🪙 The Free Token-Maxxing Guide (2026)

A messiah-guide to extracting the maximum amount of frontier-model usage — tokens and compute — for $0, legitimately.

Last verified: June 2026. Free tiers, models, and limits change weekly. Every number below was cross-checked against an official pricing/docs page or the auto-generated cheahjs/free-llm-api-resources list; where a provider hides exact caps behind a login or a JS page, it's flagged.

Contents

• TL;DR — the messiah stack
• 1. What "frontier" means in 2026
• 2. Free frontier in the browser (chat apps)
• 3. Free LLM APIs — the token buffet
• 4. Free coding assistants + the BYO-key power move
• 5. Trial credits — one-time fuel
• 6. Self-host OSS for "unlimited" tokens
• 7. The orchestration layer (wire it together)
• 8. Quota-stretching: caching & batching
• 9. Legit vs risky
• 10. The stacking playbook
• 11. The China play 🇨🇳 — cheapest tokens + free media
• 11.1 "100M tokens for ~$1" — the cache-hit trick
• 11.2 Ultra-cheap "coding plan" subscriptions
• 11.3 Free Chinese model tiers (text + multimodal)
• 11.4 Free image generation
• 11.5 Free video generation (for music videos)
• 11.6 China caveats
• Living lists to bookmark
• Sources

TL;DR — the messiah stack

For a solo creator / power user who wants frontier quality at $0:

One-liner: Chat & long-context → AI Studio. Coding → Gemini Code Assist + CLI. Agentic/overflow → Cline with Gemini-free + Groq-free + OpenRouter :free. Bulk → local models. Glue it with LiteLLM and lean on prompt caching + batch APIs.

1. What "frontier" means in 2026

As of mid-2026, the frontier tier is roughly GPT-5.5 / Gemini 3.1 Pro / Claude Opus 4.8. Free tiers usually hand you the mid model of the current generation (Flash / mini / Sonnet / Haiku) with limited or metered access to the very top model. The two big exceptions where you can touch a genuine frontier model for free: Google AI Studio (Gemini 3 Pro / 3.1 Pro in the browser) and, briefly, trial credits / student offers.

2. Free frontier in the browser (chat apps)

The zero-setup layer. Most give a frontier-ish model with dynamic, unpublished caps that silently downgrade you after a threshold.

Maximize: do long-context, multimodal, and reasoning-heavy work in AI Studio instead of burning capped ChatGPT/Claude messages. Caps on rows without a published number are approximate.

3. Free LLM APIs — the token buffet

Perpetual (or near-perpetual) free API tiers. Numbers verified against cheahjs/free-llm-api-resources and provider docs, June 2026.

Best picks: Gemini (quality) → Cerebras (speed + volume) → Groq (speed) → OpenRouter :free (variety/overflow) → Cloudflare (automation trickle).

Not really free in 2026: Together (min $5 purchase), DeepSeek direct (balance-based), Fireworks (payment profile). Treat as cheap-paid, not free.

4. Free coding assistants + the BYO-key power move

The power move: bring-your-own free key → $0 agentic coding

Run an OSS agent (Cline / Aider / Roo / Gemini CLI) and point it at a free API tier:

1. Quality: Gemini free key — aistudio.google.com/apikey. Free "Gemini 3-class"; CLI gives 60 RPM / 1,000 req/day. (Prompts may train Google.)
2. Speed: Groq free — fast Llama / gpt-oss / Qwen.
3. Overflow: OpenRouter :free — a free DeepSeek/Qwen/Llama variant when the others cap out.

Recommended free stack: Cline (or Aider) + Gemini-free, with Groq-free secondary and OpenRouter :free as overflow. Add MCP servers for tools.

5. Trial credits — one-time fuel

Use once, on one real account, as intended. (From cheahjs/free-llm-api-resources.)

For cloud credits ($300 GCP, $200 Azure, AWS's new $100+$100, Oracle $300), and student/startup programs (GitHub Student Pack, Google for Startups up to $350k, AWS Activate up to $200k), see the main README §6.

6. Self-host OSS for "unlimited" tokens

When you run the model, tokens are free (bounded by hardware). All of these speak the OpenAI-compatible API, so they slot behind the same client code / hub.

Hardware reality: Apple Silicon unified memory lets an M-series Mac address lots of RAM as effective "VRAM," which is why it's the popular local-LLM choice. A 16–24 GB machine comfortably runs 8–14B models (Llama 3 8B, Qwen3-14B, DeepSeek-R1-Distill); 32B+ needs more.

Free GPU to host bigger models / media: Hugging Face Spaces (free high-VRAM GPU time via ZeroGPU), Kaggle (30 GPU-hrs/week), Modal ($30/mo). Full details and the GPU→model map are in the main README §2–§3. Oracle's Always-Free ARM (4 OCPU / 24 GB) makes a perfect 24/7 CPU LLM backend via llama.cpp (8B at ~5–10 tok/s, permanent).

