Raw research notes

Official standing free LLM APIs

Check date: 2026-11-14. Numbers below come from each vendor’s own pricing, rate-limit, model, or terms page opened that day. If the vendor page does not print a number, the cell says not published. Blog and forum figures are not used when the vendor page is silent or disagrees.

A “light day” below, when the math is shown, is 8 agent tasks × 15 calls × 4,000 tokens = 120 requests and 480,000 tokens. No estimate is given when the vendor does not publish the quota that estimate would be compared to.

Shortlist

Best three official $0 options for a tool-calling agent on a normal laptop:

  1. Google AI Studio / Gemini Developer API, free tier. No billing account is required. Custom function calling is part of the Gemini API, and code execution plus URL context are priced “Free of charge” on the free tier. gemini-4-hydrogen publishes a 1M-token context window. This is the strongest set of free tool-calling models. The cost is data use: free-tier content is used to improve Google’s products, and per-model requests-per-minute / requests-per-day are not printed (they are shown per project in AI Studio).
  2. Groq free tier. OpenAI-compatible at https://api.groq.com/openai/v1. openai/gpt-oss-120b and openai/gpt-oss-20b are production models with documented tool use and a 131,072-token context. A card is required only if you upgrade to Developer. Inputs and outputs are not retained by default. Groq does not publish a table that is labeled as the free-tier quota (see the disagreement note under Groq), so a light-day fit cannot be calculated from the docs.
  3. Z.AI GLM-4.7-Flash. The pricing page lists input, cached input, cached storage, and output as Free, with no “limited-time” label. The model page gives a 200K context window, 128K max output, and function calling. OpenAI-compatible base URL https://api.z.ai/api/paas/v4/. Rate limits are not published (console only), so this is third: the allowance might be tighter than the $0 price suggests, and the public privacy policy does not cover the API.

Why the others lost

Standing $0 tiers

Google AI Studio / Gemini Developer API

Field What the vendor page says
Signup Google AI Studio
Credit card No for the free tier. Free qualifies as “Active project or free trial.” Tier 1 is “Set up and link an active billing account.” Moving to paid means linking billing.
Tool / function calling on the free tier The function-calling guide covers the Gemini API, including Gemini 3 and 2.5. The pricing page does not mark custom function calling as paid-only. Code execution and URL context are “Free of charge.” Computer use is “Not available” on the free tier. The custom-tools endpoint for Gemini 3.1 Pro Preview is not on the free tier.
Free chat models (price “Free of charge”) gemini-4-hydrogen, gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, gemini-3-flash-preview (legacy), Gemma 4 (no paid tier). Not on the free tier: gemini-4-carbon, gemini-4-argon.
Context window 1M tokens is printed for gemini-4-hydrogen and, on the deprecated gemini-2.5-flash section, for that model. The Gemini 3 developer guide says Gemini 3 models support 1M input / 64k output. Context for gemini-3.8-flash, gemini-3.6-flash, gemini-3.5-flash-lite, gemini-3.1-flash-lite, and Gemma 4: not published on the pricing page.
Rate limits Not published as a per-model RPM / TPM / RPD table. Dimensions are requests per minute, input tokens per minute, and requests per day. Limits are per project, not per key. Requests-per-day reset at midnight Pacific time. “Specified rate limits are not guaranteed and actual capacity may vary.” Read the live numbers in AI Studio. A pricing note says “Rate limits are subject to change.”
Stable or “may change” The free tier is a documented tier. The rate-limit page says specified limits are not guaranteed and capacity may vary.
OpenAI-compatible base URL Not confirmed this check. https://ai.google.dev/gemini-api/docs/openai timed out. The native REST host that did load is https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent.
One-line catch Free-tier rows say content is used to improve Google’s products. Paid rows say it is not. Grounding with Google Search is “Not available” on the free tier for the current Gemini 3/4 Flash rows (2.5 Flash still prints 500 grounded requests/day, and that model is deprecated).

Deprecated, do not treat as a model you can still select: gemini-2.5-flash and gemini-2.5-pro are no longer selectable in AI Studio. From 1 Dec 2026 their API calls route to gemini-3.5-flash-lite and gemini-3.8-flash. Their price tables still say “Free of charge” until that redirect.

