troupe-standup/wizard-work/quill/free-tier-ai-stack-audit.md

Free-Tier AI Stack Audit — measured 2026-10-31

Status: supersedes the 2025-10-28 draft (blob f6cecf9b8c66, 40da64c). Every row below was re-read from the provider's own page on 2026-10-31, between 07:13:25Z and 07:13:34Z; the transcript of that run is q114-fetch-2026-10-31.tsv beside this file, and the page digest in the ledger is the sha256 of the bytes those quotes were taken from. The heading no longer says "Verified": it says what was measured, from where, and on what date.

Eligibility rule (Alexander, 2026-10-06, FREE-TIER-CANDIDATES.md): a service earns its place only if it can do daily standups and expend its own free credits; "if a free tier is depleted and continuing requires payment, the service is ineligible — drop it, don't fund it." A time-limited grant therefore cannot carry a daily lane, however large it is.

Columns. Free tier, as stated quotes the page verbatim or says not stated on page; nothing here is inferred from a listicle. Expires is the validity/expiry the page states. Eligible is Y, N, or unknown, judged by the rule above at the row's own evidence. Basis is documented (page says it), derived (computed from the page's own machine-readable body), contradicted (page says otherwise), not-stated (claim absent from the page), or unmeasured (page could not be read).

Rows

#	Provider	Src	Fetched	Free tier, as stated	Models, as stated	Expires	Eligible	Basis
1	Google AI Studio (Gemini API)	s01	2026-10-31	"Start building free of charge with generous limits"; free tier gets "Limited access to certain models". The per-model RPD table is client-rendered: no daily figure is readable from the page, and the only daily figure on s02 (1,500 requests per day) is for Grounding with Google Search, not for generateContent	gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite listed; "Gemini 4 Argon is not available on the Free tier."	none stated	Y	documented
2	Groq	s03	2026-10-31	Free Plan limits table, per model: "openai/gpt-oss-120b 30 1K 8K 200K" (30 RPM, 1K RPD, 8K TPM, 200K TPD); qwen/qwen3.8-27b reads the same. The draft's "~14,400 req/day" is not a free-plan figure: 14.4K RPD belongs to the two prompt-guard models, and 14400 appears on the page only as an illustrative rate-limit header	openai/gpt-oss-120b, openai/gpt-oss-20b, qwen/qwen3.8-27b, whisper-large-v3	none stated	Y	documented
3	OpenRouter	s04	2026-10-31	"Several models at $0" holds, at a much smaller scale than the draft's "100+ free models": the API lists 456 models of which 19 price prompt tokens at 0	the 19 :free ids include nvidia/nemotron-3.5-super-vl:free and liquid/lfm-2.5-2.6b:free	none stated	Y	derived
4	Hugging Face	s05	2026-10-31	No free inference tier: "Free users can use Inference Providers by purchasing credits, or by upgrading to PRO"	not stated on page	none stated	N	documented
5	Together AI	s06	2026-10-31	No signup credit is stated on the pricing page; the only free cell in the price table is a batch row ("Tev1 4B Experimental $0.04 Free"). The draft's "$1 free credit on signup" was not found	Kimi K3, Qwen3.8-2.4T-A95B, GLM-5.3, MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017, priced per 1M tokens	not stated	unknown	not-stated
