests / limited Tab completions," no longer publishes specific numbers; new accounts include a 1-week Pro trial. Windsurf Free—unlimited Tab auto-completions, Cascade agent quota roughly enough for 2-3 days of typical coding.

Two pitfall warnings: Gemini CLI is dead (see §00), and its successor Antigravity's free tier ↗ is itself shrinking fast—250 req/day at launch in November 2025, cut to 20 req/day in December 2025, and since March 2026 overage requires purchasing extra credits at $0.01/credit. Amazon Q Developer ↗ technically still works (50 agentic requests/month), but new user registration stopped May 15, 2026 and support ends April 2027—not worth learning now. ByteDance Trae's domestic version (trae.cn) is completely free with no门槛, with built-in Doubao and DeepSeek R1/V3; the overseas version has tiered limits, with reports of sudden tightening in July 2026—test it yourself before relying on it.

Building your own agent isn't just about calling models—you also need to give it "eyes and hands" (web search, web scraping), an "execution environment" (code sandbox), and "memory" (memory layer + vector database). These infrastructure components also have free tiers, and many tutorials haven't mentioned them at all.

Search & Web Scraping APIs

Perplexity API (Sonar series) has no free tier—official pricing docs ↗ are entirely per-token/per-request billing with no mention of any free quota; new accounts must bind a payment method before generating a usable API key. The rumored "Pro subscription grants credit" claim is unconfirmed by official docs.

Code Execution Sandboxes

E2B ↗: Hobby tier one-time $100 usage credit (not monthly reset), max 1 hour/session, up to 20 concurrent sandboxes, 10GiB free storage, no credit card. Daytona ↗: one-time $200 free compute credit, no credit card. Modal ↗: $30 free compute credit/month, monthly reset non-cumulative (see §09).

Browser Automation Services

MultiOn/Multion ↗ has pivoted to consumer personal assistant product "AGI-0" and no longer offers a free tier for developer browser automation APIs; if tutorials still recommend it as an agent browser service, the info is outdated.

Vector Databases & Agent Memory Layers

General vector databases store and retrieve embeddings, while "memory layer" services like Mem0 and Zep additionally handle agent-specific logic such as memory extraction, summarization, and forgetting—different roles:

Chroma Cloud ↗ grants a one-time $5 credit, but the official FAQ notes the Starter tier "requires a credit card"—a rare exception among these options. LangSmith ↗ (LangChain observability platform) free tier includes 5,000 traces/month but no free LangGraph Platform hosted deployment—if you want to free-ride "hosted agent runtime with memory," it can't help; only useful for viewing traces.

Colab/Kaggle in §09 are for "running experiments"; this section is about writing your agent as a long-running online service accessible to others—two different scenarios that people often confuse.

Midjourney ↗ no longer has a free tier—Discord/web free trial was canceled in 2023; no credit card requirement but also no free quota. Specific numbers for Canva Magic Media and ByteDance Jimeng free quotas conflict across online sources; recommend logging in and testing yourself rather than copying numbers from any single tutorial.

Runway ↗: free tier 125 one-time credits (non-renewable non-recurring), export limited to 720p with watermark, does not include Gen-4 official model. Luma AI (Dream Machine) ↗: free tier ~30 generations/month, 5s each, 720p, no audio, with watermark. Pika: free tier limited to 480p with watermark; specific quota numbers vary by third-party source, recommend checking official site in real time.

OpenAI Sora: see §00—free quota canceled, product delisted; stop looking. Domestic Kling and Tongyi Wanxiang are listed in §06 (their video generation quotas share the same credit system as image generation).

PlayHT has shut down (see §00)—acquired by Meta in July 2025, consumer product closed end of that year, official site inaccessible; skip any tutorial still recommending it. Tencent Zhiying's free quota couldn't be first-party verified; confirm by logging into your account.