Light-day estimate: skipped. RPM, RPD, and tokens/day are not published.

Sources: pricing, rate limits (last updated 2026-10-28 UTC on that page), function calling, Gemini 3 guide, API keys.

Groq

Field What the vendor page says
Signup Groq console
Credit card No to stay on Free. A card, US bank account, or SEPA debit is required only to upgrade to Developer. You can downgrade back to Free.
Tool / function calling Tool-use overview: openai/gpt-oss-20b and openai/gpt-oss-120b support local and remote tool use (not parallel). qwen/qwen3.8-27b supports local, remote, and parallel tool use. openai/gpt-oss-safeguard-20b supports local and remote tool use.
Context window openai/gpt-oss-120b and openai/gpt-oss-20b: 131,072 (max completion 65,536). qwen/qwen3.8-27b: 131,072 (max completion 16,384).
Rate limits Free-tier RPM, RPD, TPM, and TPD: not published under a Free label. The rate-limits page prints one table (for openai/gpt-oss-120b, openai/gpt-oss-20b, and qwen/qwen3.8-27b: 30 RPM, 1,000 RPD, 8,000 TPM, 200,000 TPD) and the sentence above that table says those figures are “the base limits for the Developer plan.” The models page labels Developer-plan limits for the same gpt-oss models as 250K TPM and 1K RPM, and for qwen/qwen3.8-27b as 250K TPM and 1K RPM. Those two official pages disagree. Do not treat either row as the free quota. Groq says the exact limits are on the account Limits page.
Stable or “may change” Free is an ongoing plan, not described as a trial. The rate-limits page says its table “is a high level summary and there may be exceptions.” Preview models, including qwen/qwen3.8-27b, “may be discontinued at short notice.” openai/gpt-oss-120b and openai/gpt-oss-20b are listed as production models.
OpenAI-compatible base URL https://api.groq.com/openai/v1
One-line catch Inference inputs and outputs are not retained by default. Groq may keep them up to 30 days to troubleshoot failures or investigate abuse, unless Zero Data Retention is turned on. Data at rest is in the United States.

Light-day estimate: skipped. The free-tier token and request caps are not published as Free.

Llama 3.1 8B and Llama 3.3 70B on the models page are “Contact Sales,” not part of the free table.

Sources: rate limits, models, billing FAQ, tool use, OpenAI compatibility, your data.

OpenRouter free models

Field What the vendor page says
Signup OpenRouter
Credit card Not stated. :free variants do not require a credit purchase. A negative credit balance blocks free models until you add credits.
Tool / function calling On the free-models collection (usage ranking, November 2026), these listings say tool use in the vendor’s own text: Nemotron 3.5 Super VL (free) (“tool use”), Inkling 2 Small (free) (“supports tool calling”), North Mini Code (free) (“tool use via JSON schema”), LFM3-1.7B (free) (“tool calling”), Particle-1 (free) (trained around fewer tool-call tokens). Other $0 listings on that page are not named here, because their blurbs do not say tool or function calling. Free variant IDs end in :free. openrouter/free routes to a free model.
Context window From that same collection page: Nemotron 3.5 Super VL 1M; Inkling 2 Small about 1.05M; North Mini Code 256K; LFM3-1.7B 262K (the card also says 256K in prose); Particle-1 about 1.05M.
Rate limits For :free variants, shared across free models: 20 requests/minute. 50 requests/day if all-time credits purchased are under 10. 1,000 requests/day at 10 or more credits (the higher ceiling starts at 9 credits, to absorb top-up fees). Failed attempts count toward the daily cap. Tokens per day: not published. Other $0 models that are not :free variants skip this cap and can still hit provider limits; some of those still require a funded account.
Stable or “may change” The free-variant doc says free variants may have different rate limits or availability than paid. The collection is a live usage ranking. The 20/50 policy is the current written cap, not labeled as a temporary promo.
OpenAI-compatible base URL https://openrouter.ai/api/v1 (the limits doc calls https://openrouter.ai/api/v1/key; chat uses the same OpenAI-compatible chat-completions API).
One-line catch 50 requests/day is the whole free budget until you buy credits. Liquid’s LFM3-1.7B listing says prompts and outputs may be retained and used to train Liquid models. The Free plan column says data-policy routing is not included.