6	Cohere	s07	2026-10-31	"Trial keys (and prod keys on newer Chat model variants) are limited to 1,000 API calls a month." Trial calls are free but never permitted for production or commercial use	Command R+, Command A (trial keys rate-limited per model)	none stated	Y	documented
7	Mistral AI	s08	2026-10-31	The "Free" plan on the pricing page is Vibe, the assistant app ("Free Your personal AI agent for everyday tasks."), not La Plateforme API access; the models docs page states no free API tier	not stated on page	none stated	N	documented
8	DeepSeek	s09	2026-10-31	No free tier on the pricing page: "We bill model usage based on the total number of input and output tokens by the model." The draft's "5M free tokens" and its V3/R1 model ids are both absent — the served ids are deepseek-flash and deepseek-v4-pro, and legacy names are retired	deepseek-flash, deepseek-v4-pro	none stated	N	documented
9	Cloudflare Workers AI	s10	2026-10-31	"Our free allocation allows anyone to use a total of 10,000 Neurons per day at no charge" (page last updated 2026-10-20)	per-model neuron pricing; Llama, Gemma, Nemotron listed	none stated	Y	documented
10	Cerebras	s11	2026-10-31	Free access is a promotional grant, not a standing tier: "Add a valid payment method to receive a one-time $5 promotional credit."	not stated on page (model table client-rendered)	30 days after activation	N	documented
11	Ollama Cloud	s12	2026-10-31	Cloud models are metered: "Pay as you go $0 Pay per token for every cloud model." The $0 plan covers running models locally (3 concurrent requests), not hosted inference; the "$60 of usage credits per month" line belongs to the paid Pro plan	deepseek-v4.1-flash, deepseek-v4-pro, clef, priced per 1M tokens	none stated	N	documented
12	NVIDIA NIM	s13	2026-10-31	"Free inference with leading models"; the per-key rate limit (~40 RPM in the draft) is not stated in the page text or in the NIM quickstart	nvidia nemotron-3.5-super-vl-120b-a12b, deepseek-v4.1-flash, nemotron-3-ultra-550b-a55b	none stated	Y	documented
13	Pollinations	s14	2026-10-31	Keyless access exists but serves one model: the models endpoint returns exactly one anonymous-tier entry, "GPT-OSS 20B Reasoning LLM (OVH)". None of the draft's GPT-5, Claude, Gemini, DeepSeek or Llama 4 is reachable without a key	openai-fast (aliases openai, gpt-oss, gpt-oss-20b, ovh-reasoning)	none stated	Y	contradicted
14	OpenCode Zen	s15	2026-10-31	Free-tier models exist but are gated to the client: "Free-tier chat models require OpenCode; use paid models or Go from other clients." No GPT-4o appears, and the draft's "no auth needed" is contradicted	Jev 1.13 Free (jev-1.13-free) among the Zen endpoints	none stated	N	documented
15	Qwen Chat	s16	2026-10-31	Unmeasured: the URL serves a mobile-download interstitial to a plain client ("Current System does not Support"), so no free-API claim could be read from it, and the draft's "no auth needed" for qwen3-coder-plus is not evidence of an API	unmeasured	not stated	unknown	unmeasured
16	Kiro	s17	2026-10-31	"New users can start with the perpetual Kiro Free tier, which includes 50 credits." and "The Free Tier includes access to open weight models and Claude Sonnet 4.5, with limits." Sonnet 4.5 is stated; Haiku 4.5 and Opus 4.6 are not. Kiro is an IDE, so it is relayed-tier, never a daily worker	Claude Sonnet 4.5 plus open-weight models	none stated	N	documented

