Free AI Coding Tools
2026 Token Guide

01 · What "free" actually means in June 2026

The AI coding market in mid-2026 is not short on generosity—it is short on clarity. Vendors advertise "free" while hiding credit-card gates, subscription bundling, geographic blocks, or hard cutoffs buried in OAuth changelogs. A tool can be free to install yet expensive to run if its default model routes through premium output pricing. Conversely, a paid subscription like Claude Pro can feel "free" to Claude Code users because the CLI is bundled—until you hit weekly rate limits on Opus-tier models.

This guide treats free as: no credit card required at signup, or a permanent registration grant large enough for weeks of solo development. We exclude time-limited trials that auto-convert to paid plans without explicit consent. We also flag policy cliffs—most urgently, Google’s June 18, 2026 shutdown of subscription OAuth for Gemini CLI, covered in our Gemini CLI policy analysis.

The payoff for reading carefully: a student stacking Copilot Student, Cursor Hobby, and a SiliconFlow API key can realistically run a semester project at $0 API spend. A startup evaluating three terminal agents can burn through 70 million free tokens on Alibaba Model Studio before opening a corporate card. The tables and steps below turn those numbers into a repeatable workflow—not a marketing headline.

02 · Three adoption pain points

1. Mixing free OAuth with production credentials. Gemini CLI, Codex CLI, and Claude Code all cache OAuth tokens in your home directory. Installing all three on a MacBook that also holds production AWS keys, App Store signing certificates, and a personal Google account creates a single blast radius. One misconfigured cron job can route premium-model pricing through a "free" CLI session. The fix is not abstinence—it is isolation on a disposable macOS node, the same pattern we document for Agent Skill trials on rented Macs.

2. Treating IDE free tiers as full agent replacements. Cursor Hobby and Copilot Free are excellent for tab completion and light chat—2,000 tab completions and 50 slow premium requests per month each—but they are not substitutes for terminal agents that edit ten files, run tests, and open pull requests. Teams that standardize on IDE free tiers alone hit ceiling by week two of a sprint, then blame "AI doesn't work" when the real issue is tier mismatch.

3. Ignoring geography and ToS boundaries. Gemini CLI OAuth is personal-use only; routing that token through a third-party proxy violates Google’s terms. China-based developers cannot rely on Western OAuth flows at all without unstable VPNs. The sustainable path is pairing open tools (OpenCode, OpenClaw, Codex with config overrides) with domestic OpenAI-compatible endpoints—not tunneling a Google session through a relay.

03 · Free tier snapshot table (June 2026)

Use this matrix as your weekly sanity check. Quotas change; verify on vendor consoles before production cutover.

Tool / Platform	Free allowance	Credit card	China direct	Best free angle
Gemini CLI (OAuth)	~1,000 req/day, 60 RPM	No	VPN required	No card; Gemini 2.5 Pro/Flash until June 18 cutoff
Gemini API (AI Studio)	1,500 Flash/day; 100 Pro/day	No	VPN required	Permanent free tier on Google AI Studio
Codex CLI	ChatGPT free account (limited)	No	Proxy or domestic API	Cloud sandbox; GPT-5.x Codex route
Claude Code	Bundled with Pro/Max ($20+/mo)	Yes (sub)	Proxy or relay API	Best quality; use domestic relay for cost
OpenCode	Tool free; pay per API	No	Yes with domestic API	97K+ stars; 75+ providers
OpenClaw	Tool free; multi-provider	No	Yes with domestic API	Reuses Gemini OAuth; gateway bots
Copilot Free	2,000 completions + 50 premium/mo	No	Yes	Instant enable on GitHub
Copilot Student	300 premium req/mo (Pro-equiv)	No	Yes	$10/mo value; .edu verification
Cursor Hobby	2,000 tabs + 50 slow premium/mo	No	Yes	Full VS Code fork experience
SiliconFlow	20M tokens (permanent grant)	No	Yes	DeepSeek, Qwen, GLM routes
Alibaba Model Studio	70M tokens (70+ models)	No (real-name)	Yes	Widest domestic model catalog
Zhipu GLM	20M tokens (permanent)	No	Yes	Stable GLM-5; Claude relay option
Groq	14,400 req/day	No	VPN required	Fast Llama/Mixtral inference

Three citeable totals anchor the table: 1,000 daily Gemini CLI OAuth requests, 70 million Alibaba registration tokens, and 20 million permanent SiliconFlow tokens—enough for months of iterative coding if you avoid whole-repo scans.

04 · Gemini CLI — the generous cliff

Gemini CLI remains the most generous official terminal agent for OAuth users: a plain Google account, no credit card, roughly 1,000 requests per day at 60 requests per minute, with access to Gemini 2.5 Pro and Flash inside a 1 million token context window. Install via npm, Homebrew, or npx:

Per-model OAuth ceilings differ: Gemini 2.5 Pro allows about 100 requests per day inside the CLI quota, while Flash-Lite tolerates up to 1,000 daily requests—use Flash for drafts, Pro for merges. Check session stats with /stats model before assuming unlimited Pro access.

