Issue 001 · 2026-11-12 · SweetQuota

This week: a frontier-class model is free of charge (not "free trial") from Google, Groq's free plan quietly became the fastest way to run GPT-OSS, and two free tiers you may still be counting on — GitHub Models and Mistral's free API — are gone. Details and receipts below.

Free tier radar

1. Google Gemini API — Gemini 3.8 Flash is free of charge

Google's official pricing page lists Gemini 3.8 Flash (and 3.6 Flash) with input tokens, output tokens, and context caching all marked "Free of charge" on the Free tier. No credit card is required to start, and access runs through Google AI Studio. Two catches, both stated by Google itself: free-tier content "may be used to improve our products," and the flagship Gemini 4 models are not on the free tier at all. Quotas are enforced per project and reset at midnight Pacific; Google no longer publishes one fixed table, so the limits in your AI Studio dashboard are the ones that apply. Who it's for: anyone who needs a capable general model for prototypes, internal tools, or low-volume product features at $0. Start: create a project and API key in AI Studio, then check your per-model limits on its rate-limit page before you build around them. Official: https://ai.google.dev/gemini-api/docs/pricing

2. Groq — the free plan is a speed tier, not a toy tier

Groq's published free-plan limits for openai/gpt-oss-120b are 30 requests/minute, 1,000 requests/day, 8,000 tokens/minute, and 200,000 tokens/day. GPT-OSS 20B and the Qwen 3.8 27B preview carry the same limits. Speech-to-text is included: Whisper Large v3 gets 20 requests/minute and 28,800 audio-seconds/day (8 hours) on the free plan. Limits apply per organization, not per key, so extra keys don't multiply your allowance. Who it's for: latency-sensitive prototypes — voice agents, live transcription, interactive chat — where Groq's inference speed is the point. Start: sign up for the free plan in the Groq console, create a key, and read your exact limits on the console's Limits page (Groq notes the docs table is a summary). Official: https://console.groq.com/docs/rate-limits

3. OpenRouter — 25+ free models behind one key, with a hard daily cap

OpenRouter's pricing page lists 25+ free models. Models with the :free suffix are capped at 20 requests/minute and 50 requests/day; once your account has purchased $10 of credits in total (a one-time threshold, not a subscription), the daily cap rises to 1,000. Failed requests count against the daily cap, and a negative balance blocks free models entirely. The free roster rotates, so treat any specific :free model as temporary. Who it's for: evaluating many open models through one OpenAI-compatible endpoint without opening ten accounts. Start: grab a key, browse the free collection, and pin your code to a specific :free variant rather than the auto-router if you need predictable behavior. Official: https://openrouter.ai/pricing

4. Cloudflare Workers AI — 10,000 free Neurons a day, images and audio included

Cloudflare gives every Workers account a free allocation of 10,000 Neurons per day, resetting at 00:00 UTC, on both the Free and Paid plans. The catalog behind that allowance covers LLMs (Llama, Qwen, GPT-OSS, DeepSeek distills), embeddings, Whisper transcription, and FLUX image generation — the official pricing page prices FLUX.1 Schnell at 4.80 Neurons per 512×512 tile plus 9.60 per step, so the daily free pool covers a useful volume of small-model calls, images, or transcription for a side project. Past the allowance, requests fail until reset unless you move to Workers Paid ($0.011 per 1,000 Neurons beyond the free pool). Who it's for: edge-deployed side projects that need a bit of everything — some text, some images, some audio — on one bill of zero. Start: the allowance is already attached to a free Workers account; call models via the Workers AI binding or REST API and watch Neuron usage in the dashboard. Official: https://developers.cloudflare.com/workers-ai/platform/pricing/

5. GitHub Copilot Free — 2,000 completions a month, no card

GitHub's official Copilot page states that Free users are limited to 2,000 code completions per month, with access to entry-level models, and that no credit card is required. Chat and agent features run on a separate monthly AI Credits allowance that GitHub does not publish a figure for on the Free plan, so treat completions as the dependable part of the offer. Note that GitHub may use individual subscribers' interaction data for model training, with an opt-out in settings. Who it's for: developers who want real in-editor autocomplete without paying for a seat while evaluating. Start: sign in with a GitHub account in VS Code or a supported IDE and enable Copilot Free; check the Copilot settings page for your monthly reset. Official: https://github.com/features/copilot

One deep pick: Gemini 3.8 Flash on the free tier

If you claim one free tier this week, claim a Gemini API key. Google's pricing page lists Gemini 3.8 Flash — its current fast model, aimed at coding and agentic work — at "Free of charge" for input and output tokens on the Free tier, with context caching free as well. That covers the workload most indie projects actually have: support drafting, document Q&A, classification, summarization, and small agents, at zero token cost and with no card on file.