7. The orchestration layer (wire it together)

The whole stack is glued by the OpenAI Chat Completions schema — change only base_url + api_key to repoint any client. Pick one hub:

• OpenRouter (hosted, zero-ops): :free models, a models: [...] array for automatic fallback, and provider routing (sort, data_collection:"deny", ZDR).
• LiteLLM (self-hosted proxy, http://0.0.0.0:4000): unify 100+ providers, load-balance multiple free keys you legitimately own under one model_name, fallback across model groups, and hard-enforce RPM (enforce_model_rate_limits,
• Redis for multi-instance) so you degrade gracefully instead of getting 429-banned.

   Apps / coding tools (Aider · Cline · Roo · your scripts)   ── all speak OpenAI API
                              │
                              ▼
          ┌──────────────────────────────────────────┐
          │  ONE HUB:  LiteLLM proxy  — or — OpenRouter │
          └───────┬───────────────────────┬────────────┘
   routing /      │                       │   :free models, provider routing,
   fallback /     │                       │   model-array fallbacks
   load-balance   ▼                       ▼
   ┌────────────┬───────────────┬───────────────┬──────────────────────────┐
   │ Free tiers │ Trial-credit  │ Your OTHER    │ LOCAL (unlimited):        │
   │ (Gemini,   │ providers     │ legit free    │ Ollama, LM Studio,        │
   │ Groq, …)   │ (Modal, …)    │ keys (yours)  │ llama.cpp, vLLM           │
   └────────────┴───────────────┴───────────────┴──────────────────────────┘

Tiered fallback chain (cheapest-first): local model → free hosted tier (Groq/Gemini/Cerebras) → OpenRouter :free → small paid model as last resort.

7.1 Add a gateway, tracing & tools (all $0)

Once a hub is routing your tokens, three cheap upgrades make the stack production-grade:

• AI gateways (caching, retries, budgets, analytics in front of any provider): Cloudflare AI Gateway (managed, free — caching/rate-limit/logging/fallback; free plan keeps 100k logs, 10 gateways, and BYOK avoids markup), or self-host LiteLLM / Bifrost (Apache/MIT) for virtual keys + spend caps. Portkey & Helicone ship OSS gateways too.
• Observability / evals (see where tokens go, catch regressions): 🥇 Langfuse (MIT — self-host, or 50k-units/mo free cloud) for traces, cost, prompt management and LLM-as-judge; lighter options are Helicone (Apache, 10k req/mo free), Arize Phoenix (local OTel) and OpenLLMetry (Apache, export to any OTel backend).
• MCP (Model Context Protocol) — the open "USB-C for AI": one tool server (files, DBs, APIs, search) plugs into Claude, Copilot, Cline, Cursor. Find free servers in the official modelcontextprotocol/servers repo, the GitHub MCP Registry, Smithery, mcp.so, or the Docker MCP Catalog (300+ containerized — safest). ⚠️ Only attach trusted servers with least-privilege tokens — a malicious MCP server is a prompt-injection / secret-exfil vector.

$0 glue pick: LiteLLM (route) + Langfuse self-host (trace) + Docker MCP Toolkit (tools), plus n8n (self-host) or Dify (OSS) for visual automation / RAG. Need the data + hosting layer too? See the Free Ship-It Stack 🚀 (DBs, vector/RAG, embeddings, hosting, storage, auth — all $0, commercial-use traps flagged).

8. Quota-stretching: caching & batching

Make each free token go further:

• Prompt caching (reuse a fixed prefix cheaply):
• Anthropic — cache_control: {type:"ephemeral"}; cache reads ≈10% of base input (~90% off).
• Gemini — implicit caching auto-on for 2.5+; explicit via caches.create() (guaranteed savings). Put stable content at the start of the prompt.
• DeepSeek — disk caching on by default, no code changes (prompt_cache_hit_tokens in usage).
• Batch APIs (50% off + higher limits) for non-interactive jobs: OpenAI Batch (24-hr turnaround) and Anthropic Message Batches (most finish < 1 hr).
• Trim & stabilize: keep a stable system-prompt prefix (maximizes cache hits + batch efficiency), summarize old turns, pin num_ctx locally to avoid silent truncation/OOM.

9. Legit vs risky

✅ Clearly legit & sustainable

Official free tiers (AI Studio, Groq, Cerebras, Cohere, NIM, Mistral, Cloudflare, GitHub Models) · trial credits used as intended on one real account · OpenRouter :free · OSS self-hosting · student/startup programs you actually qualify for.

Legit ≠ private. Many free tiers train on your inputs (Google AI Studio free, Mistral's free Experiment plan, etc.). Never paste secrets, proprietary, or sensitive code into a free tier.