No extra steps needed—register an account (some don't even require that) and start using; quota ceilings generally don't publish fixed numbers:

ChatGPT Free: official wording "daily text conversations unlimited," subject to abuse-protection limits; image generation/voice/file analysis billed separately. Claude.ai Free: rolling 5-hour window, specific message count not fixed/published, only Sonnet series available, Opus not open. Gemini App / Microsoft Copilot (Bing/Edge sidebar): free, hit ceiling and get "try later" prompt, no public hard number. Perplexity Free: standard web search basically unlimited, Pro Search has limited daily quota (3-5 per source variation). Meta AI: still free; paid subscription "Meta One" launched May 27, 2026 ↗, but official stance: "more casual users will continue to get it free."

GitHub Student Developer Pack ↗—after school email verification, bundles JetBrains full suite (free, ~$289/year value), DigitalOcean $200 credit, Azure $100 credit, plus newly added Camber (40 CPU-hours + 5 GPU-hours + 50 agent messages/month, connectable to Cursor/Claude Code via MCP). Copilot portion has changed dramatically: unlimited Copilot Pro for students canceled March 2026, replaced by 200 AI credits/month; new applications suspended April 20 (reason: excessive agent-mode consumption); existing users retain original benefits, new students currently get standard Copilot Free.

Azure for Students ↗: no credit card, $100 credit renewable once/year, applied against Azure OpenAI/Foundry pay-as-you-go (not an independent free quota).

Independent student verification channels are tightening across the board: Perplexity's ".edu one year free" ↗ expired January 2026, now Education Pro at half-price $10/month; Cursor student free Pro year closed to new users June 25, 2026 (see §00). Domestic Alibaba Cloud "Yun Gong Kai Wu" ↗ grants ¥300 voucher/year + 1-core 2GB ECS free 12 months after identity + Xuexin verification, but note that most "AI large model token quotas" from domestic cloud providers are universal benefits for all new users, not student-exclusive—don't conflate them with student programs.

Hackathon channels: there is no perennial AI credit pool anyone can self-serve at any time; the mechanism is event organizers apply first, then distribute credit codes to registered participants. OpenAI has an official application form ↗ for organizers; Anthropic's Claude Campus Program ↗ targets campus organizations to apply for credits to run events; Groq's own free developer tier is already generous (30,000 tokens/min), many hackathons just use the free tier; MLH partnership page ↗ offers $300 Gemini credits.

No institutional endorsement needed; individuals can self-serve apply:

AWS Activate Founders ↗: $1,000, no VC endorsement needed, just complete company documentation and ≤10 years since founding; credits directly convertible to third-party model usage on Bedrock. Modal: $30 free compute credit/month, long-term valid, included on registration. Google Cloud ↗: $300 trial for new users, 90-day validity, but cannot cover Gemini API's independent billing.

Requires institutional/academic endorsement; individual developers most likely can't get these—skip:

OpenAI Researcher Access Program ↗ ($1,000, requires academic institution endorsement, only 4 approval windows/year); Anthropic Claude for Startups ↗ (amount undisclosed, requires institutional investor funding + company ≤4 years); Together AI/Fireworks AI/Replicate large Startup Programs ($10K-$50K level, all require company identity).

Opera AI (formerly Aria): official page states "free to use without login or registration"; restructuring upgrade since October 2025 remains completely free; what actually costs money is the separate agentic browser Opera Neon at $19.90/month. Brave Leo: free tier requires no account, open and use directly; can call Llama 3.1, Qwen 3, Claude Haiku 3.5, GLM and other models; conversation history stored locally.

Beyond free, there's a category worth including: not $0, but cheap enough that "one free-ride lasts a long time" at standard pricing (excluding first-month promotional pricing). Conclusion first: the threshold (monthly fee ≤¥10 or ≤$3) is strict; image/video generation and cloud services currently have no qualifying standard-price options—everything found is $4+; not forcing entries. The solid finds are concentrated in coding IDEs and model API pay-as-you-go.