Light day: 8 × 15 = 120 requests. The no-purchase cap is 50 requests/day. 120 > 50, so a light day does not fit. 50/15 ≈ 3 tasks before the daily cap.

Sources: limits, free models, pricing, free variant.

Cloudflare Workers AI

Field What the vendor page says
Signup Cloudflare dashboard, Workers AI. Pricing
Credit card No for the daily free allotment. Workers Free includes 10,000 neurons/day. A paid Workers plan or prepaid AI Gateway credits are required only above that allotment, and for the named paid-only models.
Tool / function calling The function-calling docs (page updated 21 Apr 2026) say to look for a “function calling” property in the catalog. The only model they name is @hf/nousresearch/hermes-2-pro-mistral-7b. That id is not in the LLM price table updated 12 Nov 2026. No current priced model is named there as tool-capable.
Context window Not published on the pricing or limits pages.
Rate limits 10,000 neurons per day, reset 00:00 UTC. Text generation: 300 requests/minute, unless the model requires Workers Paid. Paid-only models: 20 requests/minute on standard billing, or 50 with prepaid AI Gateway credits. Requests per day and tokens per day: not published as those units. The neuron prices are the token budget.
Stable or “may change” Workers AI is documented as Generally Available. Beta models may have lower rate limits. The free allotment is current pricing, not labeled “may change.”
OpenAI-compatible base URL https://api.cloudflare.com/client/v4/accounts/{account_id}/ai/v1
One-line catch Frontier models on the paid-only list (kimi-k2.6, kimi-k2.7-code, glm-5.2, glm-5.3, glm-5.3-flash, DeepSeek V4 variants, inkling-2-small) are outside the free plan. Training use of prompts: not stated on the pricing or limits pages.

Neuron math for a light day of 480,000 tokens, using published neurons-per-million on @cf/openai/gpt-oss-20b (18,182 per million input, 27,273 per million output). This shows budget only. It does not show that this model does tool calling.

A light day fits inside 10,000 neurons only when enough of those tokens are input. The same bound applies, with different neuron prices, to other models on the free price table.

Sources: pricing (updated 12 Nov 2026), limits (updated 17 Sep 2026), function calling, OpenAI compatibility.

SambaNova Cloud

Field What the vendor page says
Signup SambaNova Cloud plans. API base is documented at API keys and URLs.
Credit card The vendor’s own pages disagree. Rate limits: free tier applies “when there is no payment method linked.” Plans: Free says “Add a payment method and purchase credits to run your first requests.”
Tool / function calling Function-calling doc lists Meta-Llama-3.3-70B-Instruct, gpt-oss-120b, DeepSeek-V3.1, DeepSeek-V3.2, gemma-4-31B-it. All five are also on the free-tier rate table. For gpt-oss-120b, the doc says to set reasoning_effort to high for better tool calls.
Context window Production, from the models page: DeepSeek-V3.1, Meta-Llama-3.3-70B-Instruct, gpt-oss-120b are 128k. Preview: DeepSeek-V3.2 32k; gemma-4-31B-it 128k on the models page. A release note also says Gemma 4 31B text-only supports up to 256k and video up to 128k.
Rate limits Free tier, per model, for every model in the free tables: 20 requests/minute, 20 requests/day, 200,000 tokens/day. Developer tier (payment method linked) is higher and is not free. Preview models “may be removed at short notice.”
Stable or “may change” Production free-tier numbers are written as the current policy. Preview models are evaluation-only and may be removed at short notice.
OpenAI-compatible base URL https://api.sambanova.ai/v1
One-line catch 20 requests/day per model. The Plans page and the rate-limit page do not agree on whether the first request needs a card.

Light day: 120 requests against 20 requests/day. The request cap binds first. At 4,000 tokens each, 20 requests are 80,000 tokens, which is under the 200,000 tokens/day cap. 20/15 ≈ 1 task per model per day.

Sources: rate limits, models, function calling, API URLs, plans.

