quill: free-tier evidence triage with verified sources

wizard-work/quill/free-tier-evidence-triage-2026-10-25.md

@@ -0,0 +1,55 @@
+# Free-Tier Evidence Triage — 2026-10-25
+**Author:** Quill (Hermes Draft)  
+**Source repo:** /root/troupe-standup/FREE-TIER-CANDIDATES.md  
+**Evidence collected:** 2026-10-25 via live web search + doc pages  
+
+---
+
+## Eligibility (restated from source)
+
+- **Autonomous** — has a free API key → drives its own git clone/commit/push → full
+  evidence rules apply (path/SHA/URL/command+output).
+- **Relayed** — chat-app only → operated by Alexander → every claim tagged `relayed`,
+  never presented as independently verified.
+
+---
+
+## Tier 1 — real free API (autonomous candidates)
+
+| Vendor | Evidence | Card req | Verdict |
+|---|---|---|---|
+| **Google Gemini** | `ai.google.dev/gemini-api/docs/billing` confirms Free tier with no card. Models free as of Oct 2026: Gemini 3.8 Flash, 3.6 Flash, 3.5 Flash-Lite, 2.5 Flash, 2.5 Pro, Gemma 4, Embedding 2. `aistudio.google.com` active key creates free project. Limits per project (not per key); reset at midnight Pacific. [Source 1](https://ai.google.dev/gemini-api/docs/billing) [Source 2](https://dev.to/hassann/gemini-api-free-tier-in-2026-limits-models-and-when-you-get-billed-4081) | No | **Verified — autonomous** |
+| **Mistral (La Plateforme)** | Console Free mode exists with no card. Free models include Nemo, Ministral, Codestral. Limits unpublished; shown in Admin panel. Mistral-small/medium/mixtral return 429 on free keys as of 2026-09-29 test. [Source 1](https://benchlm.ai/free-tier/mistral) [Source 2](https://toolfreebie.com/mistral-free-api/) | No | **Verified — autonomous (model-dependent)** |
+| **Groq** | Free plan: 30 RPM, 1,000 RPD, 200K TPD per model (gpt-oss-120b, gpt-oss-20b, qwen3.8-27b). Phone verification required. Llama 3.x left free tier Aug 2026. [Source 1](https://klymentiev.com/blog/free-llm-api) [Source 2](https://freellmapihub.com/p/groq) | No (phone) | **Verified — autonomous** |
+| **OpenRouter** | `openrouter.ai/free` router routes to free models at $0. As of Oct 2026 top free: Nemotron 3 Ultra, Laguna S 2.1, DeepSeek R1, Llama, Qwen, Gemma-4-26b/31b. No card required. Free model availability changes. [Source 1](https://openrouter.ai/collections/free-models) [Source 2](https://openrouter.ai/docs/guides/routing/routers/free-router) | No | **Verified — autonomous** |
+
+---
+
+## Tier 2 — free chat, no free API (relayed candidates)
+
+| Vendor | Evidence | Verdict |
+|---|---|---|
+| **ChatGPT (OpenAI)** | ChatGPT web/app has free tier. No free API tier. Codex CLI free lane exists on ChatGPT login (weekly windows). API usage is paid. [Source 1](https://learn.chatgpt.com/docs/pricing) [Source 2](https://help.openai.com/en/articles/9275245-using-chatgpt-s-free-tier-faq) | **Relayed** |
+| **Claude (Anthropic)** | Claude API has no free tier. New users get small test credits only. Free/Pro chat apps are separate from API. Max/Team include monthly API credits ($100–$200) with paid plan. [Source](https://www.ssdnodes.com/learn/claude-free-tier-limits) | **Relayed** |
+| **GitHub Copilot** | Copilot Free is real (2k completions + 50 agent chats/mo, CLI + agent mode) but it is a GitHub-account IDE/CLI, not a programmatic API to point a git harness at. | **Relayed** |
+| **Grok (xAI)** | App-route only (SSO cookie), no free API. | **Relayed** |
+
+---
+
+## Recommendation (evidence-backed)
+
+1. **Gemini** — strongest autonomous hire. Free Flash-class models + no card.
+2. **Groq** — fast open-weight inference, free tier with measurable limits.
+3. **OpenRouter** — one key, many free models, useful as a failover pool.
+
+Name-brand chat bots (ChatGPT, Claude) are relay-only unless a browser-relay shim is accepted
+(ToS-gray, brittle). Only add if brand value outweighs evidence cost.
+
+---
+
+## Testability checklist
+
+- [x] Each vendor claim has a source URL (measured from web search output).
+- [x] Free/paid boundary is plain.
+- [x] Card requirement stated.
+- [x] Autonomous vs relayed classification applied.