Coding IDEs—the only confirmed standard low-price plan: ByteDance Trae's paid entry tier Trae Lite, $3/month ↗, includes base quota + bonus usage, unlimited auto-completions, 2 concurrent cloud tasks; effective since February 2026 and unchanged. Compared to Cursor and Windsurf starting at $20/month, this is the only "long-term standard price" rather than "promotional price" low-cost plan.

Model API pay-as-you-go—converting to "how much does ¥10 buy" is more practical: per-token pricing varies significantly across providers; rather than memorizing free quotas, memorize unit-price conversions:

The conclusion is straightforward: the free GLM-4.7-Flash is itself the best value option; if you must pick a paid one, off-peak DeepSeek ↗ is several times more cost-effective than always-online Kimi ↗, and Qwen-flash ↗ is the cheapest paid option.

Borderline but not fully qualifying, for reference only: iFlytek Spark App and iFlytek Tingxie transcription continuous-pack standard renewal prices are ¥15-18/month (~$2-2.5), meeting the "$3" threshold but exceeding the "¥10" threshold, and first-month promos are often as low as ¥9.9 or ¥6—promotional prices don't count as long-term free-tier; always check the standard price before renewing; Alibaba Tongyi Qianwen's newly launched "Premium Member" in August 2026 is ¥19/month standard, also on the threshold edge—these two pieces of info ↗ couldn't be second-verified within the official app, suggest opening and confirming yourself before subscribing. Closest to the threshold in cloud services is DigitalOcean smallest instance at $4/month ↗, still no standard option truly below $3.

The previous two sections were about "how to spend less"; this section flips the calculation: if you're already using a coding subscription plan like Cursor or Claude Code, is it actually worth it, and is it cheaper than connecting to APIs yourself? The core variable is cache hit rate—each round of a coding agent request is typically "a large chunk of reused code context (should hit cache) + a small new instruction"; the real unit price must be weighted by hit rate, not just the base price listed on the website.

Subscription plans: who's quantifying publicly, who's deliberately obfuscating

No official source has published "this dollar amount equals how many API calls"—third-party estimates suggest Claude Code Max 20x heavy users' actual API-equivalent cost could reach $600-1,500/month, but this is community blog modeling, not official data (third-party estimate ↗).

API + Cache Pricing Panorama (per million tokens, USD)

Sources: Anthropic official pricing ↗ and cache docs ↗, OpenAI pricing page ↗, Google Gemini pricing page ↗, xAI pricing page ↗, Zhipu open platform ↗, Moonshot open platform ↗, DeepSeek pricing page ↗. DeepSeek's hit price is only ~3% of the miss price, the largest discount with the most transparent mechanism; Anthropic and OpenAI are the only two where "cache writes also incur a surcharge" (1.25× base), meaning you pay extra once before hitting.

How cache hit rate changes real cost: assuming 90% hit rate (typical for normal agentic coding sessions), effective input unit price = 0.9 × hit price + 0.1 × base price. By this formula, Claude Sonnet 5 / GPT-5.6 Terra / Gemini 3 Pro effective unit price all drop to $0.38 (19% of base), and DeepSeek V4-Flash (off-peak) drops to $0.028 (13% of base)—each additional 10 percentage points of hit rate benefits models like DeepSeek where "hit price = 3% of base price" the most.

Applied to one conversation round: assume a typical agent round is "50K tokens reused context (90% hit) + 1,500 tokens new output":

Conclusion: Light users (~300 rounds/month) on Sonnet-class models pay only $10-11 in direct API costs—compared to Claude Code Pro at $20/month, pure pay-as-you-go is actually cheaper; the subscription buys "not worrying about bills + higher caps" peace of mind, not savings. Heavy users (1,500 rounds/month) see Sonnet-class API-equivalent cost rise to $51-56, at which point Claude Code Pro or Cursor Pro (both $20) are clearly better deals, and Max/Ultra tiers ($100-200) also start making sense. GitHub Copilot is the only one where you can precisely account: Pro's $15 credit buys ~440 rounds at Sonnet pricing or ~160 rounds at GPT-5.6 Sol—same dollar amount, choosing the cheaper model gets you nearly 3× more usage.