Z.AI / Zhipu

Field What the vendor page says
Signup Z.AI. Rate-limit console: z.ai/manage-apikey/rate-limits.
Credit card Not stated on the pricing page.
Tool / function calling GLM-4.7-Flash is on the GLM-4.7 model page, which documents function calling (“Function Call”) and interleaved thinking with tools. GLM-4.5-Flash and GLM-4.6V-Flash are priced Free. Their context and tool support were not on a page opened in this check.
Context window GLM-4.7-Flash: 200K context, 128K max output. The other two free models: not published on the pages opened.
Rate limits Not published. The FAQ says to read them in the console linked above. Error code 1302 is “Rate limit reached,” with no public number attached.
Stable or “may change” The pricing page lists GLM-4.7-Flash, GLM-4.5-Flash, and GLM-4.6V-Flash as Free for input, cached input, storage, and output. It does not mark that price “limited-time.” Cached-input storage on paid models is separately labeled “Limited-time Free.”
OpenAI-compatible base URL https://api.z.ai/api/paas/v4/
One-line catch The public privacy policy says it does not cover API Services and points developers to a Data Processing Addendum, which was not opened. Do not assume the consumer “improve our models” wording applies to this API. The Coding Plan is a separate paid subscription and is not this free model.

Light-day estimate: skipped. Requests and tokens per day are not published.

GLM-5.3-Flash is not free. The pricing page lists it at $0.15 / $0.50 per million input/output tokens.

Sources: pricing, GLM-4.7, OpenAI SDK, FAQ, privacy policy.

Cohere trial keys

Field What the vendor page says
Signup A trial key is created when the account is created. Dashboard API keys, pricing FAQ.
Credit card Not stated for a trial key. Production keys go through a Go to Production billing flow.
Tool / function calling The models overview says the Command family powers tool-using agents. Command A (command-a-03-2025) is documented for tool use, 256k context, 8k max output. Also live and on the trial chat table: command-a-plus-05-2026 (128k, agentic), command-r7b-12-2024 (128k, tool use), command-a-reasoning-08-2025 (256k, agent tasks), command-a-vision-07-2025 (128k), command-r-08-2024 and command-r-plus-08-2024 (128k). North Mini Code is on the trial chat table at the same 20 requests/minute. Its own tool-use sentence was not on the pages opened.
Context window See the model rows above. Tokens per day: not published.
Rate limits Trial keys: 1,000 API calls a month, and 20 requests/minute per listed chat model. Embed, rerank, and other endpoints have separate trial caps on the same page.
Stable or “may change” Trial keys are the standing free path. They are limited, and they are not a production key. The page does not say the 1,000-call cap “may change.” A warning says production keys behave like trial keys on newer variants such as Command A Reasoning.
OpenAI-compatible base URL Not an OpenAI-compatible API on the pages opened. Calls go through Cohere’s Chat API and SDK.
One-line catch Trial calls are free and not permitted for production or commercial use.

Light day: 120 calls against 1,000 calls/month. 1,000 / 120 ≈ 8.3 light days per month. The 20 requests/minute cap can hold a sequential 15-call task. It cannot hold 120 calls in one minute.

Sources: rate limits, pricing, models, Command A, tool use quickstart.

Time-bounded or card-gated offers (not standing $0)

These were on the list, or they are large enough to mention. They are not a renewing no-card API.

Cerebras free trial

Cerebras’s own FAQ: “Is there a permanently free tier? No.” New accounts get $5 after a verified payment method. Credits expire 30 days after they are granted. Skip the card and Playground and API access stay inactive. There is no charge until you buy more credits.

Free-trial limits while the $5 lasts:

Model RPM Uncached TPM Total TPM Tokens/hour Tokens/day
gpt-oss-120b 5 30K 90K 1M 1M
qwen-3.8-27b 300 150K 750K (temporary; standard total is 450K) none none

qwen-3.8-27b documents tool calling and parallel tool calling. Its context is 65,536 on the free trial and 131,072 “for paid tiers and trial customers” (the model page prints both). gpt-oss-120b context: not published on the rate-limits page. OpenAI base URL: not confirmed on the pages opened. Pricing on the Qwen page is $0.99 / $1.49 per million input/output tokens, so the $5 is a small prepaid balance, not a $0 rate.