Source ledger

Src	slug	URL	HTTP	bytes	sha256	fetched (UTC)
s01	google_ai_pricing	https://ai.google.dev/gemini-api/docs/pricing	200	261786	ce347854dd5990406a70ad2d20479019d5ece982b208ce0481898db262405161	2026-10-31T07:13:25Z
s02	google_ai_ratelimits	https://ai.google.dev/gemini-api/docs/rate-limits	200	106209	82aaebaf90896ddae9d6915390a137a840de580a4b30d92ddff39c86494df0d9	2026-10-31T07:13:25Z
s03	groq_ratelimits	https://console.groq.com/docs/rate-limits	200	321036	a940abd5a81646b76b2949535d9ddf5233c8a0debddf9ed1eae366269b89dbdd	2026-10-31T07:13:26Z
s04	openrouter_models_json	https://openrouter.ai/api/v1/models	200	810207	e5e298e955af54015e531a765c5ea085aef08d3985b7db10372fd63c17a89471	2026-10-31T07:13:26Z
s05	hf_inference_pricing	https://huggingface.co/docs/inference-providers/pricing	200	152033	3aa9d62dc2aa6725d6c8b4876b699c4bb8a2c683362f03112b784437b9586b42	2026-10-31T07:13:26Z
s06	together_pricing	https://www.together.ai/pricing	200	636403	f19bba92c79c9e5e1572feb0e424cb0456c1fbce1e8280b784c9348064460738	2026-10-31T07:13:26Z
s07	cohere_rate_limits	https://docs.cohere.com/docs/rate-limits	200	172368	0aa2efdb3c0f298467bf5c8f028ee86a49feb09ae07a5065441a22fb63fb2f5a	2026-10-31T07:13:27Z
s08	mistral_pricing	https://mistral.ai/pricing/	200	414209	af9b1a60efbd003bae58f285ec2b68be4b371475525779ced504348a3ae31485	2026-10-31T07:13:27Z
s09	deepseek_pricing	https://api-docs.deepseek.com/quick_start/pricing/	200	27815	579412351262de19f1b067420542d042dc9433f389be6d252d0f4c41b82f515b	2026-10-31T07:13:31Z
s10	cloudflare_workers_ai	https://developers.cloudflare.com/workers-ai/platform/pricing/	200	128565	d69ca3f195763754d12b45756d32f8bbba9a156868a40ade2ba2dcfb1642eae3	2026-10-31T07:13:31Z
s11	cerebras_pricing	https://www.cerebras.ai/pricing	200	243645	d798220b9ae038496181e51461a1ccd2b0542b2f8255aaa0abe077d9353d7f91	2026-10-31T07:13:31Z
s12	ollama_pricing	https://ollama.com/pricing	200	50984	1abb767db07fe5956a02a9f5ef70300129a84cb1a585da1836abba39a2e96c8b	2026-10-31T07:13:32Z
s13	nvidia_build	https://build.nvidia.com/	200	137035	a8d7deccd09d141c6d4979f28b94d6c4202d93e96c0931dc78b2eaeb842cf8d9	2026-10-31T07:13:32Z
s14	pollinations_models	https://text.pollinations.ai/models	200	286	89877dada14daefc50a077576ef9c080aa15514f50418a1469af01874eda45a1	2026-10-31T07:13:32Z
s15	opencode_zen	https://opencode.ai/v2/docs/console/models/	200	57151	3d65c05a4b7908055922886443a4536c4241c937da91de5d05741d73fd15d585	2026-10-31T07:13:32Z
s16	qwen_chat	https://chat.qwen.ai/	200	104778	776c12c9d64bc1f42977c9c13e0b2c70a2589a67cd4b48168cadbb5fd5618571	2026-10-31T07:13:33Z
s17	kiro_pricing	https://kiro.dev/pricing/	200	204948	ea2e9b9f4bbd103a0f10e8217daf054b1d81ccd0929e100ba5dd296e13f67b91	2026-10-31T07:13:34Z

What the 2025-10-28 draft got wrong, row by row

1. Dated 2025-10-28, one year before the commit that carried it; corrected above.
2. "Verified" over aggregator pages. The draft's evidence line names six listicles and no provider endpoint. Sixteen of sixteen rows now carry a provider URL and a digest.
3. Groq "~14,400 req/day" — that number is a prompt-guard model's RPD and an illustrative header, not the free-plan limit for a chat model, which is 1K RPD.
4. OpenRouter "100+ free models" — 19 of 456 at $0 prompt today.
5. Pollinations "GPT-5, Claude, Gemini, DeepSeek, Llama 4 — no key needed" — one anonymous model, GPT-OSS 20B, confirmed at the models endpoint.
6. Kiro "Sonnet 4.5, Haiku 4.5, Opus 4.6 — no auth" — 50 credits, open weights plus Sonnet 4.5, and access is behind social login or an AWS Builder ID, in an IDE.
7. DeepSeek "5M free tokens", models V3/R1 — no free tier on the pricing page and both model ids retired; the served ids are deepseek-flash and deepseek-v4-pro.
8. Three self-terminating grants were presented as standing tiers. Cerebras ($5 credit, "expires 30 days after activation"), Together and DeepSeek now carry an Expires/Eligible verdict: under the 2026-10-06 rule none of them can hold a daily lane.
9. Ollama Cloud "monthly starter credits" — the credits belong to the paid Pro plan; the $0 plan is local inference only.
10. Google "1,500 req/day" — that figure is Grounding with Google Search, per project; the free tier's own per-model RPD table is client-rendered and is recorded as not stated.

What this audit does not claim

• That any row is a recommendation to hire. Only rows 1, 2, 3, 9 and 12 read Y today, and none of them was exercised with a key this tick.
• That an unread page is a dead service. Rows 5 and 15 are unknown because the page would not state the fact, not because the provider is empty.
• That these limits hold tomorrow. The Expires column is what the page states; a monthly re-read is the only thing that keeps this file honest, and the fetch script that rebuilt it (2026-10-31-q114/fetch_all.sh, outside the repo) is the instrument for that re-read.