# Quota Notes -- Paid AI, 2026 *Last reviewed: 10 September 2026.* This file tracks **how paid AI tools meter usage**, not what they cost. Headline prices have barely moved in 2026 -- $20/month is still the anchor for a pro plan. What changed is the unit underneath it. It exists because the free/paid boundary this repo tracks is no longer drawn by price. Several tools on the free list are free in the sense that matters (you can start without a card) while quietly metering the part you actually want, and the `free_tier` column can't carry that nuance on its own. --- ## The shift, in one paragraph Flat-rate ended in 2026. Through 2025 a subscription bought an unlimited-ish allowance of *requests*; by August 2026 nearly every major vendor meters either **dollars of inference** or **time-windowed compute**, and most run two limits at once -- a short rolling window to smooth load, plus a weekly or monthly cap to bound cost. The second consequence matters more than the first: cheap, high-volume features (inline completions, autocomplete) are being kept free and unlimited, while *agentic* work -- chat, agents, code review, headless runs -- is what gets metered. The meter moved from "how often do you ask" to "how much compute did you consume". --- ## Timeline | Date | Change | |---|---| | **14 May 2024** | Google AI Studio introduces Gemini 1.5 Flash free tier: 15 RPM, 1M TPM, 1,500 RPD, and 1M context. | | **Mar 2026** | Anthropic reduces Claude Code 5-hour limits during weekday peak hours (5-11 AM PT). | | **19 Mar 2026** | Windsurf retires credits entirely, moving to daily + weekly quotas. Pro goes $15 -> $20. | | **1-7 Apr 2026** | Google removes Gemini Pro models (2.5 Pro / 3 Pro / 3.1 Pro) from the AI Studio free tier; free tier becomes Flash/Flash-Lite/Gemma only. | | **Apr 2026** | GitHub pauses new individual Copilot signups; reopens gradually from 17 June. | | **6 May 2026** | Anthropic permanently doubles Claude Code's 5-hour limits (weekly cap unchanged). | | **13 May 2026** | Claude Code weekly limits run 50% above published standard -- extended three times, most recently on 19 Aug through 31 Aug 2026, with Anthropic saying it hopes to make it permanent. | | **15 May 2026** | Amazon closes new Amazon Q Developer free-tier signups; full end of support 30 Apr 2027. Kiro is the replacement. | | **1 Jun 2026** | GitHub Copilot replaces premium-request quotas with token-based **AI Credits** on all plans. Inline completions and next-edit suggestions stay free and never touch the balance. | | **15 Jun 2026** | Anthropic splits programmatic use off the subscription: Agent SDK calls, `claude -p` headless mode and third-party Agent-SDK tools draw from a **separate monthly credit pool** ($20 Pro / $100 Max 5x / $200 Max 20x). | | **17 Jun 2026** | Google AI Studio standardises active Flash models (3.6 Flash, 2.5 Flash-Lite) at 15 RPM, 1M TPM, 1,500 RPD with up to 2M context. (An earlier version of this file claimed 2.5 Pro/Flash endpoints were *retired* on this date -- wrong; that date traces to the June 2025 release. The real tightening was 1-7 Apr above and 19 Jun below.) | | **19 Jun 2026** | Google blocks unrestricted AI Studio API keys and stops publishing static free-tier quotas; limits are in-account only from here on. | | **Mid-Jul 2026** | OpenAI drops the 5-hour Codex cap alongside GPT-5.6 Sol, leaving a single weekly quota. | | **25 Aug 2026** | OpenAI restores the 5-hour cap for Codex and ChatGPT Work on **Plus only**; Pro ($100/$200) keeps it disabled. | | **Sep 2026** | Hugging Face Inference Providers moves free access from open rate limits to credits: **$0.10/month** of compute credits, pay-as-you-go beyond. | | **Sep 2026** | Kling ties daily login credit grants to paying subscribers; the free tier drops to ~66 credits/month (the old "66 