Light-day estimate against the $5 balance: skipped. The $5 is money, not a published request or token allowance that the 120-call formula maps onto without using the paid token price.

Sources: rate limits, Qwen 3.8 27B, pricing.

NVIDIA NIM at build.nvidia.com

Field What the vendor page says
Signup build.nvidia.com. Keys: build.nvidia.com/settings.
Credit card No. The API catalog llms.txt says every model has a free trial tier and no credit card is required.
Tool / function calling Not published as a free-tier model list. Endpoints implement OpenAI Chat Completions. Which of those models accept tools is per model and was not enumerated on the pages opened.
Context window Not published as a single catalog number.
Rate limits Not published. NVIDIA staff on the developer forum say the catalog is a trial, the limit varies by model and by how busy the endpoint is, and the number is shown in the account menu on build.nvidia.com. Forum posts that say “40 RPM” are not a docs-page figure and are not used here.
Stable or “may change” Trial Terms: access is for limited trial purposes, not production, and a pre-release service may be withdrawn. The same terms still describe credits and a limited time. NVIDIA staff (forum, 10 Sep 2025 and 17 Sep 2025) said the credit system was replaced by rate limits and that the trial is not limited by a time period. Treat the hosted API as a trial whose limits are account-specific.
OpenAI-compatible base URL https://integrate.api.nvidia.com/v1
One-line catch Not for production or for serving real end users. Trial Terms section 3.3: NVIDIA may use User Content and Generated Content to improve NVIDIA products and services. Section 2.4 also says NVIDIA will not store that content at the end of each API session, with stated exceptions for uploads and security logs.

Light-day estimate: skipped. No published RPM, RPD, or token cap.

Sources: API catalog llms.txt, Trial Terms (PDF), NIM FAQ (staff answers on rate limits and the meaning of production).

Vercel AI Gateway free credit

Field What the vendor page says
Signup A Vercel team. Getting started.
Credit card Yes. Getting started: add a valid payment method to use free credits. FAQ: 403 customer_verification_required until you do.
Tool / function calling Not published as a model list on the pricing page. Free credit covers a subset of models. The page says to browse “Free Tier models” in the dashboard. Names of those models were not on the page opened.
Context window Not published on the pricing or rate-limit pages. It follows the chosen model.
Rate limits Free tier: a lower limit per model. The number is not published. Paid tier: AI Gateway adds no limit of its own. Tokens per day: not published.
Credit amount Not published on the current pricing page or FAQ. Those pages say a monthly included credit, that it starts at the first AI Gateway request, and that it is not an expiring trial. An older Vercel blog post said $5 every 30 days. That dollar figure is not restated on the current pricing page, so it is not recorded here as the live amount.
Stable or “may change” The FAQ says the free tier is a monthly included credit, not a time-limited trial. Buying AI Gateway Credits moves the team to the paid tier and the monthly free credit stops.
Host https://ai-gateway.vercel.sh is the host named on the pricing page (regional hosts are separate and are not on the free/Hobby path). The /v1 OpenAI path was not printed on the pages opened.
One-line catch A card is required, and the first credit purchase ends the monthly free credit. AI Gateway says it does not log prompts or responses. Zero Data Retention is not available on free credits.

Light-day estimate: skipped. The credit amount and the per-model rate limit are not published.

Sources: pricing, FAQ, rate limits, getting started.

Alibaba Cloud Model Studio (new-user quota, not standing)

Official page, last updated 14 Nov 2026. First activation in the Singapore region, International deployment scope, grants a free quota per model, valid 90 days. The page says each model’s independent quota is typically 1,000,000 tokens, shared by input and output. After it is used or expires, calls bill pay-as-you-go unless Free Quota Only is on, in which case calls stop. A card is not named on that page. “Complete your account information” is required before activation.

This is a new-user grant, not a renewing free tier. Using the page’s “typically 1,000,000 tokens”: 1,000,000 / 480,000 ≈ 2.1 light days per model, then it bills. A separate OAuth path on the same page is 2,000 calls/day. Tool calling on that OAuth path was not checked.