⚠️ Grey / risky — flagged, not endorsed (no how-to)

10. The stacking playbook

A realistic $0 routine for a solo AI creator:

1. Daily driver: Google AI Studio for ideation/long-context/multimodal; Gemini Code Assist + CLI for coding.
2. Agentic coding: Cline/Aider → Gemini-free → Groq-free → OpenRouter :free, behind LiteLLM with per-key RPM caps + fallbacks.
3. Bulk / iterative grunt work: local Ollama/LM Studio (lint fixes, drafts).
4. Speed bursts: Cerebras (1M tok/day) / Groq.
5. GPU jobs (media, fine-tunes): Modal ($30/mo) for serverless bursts; Kaggle (30 h/wk) for long headless runs; Oracle A1 as a free 24/7 backend.
6. Stretch quotas: prompt caching (~90% off) + batch APIs (50% off) on stable prompts.
7. One-time fuel: spend trial credits (Modal/Baseten/SambaNova/Scaleway…) on real deployable workloads; claim GitHub Student Pack / startup programs if eligible.
8. Hygiene: set budgets, disable auto-recharge, delete idle GPU/volumes, and never feed secrets to a training-on free tier.

11. The China play 🇨🇳 — cheapest tokens + free media

Chinese labs are in an all-out price/feature war, which is fantastic for a token-maxxer. Three things stand out: near-free token pricing via context caching, flat coding subscriptions at a fraction of Claude's price, and genuinely free image/video (especially via open weights). The trade-offs — data jurisdiction, watermarks, ToS — are real; see §11.6.

Versions move fast. As of June 2026 the current models are GLM-5.1 / 4.7, Kimi K2.7, MiniMax M3, DeepSeek V4, Kling 3.0, Hailuo 2.3, Seedance 2.0, CogVideoX-3, Wan 2.2, HunyuanVideo 1.5. Re-verify names before relying on them.

11.1 "100M tokens for ~$1" — the cache-hit trick

The viral claim is real, but only for input tokens served from cache — not output. The mechanism is context caching: a provider bills a repeated prompt prefix (system prompt, long doc, codebase, chat history) at ~1–2% of the normal price. Agentic coding naturally creates this, because every tool-call resends the same growing context. (DeepSeek prices verified on its official pricing page, 2026-06-12; the prefix cache is automatic — it only matches an identical prefix from token 0.)

Where the meme came from: Hacker News users (May 2026) reported running DeepSeek V4 Pro + opencode and burning ~100M tokens for ~$2 in a day, "majority cache tokens." Plausible against official pricing with a high cache-hit ratio and limited output.

The honest version: fresh input is ~$0.14/M (≈7M/$1) and output is 100–300× the cache-hit price, so a real mixed bill is higher. Mechanisms that actually work: automatic context cache (DeepSeek, Doubao, Kimi) + batch APIs (~50–60% off, non-real-time). ⚠️ A specific off-peak discount could not be confirmed on DeepSeek's live 2026 page (it existed historically) — treat off-peak as unverified.

11.2 Ultra-cheap "coding plan" subscriptions

Chinese labs sell flat monthly coding plans that undercut Claude Code massively. The wiring is officially supported, not a hack: most expose an Anthropic-compatible endpoint, so you point Claude Code / Cline / Roo at it:

// ~/.claude/settings.json  (Z.ai GLM example, per docs.z.ai)
{ "env": {
  "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
  "ANTHROPIC_AUTH_TOKEN": "<your_key>",
  "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.1",
  "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-4.7"
}}

⚠️ Alibaba Qwen's free coding tier was discontinued 2026-04-15 (1,000→100→0 req/day); Qwen Code now needs a paid plan or BYO key.

11.3 Free Chinese model tiers (text + multimodal)

🏆 The standout: Zhipu's *-Flash family — the only major Chinese lab with a genuinely $0 API spanning text, vision, and image (and historically video). Sign up on the international Z.ai portal (intl email/card OK).

Best $0 self-host weights (run free on free GPU — see §6): Qwen3, DeepSeek-V3/R1, GLM-4.5/4.7, Hunyuan-A13B, ERNIE 4.5, Kimi K2, Yi, MiniMax-M1.

11.4 Free image generation

Verdict: self-host Kolors (top quality, clean license, runs on free GPU) or hit the CogView-3-Flash API (no server needed) for commercial-grade cover art and thumbnails at $0.

11.5 Free video generation (for music videos)

The key insight: free app tiers almost all watermark and forbid commercial use — so for monetized music videos, the free and legally-clean path is OPEN WEIGHTS, run free on free GPU (HF ZeroGPU / Kaggle / Colab / Modal) via ComfyUI or Diffusers.