credits/day" figure traced to the $1=66-credit purchase rate). | | **Sep 2026** | ElevenLabs re-denominates the free tier from 10k chars/month to 10k unified credits usable across TTS, STT, SFX and music. | --- ## How each tool meters, as of August 2026 | Tool | Plans | Metered by | Notes | |---|---|---|---| | **Google AI Studio (Gemini)** | Free API key (per-project) / Pay-as-you-go | **Tri-metered**: RPM (15), TPM (1M), RPD (1,500) | Limits apply at the **Google Cloud project level** across all keys. Flash models get 1,500 requests/day, up to 2M tokens context, and free 2GB temporary file uploads. **Sep 2026 caveats**: Pro models are no longer on the free tier at all (since Apr 2026); unrestricted API keys are blocked (since 19 Jun 2026); Google no longer publishes these static figures -- verify them in-account. **Privacy tradeoff**: on the free tier, prompts and responses may be human-reviewed and used to train Google products; attaching billing removes training use but converts calls to pay-as-you-go. | | **Claude Code** | Pro $20 / Max 5x $100 / Max 20x $200 | 5-hour rolling window **and** a weekly cap on active compute hours | Window starts on your first prompt, not a fixed clock. The bucket is **shared** across Claude Code, claude.ai and Cowork. Roughly 10-45 prompts per window on Pro, up to ~900 on Max 20x. Programmatic use is a separate pool (see 15 Jun). | | **ChatGPT / Codex** | Plus $20 / Pro $100 / $200 | Rolling window + weekly quota, per surface | Plus: ~160 GPT-5.5 messages/3h, ~3,000 GPT-5.5 Thinking/week. Codex 5-hour cap is back on Plus as of 25 Aug; Pro is exempt. Overflow is buyable as credits. | | **GitHub Copilot** | Free / Pro $10 / Pro+ $39 / Max $100 | Token-based **AI Credits** | Pro includes $10/mo of credits, Pro+ $70. Chat, agent mode, code review and Copilot CLI draw down; **completions do not**. Free plan: 2,000 completions/month unmetered -- chat/agent/CLI draw on a monthly AI Credits allowance whose size GitHub no longer publishes (the old "50 chats/month" figure has been dropped from the docs). | | **Cursor** | Free (Hobby) / Hobby $10 / Pro $20 / Pro+ $60 / Business $40/seat / Ultra $200 | Monthly credit pool | Replaced the old "fast requests" count -- the change that caused the most confusion. Frontier models sit behind Max Mode multipliers, which accelerates burn. Hobby limits are account-specific and not published as fixed numbers. | | **Windsurf** | Free / Pro $20 / Max $200 / Teams $40/user | Daily + weekly quotas | Credits gone since 19 Mar. Free keeps unlimited Tab autocomplete and inline edits plus ~25 Cascade Flow Actions/month. | | **OpenCode Go** | $5 first month, then $10/mo | **Dollar-denominated**: $12 per 5 hours, $30/week, $60/month | The clearest example of the shift -- the quota is literally priced in dollars of inference. One key, 18 open models via OpenCode Zen. Overflow falls back to your Zen balance if "Use balance" is on. Current model agreement runs through 31 Aug 2026. | --- ## Claude & Claude Code -- why it's worth a note Claude’s quota system is the most sophisticated — and frequently misunderstood — dual-metered system in production: 1. **The Shared Pool Trap:** A single usage bucket is shared across **claude.ai web chat**, **Claude Code** terminal CLI, and **Cowork**. Intensive agentic refactors or test-fix loops in Claude Code will deplete your 5-hour interactive allowance on the web interface. 2. **Compute-Weighted Consumption:** The 5-hour rolling window is **not a message counter**. It meters active compute tokens. A prompt with a 150k-token repository context or extended thinking burns through the window dramatically faster (often 10–15 prompts) than lightweight text requests (up to 45 prompts). 