Critical deadline: On June 18, 2026, Google disables subscription OAuth for free, Pro, and Ultra Google AI accounts. After that date you need a paid GEMINI_API_KEY, an enterprise license, or a migration to Antigravity (closed source, ~20 agent turns/day on free tier versus the old ~1,000). Plan B should be live before the cutoff—see our policy article for the five-step migration HowTo.

05 · Codex CLI — ChatGPT OAuth plus domestic escape hatch

OpenAI’s Codex CLI is a Rust-based terminal agent with OS-level sandboxing—roughly 83,000 GitHub stars in June 2026. Free-tier access flows through Sign in with ChatGPT on a free OpenAI account, though limits are tighter than Plus/Pro bundles and can change without notice.

Developers without reliable Western connectivity can point Codex at domestic OpenAI-compatible endpoints. Edit ~/.codex/config.toml:

Export OPENAI_API_KEY with your SiliconFlow or Zhipu key, then run codex doctor to validate routing. You lose OpenAI’s cloud sandbox but gain stable latency and zero VPN dependency.

06 · Claude Code — quality tier, not zero-dollar tier

Claude Code is not genuinely free: it ships inside Claude Pro ($20/month) or Max subscriptions. That said, it remains the quality ceiling many teams escalate to after exhausting free routes—especially for multi-file refactors that show up as 606 billion weekly tokens on OpenRouter’s CLI board per our June 2026 CLI rankings article.

Cost-sensitive teams route Claude Code through domestic relays by setting ANTHROPIC_BASE_URL and ANTHROPIC_API_KEY in ~/.claude/settings.json or shell exports. Zhipu and SiliconFlow both expose Anthropic-compatible gateways at fractions of direct Opus pricing. Never run /init on a monorepo—it ingests the entire tree in one shot and can burn hundreds of thousands of tokens before the first edit.

07 · OpenCode — open router, your API bill

OpenCode (MIT, 97,500+ GitHub stars) is free software that bills only what your chosen provider charges. It supports 75+ providers, Docker sandboxes, and headless CI—making it the default recommendation for China-based developers who want a terminal IDE without Western OAuth.

Recommended zero-cost stacks in June 2026: Gemini API free tier (1,500 Flash requests/day), SiliconFlow registration grant, local Ollama with Qwen3 8B on Apple Silicon, or OpenRouter’s permanently free model routes. Use Tab to switch between Plan and Build modes—Plan on cheap models, Build on capable ones.

08 · OpenClaw — gateway that recycles free OAuth

OpenClaw is an open-source agent gateway—not a pure coding CLI—that shines when you want Telegram bots, cron skills, and multi-model routing in one daemon. It can reuse Gemini CLI OAuth (while that path still works) and accept pasted Anthropic setup tokens from Claude Code.

For persistent 7×24 gateways, pair OpenClaw with a rented Mac mini M4 rather than a laptop that sleeps—see our OpenClaw on rented Mac mini M4 guide. OpenClaw’s models scan command surfaces OpenRouter free models automatically, complementing the billing insights from weekly OpenRouter token rankings.

09 · GitHub Copilot and Cursor — IDE-native free tiers

GitHub Copilot

Copilot Free requires only a GitHub account: enable it under Settings → Copilot, install the VS Code or JetBrains extension, and you receive 2,000 completions plus 50 premium requests per month. Copilot Student (GitHub Education verification) upgrades to Pro-equivalent features—300 premium requests monthly, a $10/month value—without payment. Note: Copilot Pro new signups were temporarily paused around April 2026; confirm the live button state before promising student onboarding.

Cursor

Cursor Hobby mirrors Copilot’s generosity: 2,000 Tab completions and 50 slow premium requests monthly, full VS Code compatibility, no credit card. Students can claim one year of Cursor Pro via cursor.com/students with a .edu email or GitHub Education status. Track usage at app.cursor.sh/account/usage—heavy agent mode exhausts the 50-request pool in days if used as a full CLI replacement.

IDE tier	Tab completions	Premium agent requests	Best for
Copilot Free	2,000 / month	50 / month	GitHub-centric light coding
Copilot Student	Unlimited completions	300 / month	University projects
Cursor Hobby	2,000 / month	50 slow / month	VS Code power users evaluating AI
Cursor Student Pro	Unlimited tabs	500 fast / month	One-year Pro trial

10 · China-region free APIs — no VPN required

When Western OAuth is unreachable, domestic OpenAI-compatible platforms carry the stack. All figures below reflect June 2026 registration grants; promotional campaigns can temporarily increase them.

Provider	Registration grant	Endpoint	Notable models
SiliconFlow	20M tokens (permanent)	api.siliconflow.cn/v1	DeepSeek-V3, Qwen3.5, GLM-5
Alibaba Model Studio	70M tokens	dashscope.aliyuncs.com/compatible-mode/v1	Qwen3.5-Max, DeepSeek-V3
Zhipu GLM	20M tokens (permanent)	open.bigmodel.cn/api/paas/v4	GLM-5, GLM-4.7-Flash
Infinigence AI	Promotional billions-tier grants	Platform-specific	GenStudio free API tier
NVIDIA NIM	Free after phone verify	build.nvidia.com	Llama, Nemotron, DeepSeek

Set 80% usage alerts in each vendor console. A single benchmark runaway is cheaper to catch at 16 million tokens than at 20 million.