🔓 Self-host open weights — free, no watermark, commercial:

📱 Free app tiers (easy, but watermark + personal/non-commercial — good for drafts):

Verdict for music videos: 1. 🥇 Wan 2.2 TI2V-5B (self-host) — cleanest license (Apache-2.0), 720p/24fps, runs on a free HF ZeroGPU Space or a 4090. Publishable & monetizable. 2. 🥇 CogVideoX-2B (self-host) — runs on a free Colab T4; churn many short clips to cut to a beat. 3. Kling / Hailuo free apps — fast drafts & storyboards (watermarked). 4. Cheapest paid fallback: Zhipu CogVideoX-3 API ~$0.20/video (4K/60fps/audio).

🎙️ Need the soundtrack too? Video is only half a music video — see the companion Free Audio Generation Guide for free, commercially-clean music (ACE-Step), voice/TTS, and lyric-timing/transcription (WhisperX), and the Media Post-Production Guide 🎬 to finish it (4K upscale, 60fps, mastering, subtitles, dubbing).

⚠️ "CogVideoX-Flash free" is not confirmed for 2026 — there's no free video model on international Z.ai (CogVideoX-3 is paid). The free Flash video tier was a Chinese-BigModel feature and may be retired; verify before relying on it.

11.6 China caveats

• Data jurisdiction: these run under PRC jurisdiction and inputs may be used to train models. Don't send secrets, proprietary, or regulated code/codebases.
• Watermark + non-commercial on free app tiers: for anything you'll monetize (incl. YouTube music videos), use open weights (Wan / CogVideoX / Hunyuan / Kolors) — free and commercially clean.
• Don't prompt copyrighted characters: Disney/Universal/Warner sued MiniMax (Sept 2025) over Hailuo reproducing IP. Generate original characters only.
• ToS: Z.ai's coding plan restricts use to officially supported tools and is individual-use only (repeated violations → ban). One legit account per service — Anthropic publicly accused MiniMax (Feb 2026) of using thousands of fraudulent accounts; don't be a cautionary tale.
• Geopolitics/availability: some orgs prohibit sending code to Chinese endpoints; access and pricing can change (Qwen's free tier vanished; GLM's entry price rose). Verify current terms before committing.

Living lists to bookmark

• cheahjs/free-llm-api-resources — the auto-generated, legit-only list of free LLM API tiers + trial credits.
• ripienaar/free-for-dev — huge general free-tier catalog for developers (incl. AI/ML).

Sources

Compiled June 2026 from a multi-model research sweep (GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro), with every number cross-checked against primary docs:

• Free LLM tiers & limits: github.com/cheahjs/free-llm-api-resources
• Gemini API pricing / rate limits: ai.google.dev/gemini-api/docs/pricing, /rate-limits
• Gemini Code Assist: blog.google/.../gemini-code-assist-free/, developers.google.com/gemini-code-assist/resources/quotas
• Gemini CLI: github.com/google-gemini/gemini-cli
• GitHub Copilot Free + signup pause: docs.github.com/copilot/get-started/plans
• OpenRouter: openrouter.ai/docs (free variants, fallbacks, provider routing)
• LiteLLM: docs.litellm.ai/docs/proxy/load_balancing, /reliability
• Cerebras: inference-docs.cerebras.ai; Groq: console.groq.com/docs/rate-limits
• Cloudflare Workers AI: developers.cloudflare.com/workers-ai/platform/pricing
• Prompt caching / batch: platform.claude.com/docs, ai.google.dev/gemini-api/docs/caching, developers.openai.com/api/docs/guides/batch
• Local runtimes: ollama.com, lmstudio.ai, github.com/ggml-org/llama.cpp, github.com/vllm-project/vllm
• Modal pricing: modal.com/pricing
• China — token pricing: api-docs.deepseek.com/quick_start/pricing, /news/news0802 (context cache), /news/news250929 (V3.2 DSA), volcengine.com/docs/82379/1544106 (Doubao)
• China — coding plans: docs.z.ai/devpack, platform.minimax.io/docs/token-plan, platform.kimi.ai/docs/pricing, api-docs.deepseek.com/guides/anthropic_api
• China — free model tiers: docs.bigmodel.cn/cn/guide/models/free/* (GLM/CogView/CogVideoX Flash), cloud.tencent.com/document/product/1729/97731 (Hunyuan 1M/yr), alibabacloud.com/help/en/model-studio/new-free-quota
• China — image/video (open weights): github.com/Kwai-Kolors/Kolors, github.com/Tencent/HunyuanDiT, github.com/Wan-Video/Wan2.2, github.com/THUDM/CogVideo, github.com/Tencent-Hunyuan/HunyuanVideo; apps: kling.ai, hailuoai.video

Numbers are point-in-time (June 2026) and change frequently — verify before relying on them. Corrections via PR welcome.

📝 Spotted a stale quota or a license that changed? This guide is open source — edit it on GitHub.