The boundaries are real. Google states that free-tier content may be used to improve its products, so keep client data and secrets off it. Search and Maps grounding are unavailable on the free tier for this model, and the Gemini 4 flagships are paid-only. Quotas are enforced per project, reset at midnight Pacific, and Google says specified limits are "not guaranteed" — your AI Studio dashboard, not a blog post, is the source of truth.

Against the paid tier, you give up the data-use exclusion, higher and more predictable limits, and grounding. The paid Flash rate is $0.75 input / $3.75 output per million tokens through December 31, 2026 — cheap, but not zero. Rational setup: prototype and run low-volume glue on free, and move a workload to paid only when a specific boundary — data policy, grounding, quota — actually bites.

Free model access

Legitimate, no-card routes to real models right now. Availability and limits verified 2026-11-12 against the official pages linked in the appendix.

• GPT-OSS 120B (OpenAI, open weights) — via Groq's free plan: 30 requests/min, 1,000/day, 200,000 tokens/day. The binding limit is the 8,000 tokens/minute cap; design for short, bursty calls.
• Gemini 3.8 Flash — via the Gemini API free tier in AI Studio: input and output priced "Free of charge"; quotas are per project and shown in AI Studio. Main restriction: free-tier content may be used to improve Google's products.
• Rotating :free roster (Gemma, Nemotron, GPT-OSS and others) — via OpenRouter: 20 requests/min and 50/day, rising to 1,000/day after $10 of lifetime credit purchases. Main restriction: the roster changes without notice and failed requests consume quota.
• Llama 3.1 8B Instruct, FLUX.1 Schnell, Whisper — via Cloudflare Workers AI: everything draws on one shared pool of 10,000 free Neurons/day. Main restriction: the shared pool — heavy image or 70B usage drains the same budget as cheap 8B calls.
• 100+ hosted models (DeepSeek, Llama, Nemotron, Kimi and more) — via NVIDIA Build (NIM API catalog): free access for NVIDIA Developer Program members, scoped to prototyping and evaluation. Main restriction: NVIDIA publishes no single fixed quota; limits vary by model and account, and production use requires an AI Enterprise license.

Quick links

• Google Colab — free browser notebooks with GPU access; Google's FAQ caps free sessions at 12 hours and warns resources are not guaranteed. https://research.google.com/colaboratory/faq.html
• OpenRouter free models collection — the live, current list of :free models, so you can check what's actually free today instead of trusting a blog post. https://openrouter.ai/collections/free-models/
• Gemini skills repository — Google's free, official agent skills and plugins for wiring Gemini into coding assistants like Claude Code and Cursor. https://github.com/google-gemini/gemini-skills
• NVIDIA Build model catalog — browse the hosted NIM models you can call free for prototyping with a developer account. https://build.nvidia.com

Enjoying the Wire? Subscribe to get one issue a week — every free tier in it re-verified against the vendor's own pages before it lands in your inbox.

Prices and free tiers are as published on the official pages linked below; vendors change them without notice. Verification date for this issue: 2026-11-12.

Verification appendix (official sources only)

Cut from this issue, and why

• GitHub Models — cut. GitHub's official docs now state the service "has been fully retired" as of July 30, 2026; the playground, catalog, and inference API are unavailable. Any guide still recommending it is out of date. Source: https://docs.github.com/en/github-models/prototyping-with-ai-models
• Mistral free API tier — cut as a free-API recommendation. Mistral's current official docs (Usage and limits) state that Free mode "doesn't include API access" and that API use requires billing activation and pay-as-you-go. The widely cited "Experiment plan" figures come from older, now-removed pages. Source: https://docs.mistral.ai/admin/billing-usage/usage-limits
• Cerebras free tier — cut. No no-card free tier could be confirmed on Cerebras' official pages in this pass; third-party reports describe a move to a card-gated trial. Not listed until an official page confirms otherwise.
• NVIDIA "1,000 free credits" figure — not quoted. The number circulates in third-party blogs but does not appear on an official NVIDIA page we could verify, so the entry above describes the access without it.