11 · Ten token-saving tactics

1. Never /init entire repositories in Claude Code or Codex—scope to single files or directories.
2. Default to Flash-tier models for exploration; promote to Pro/Opus only for merge-ready output.
3. Enable provider alerts at 80% of free grants on SiliconFlow, Alibaba, and Zhipu consoles.
4. Rotate separate Google accounts only where ToS permits (AI Studio API keys)—never proxy Gemini CLI OAuth.
5. Run local Ollama (Qwen3 8B, DeepSeek Coder V2 16B) for unlimited inner loops on 16 GB+ Macs.
6. Shorten system prompts in OpenClaw and OpenCode configs—every turn pays the overhead tax.
7. Scan OpenRouter free models via openclaw models scan before paying for premium routes.
8. Disable web search in Codex when using domestic endpoints with unstable outbound connectivity.
9. Batch related edits into one agent session instead of ten micro-sessions that re-upload context.
10. Track vendor promo calendars—Chinese cloud providers run seasonal grants that temporarily double free pools.

12 · Five-step isolated Mac trial (HowTo)

Free tiers are easiest to abuse—and easiest to misconfigure—on your daily driver. Run this protocol on a rented Mac before promoting any CLI to your team standard.

1. Rent an isolated macOS node. Book Mac mini M4 or MacBook Pro M3 through MacDate; SSH per the daily Mac rental FAQ. Create a local user with no production Apple ID.
2. Install two to three CLI candidates. Example stack: Gemini CLI (OAuth, pre-June 18), Codex CLI with SiliconFlow config, and OpenCode. Store keys in project-scoped .env files—never global shell profiles.
3. Run identical benchmark tasks. Pick a 12k-token task with three tool calls (read, edit, test). Log prompt tokens, completion tokens, wall time, and rate-limit errors per CLI.
4. Document free-tier ceilings. Record daily request counts from /stats model, provider dashboards, and OpenRouter usage CSV if proxied.
5. Revoke and release. Delete OAuth tokens, uninstall CLIs, revoke test API keys, export cli-benchmark-YYYYMMDD.csv, and wipe the rental per MacDate return checklist.

13 · Mac rental hardware tiers for free-tier trials

Tier	Hardware	RAM	Ideal free-tier workload	Limitations
Light	MacBook Air M3	16 GB	Copilot/Cursor eval, Aider, API-only CLIs	No Docker; single agent loop
Medium	Mac mini M4	16–24 GB	Gemini CLI + OpenCode side-by-side	Ollama 8B only at 16 GB
Heavy	MacBook Pro M4 Pro	32 GB+	OpenClaw gateway + Ollama 14B fallback	Local 32B models still impractical
Lab	Mac Studio M4 Ultra	64 GB+	Parallel CLI shootouts + local Qwen3 32B	Overkill for IDE-only trials

Pricing lives on bare-metal macOS pricing. Most free-tier evaluations finish in one to three rental days on Mac mini M4 16 GB—enough to stack Gemini OAuth, SiliconFlow, and Copilot Student without CapEx commitment.

14 · FAQ

Q: Is Gemini CLI still free after June 18, 2026? Subscription OAuth ends on that date for free, Pro, and Ultra Google AI accounts. You can keep the open-source binary with a paid GEMINI_API_KEY or migrate to Antigravity, Claude Code, or Codex. Plan the switch now—not on June 17 at midnight.

Q: Can I use Copilot Free for serious development? For autocomplete and occasional chat, yes. For multi-file agent refactors, you will exhaust 50 premium requests quickly. Pair Copilot Free with a terminal CLI on free API credits for agent work.

Q: What is the fastest zero-dollar stack for China? Register SiliconFlow and Alibaba Model Studio, install OpenCode, export domestic base URLs, and optionally add Cursor Hobby for IDE tabs—all without VPN.

You can install every tool on a personal MacBook and hope discipline prevents credential collisions—but free OAuth tokens, Gemini cached sessions, and OpenClaw LaunchAgents do not respect hope. WSL and Linux VPS handle some Node gateways, yet they break down on TestFlight signing, Keychain certificates, and Claude Code’s macOS Seatbelt sandbox. A rented Mac is a forensic clean room: if a free tier misroutes into Opus pricing during a cron loop, the blast radius ends when you release the instance. If Gemini’s June 18 cliff pushes you toward domestic APIs, you can validate Codex and OpenCode routing on native macOS paths before touching production laptops.

For Apple-platform teams, the rental loop also keeps pace with toolchain churn—Swift 6 concurrency, Xcode 26 betas, and OpenClaw gateway updates ship faster than vendor free-tier changelogs. Convert zero-budget curiosity into auditable evidence on disposable silicon; promote one winner only after CSV benchmarks, not after a README promise of "unlimited free."