3. **The 5-Hour + Weekly Ceiling:** Pro ($20) enforces both a 5-hour rolling prompt window and a weekly active compute hour cap. Max plans (Max 5x at $100/mo and Max 20x at $200/mo) scale the rolling window up to ~900 prompts per window and scale the weekly ceiling accordingly. 4. **Programmatic / Headless Split (15 Jun 2026):** To prevent automated CI/CD loops from paralyzing interactive subscriptions, Anthropic split programmatic use (`claude -p`, Agent SDK) into dedicated monthly dollar pools ($20 on Pro, $100 on Max 5x, $200 on Max 20x). 5. **Free Tier:** Claude's free tier on claude.ai enforces dynamic 5-hour rolling limits (~10–30 messages per 5h) that throttle down during US weekday peak hours, falling back to smaller models when capacity is constrained. --- ## ChatGPT & Codex -- why it's worth a note OpenAI meters ChatGPT and its Codex coding agent through a combination of rolling windows and weekly model-specific quotas: 1. **The 5-Hour Cap Fluctuation on Plus ($20):** - **Mid-Jul 2026:** OpenAI removed the 5-hour rolling Codex cap during the GPT-5.6 Sol launch, consolidating usage into a single weekly quota. - **25 Aug 2026:** OpenAI **restored the rolling window cap on Plus only** (~160 messages every 3 hours) after automated agent usage surged. - **Pro Exemption:** Pro users ($100 / $200/mo) remain completely exempt from rolling 3h/5h caps, running unthrottled compute against monthly fair-use boundaries. 2. **Weekly Reasoning Quotas:** Standard messaging is separated from deep reasoning models (e.g. GPT-5.5 Thinking), which carry strict weekly caps on Plus (~3,000 thinking messages/week). 3. **Free Tier:** The free tier on chatgpt.com provides dynamic access to frontier models (falling back to mini models after ~10–16 messages per 3 hours during peak times). --- ## OpenCode & OpenCode Go -- why it's worth a note OpenCode demonstrates how to cleanly separate open-source tooling from inference subscriptions: 1. **Free Open-Source Agent:** The OpenCode CLI/TUI is open source, completely free, and unmetered. It supports 75+ LLM providers via Bring-Your-Own-Key (BYOK), including zero-cost local inference through **Ollama**. 2. **OpenCode Go ($10/mo):** An optional flat-rate subscription providing access to 18 hosted frontier open coding models (DeepSeek V4 Pro, Qwen3.8 Max, Grok 4.5, GLM-5.2, Kimi K3, GPT 5.6 Luna) through OpenCode Zen. 3. **Dollar-Denominated Metering:** Unlike artificial "credits," OpenCode Go is metered directly in dollars of compute: **$12 per 5 hours**, **$30 per week**, and **$60 per month**. When enabled, overflow draws smoothly against a prepaid OpenCode Zen balance instead of hard-stopping midway through a code refactor. --- ## Google AI Studio -- why it's worth a note Google AI Studio represents the opposite extreme of the subscription clampdown: while developer coding subscriptions moved to restrictive 5-hour rolling windows or token credit burn-downs, Google continues to offer one of the most generous free developer tiers in the industry. *Sep 2026 update: that claim is softening. Pro models left the free tier (Apr 2026) and unrestricted API keys were blocked (19 Jun 2026); the static figures below are the last published values -- Google now shows quotas in-account only.* Current Flash models (such as 3.6 Flash and 2.5 Flash-Lite) provide **15 RPM**, **1,000,000 TPM**, and **1,500 requests per day** with up to a **2M token context window**, free multimodal ingestion (audio and up to 1h video), structured JSON output, function calling, and a free 2GB temporary **Files API** (48h retention). However, two architectural nuances are essential to understand: 1. **Limits are per Google Cloud Project, not per key.** Creating 5 API keys in the same project still draws from the single shared 1,500 RPD bucket. 2. **The Data Privacy Trade-off.