Source: new-user free quota.

Scaleway Generative APIs

The pricing page says: “Try out new models with our free tier: 1 million tokens and 60 minutes of audio transcription.” It does not say that this resets. Tool calling and a card requirement were not on the snippet of that pricing page. 1,000,000 / 480,000 ≈ 2.1 light days if the million tokens are a single pool. Not treated as a standing $0 agent API.

Source: Scaleway model-as-a-service pricing.

Checked and dropped

Mistral

Dropped as a standing free API. Usage and limits: “Free mode includes a monthly usage budget for trying Vibe and Studio. It doesn't include API access. Creating API keys and using the API require billing activation and pay-as-you-go.” The public pricing page free plan is the Vibe chat product (limited messages and coding sessions), not a developer token API. API pricing lists Mistral Moderation 2 as Free, which does not override the billing-activation requirement for API keys, and it is a moderation model rather than a tool-calling agent model. Numeric free-tier RPM/TPM: not published (the limits page says they are shown per model in the admin console).

GitHub Models

Dropped. Retired. GitHub docs: as of 30 Jul 2026, the playground, model catalog, inference API, and bring-your-own-key are no longer available. Changelog.

Hugging Face Inference Providers

Dropped for a free account. Pricing: Free users’ monthly credits are None. Extra usage requires a credits purchase. PRO and Team/Enterprise get $2.00 per month (per seat for Team/Enterprise), which is a paid subscription, not a $0 tier. The inference overview still says “includes a generous free tier.” That sentence disagrees with the pricing table. The pricing table is the one used here. Hub rate limits (requests to the website and file resolvers) are a different quota and are not free inference.

Together

Dropped as $0-to-start. Models: Prism-ML/Ternary-Bonsai-27B is Free input and Free output, context 262,144, function calling Yes. Credits: “Together AI doesn't currently offer free trials. To use the platform, you must purchase at least $15 of credits.” Until that purchase, you cannot use the playground or create API keys. A $0 token price behind a $15 gate is not a standing free API. Default data use, if you do pay to enter: prompts and outputs are stored and may be used for product improvements unless Zero Data Retention is enabled. Model training is off by default.

Rate-limit pages that were not copied from blogs

Gemini per-model RPM/RPD tables on third-party sites, Groq “free tier is 30 RPM / 200K TPD” write-ups, NVIDIA “40 RPM,” Z.AI “about 1 request/second,” and Vercel “$5” on non-pricing pages were not copied. Where a vendor page is silent, this file says not published and links that page.

Sources

All checked 2026-11-14.

Addendum, 14 November 2026

On 14 November 2026 the free Gemini model to select is gemini-4-hydrogen, free of charge on the free tier and $1.50 input / $7.50 output per million on the paid standard tier. gemini-4-argon and gemini-4-carbon are not on the free tier. A previous-generation Flash id still shows a paid standard price of $0.75 / $3.75 per million through 31 December 2026, then $1.50 / $7.50, and that id is not the model to select. GLM-4.7-Flash remains the free Z.AI model, listed as Free, while GLM-5.3-Flash is $0.15 input and $0.50 output per million. Groq's models page does not label a free model, so this note does not treat a priced production model as the free plan. The cheap API id stays deepseek-flash, the version string is DeepSeek-V4.1-Flash-1017, and cache-miss rates are $0.15 / $0.60 off-peak and $0.30 / $1.20 peak per million. The daily-driver Claude model is claude-opus-5-5, Claude Opus 5.5, at $4 input and $20 output per million. The hybrid whole dollars stay about $17 and about $21 for a light month, and about $185 and about $201 for a heavy month. https://ollama.com/library/deepseek-v4 was a 404 on this date, so that library path is not a local file for the API id. Publisher score pages from the same check are local-model cards, not these API prices, and any score on them is the publisher's claim: https://huggingface.co/Qwen/Qwen3.5-4B, https://ollama.com/library/mimo-v2.6, https://huggingface.co/Qwen/Qwen3.8-27B, and https://huggingface.co/Qwen/Qwen3.5-122B-A10B. Local Agent Bench at https://localagentbench.com/results/ is not a paper and is not a free-tier score.