** On the Free tier, prompts and responses **may be reviewed by human reviewers and used by Google to train and improve models**. Attaching billing removes the training clause, but immediately removes the free tier allowance, switching all calls to pay-as-you-go. For clean separation, keep separate free prototyping projects and paid production projects. --- ## Quantitative Quota Tracking (`quota_history.csv`) To quantitatively explore how quotas, rate limits, credit pools, and token buckets have shifted across vendors over time, this repository maintains an append-only time-series dataset in [`quota_history.csv`](quota_history.csv). ### Using the CLI (`quotas.py`) You can inspect, query, and compare historical and active quotas using `quotas.py`: ```bash # 1. View current active quotas across all providers and tiers: python3 quotas.py summary # 2. View full timeline of changes for a specific vendor or tool: python3 quotas.py history --product "Google" python3 quotas.py history --product "Claude" # 3. Filter by metric type (requests, tokens, credits_usd, compute_hours): python3 quotas.py history --metric requests # 4. Compare free-tier vs paid-tier rate limits side-by-side: python3 quotas.py compare # 5. Output structured JSON for graphing or analysis: python3 quotas.py history --product "Google" --json ``` --- ## What this means for the free list - **"Free tier" increasingly means "free completions."** Copilot and Windsurf both give away the cheap, high-frequency feature and meter the agent. When reading a `free_tier` value, check *which* capability is free. - **Shared pools are easy to miss.** Claude's quota spans three products; using one silently drains the others. Free tiers built on the same infrastructure tend to inherit this. - **Dollar-denominated limits are the honest ones.** Where a vendor publishes a dollar figure (OpenCode Go, Copilot's AI Credits), the tier is legible. Where it publishes "credits" without a rate, or nothing at all (Cursor Hobby), it isn't -- and those are the rows most likely to need an `edit` next refresh. - **Boosts are temporary until stated otherwise.** Claude Code's +50% weekly limit has been extended three times and is still, formally, an extension. Don't write a temporary allowance into a `free_tier` or `pricing` value without dating it. --- ## Sources Reviewed September 2026. These change often -- re-verify before relying on a number. - [Claude Code rate limits & usage quotas](https://www.truefoundry.com/blog/claude-code-limits-explained) | [Claude usage limits, dated timeline](https://explainx.ai/blog/claude-usage-limits-2026-timeline-explained) - [The flat-rate AI coding subscription era is ending](https://medium.com/activated-thinker/the-flat-rate-ai-coding-subscription-era-is-ending-what-github-copilot-claude-code-and-cursor-9763e043a63a) - [GitHub Copilot pricing 2026: credits explained](https://questloops.com/blog/github-copilot-pricing-2026-free-pro-pro-and-the-new-ai-credits-system) - [OpenAI restores 5-hour Codex limits for Plus](https://9to5mac.com/2026/08/24/openai-restores-5-hour-codex-and-work-limits-for-chatgpt-plus-users/) | [ChatGPT Plus limits 2026](https://customgpt.ai/chatgpt-plus-limits-2026/) - [Cursor pricing 2026](https://www.nxcode.io/resources/news/cursor-ai-pricing-plans-guide-2026) | [Windsurf pricing 2026](https://www.nocode.mba/articles/windsurf-pricing) - [OpenCode Go docs](https://opencode.ai/go) | [OpenCode Go pricing, limits & models](https://hackup.ai/ai-plans/opencode/) - [Gemini API deprecations](https://ai.google.dev/gemini-api/docs/deprecations) | [Hugging Face Inference Providers pricing](https://huggingface.co/docs/inference-providers/en/pricing) - [Kling credits policy](https://kling.ai/docs/point-policy) | [Amazon Q Developer end-of-support announcement](https://aws.amazon.com/blogs/devops/amazon-q-developer-end-of-support-announcement/)