Whenever the industry starts preaching efficiency again, the same worry resurfaces in developer communities: is this the round where free tiers finally disappear? As of November 9, 2026, we walked through the official pages of five hosted inference services one by one: Google (Gemini API Free tier), Groq (Free Plan), Cerebras (Free Trial tier), Mistral (La Plateforme Free plan), and LG (EXAONE API Platform). The headline finding is simple: all five free tiers are still live. But "still live" and "put down in writing" are two different things. One vendor prints exact rate numbers into its documentation; another publishes a rate-limit page with a policy but no numbers; a third's data-policy page simply fails to load. The price of free sits in the fine print, and this piece lines those rows up side by side in a mechanism comparison with zero benchmarks: what models each free tier hands out, how hard each one throttles you, what it takes from your data, and where the official record goes quiet.
A note on positioning. How to read paid unit prices is the subject of llm-api-pricing-comparison-review; how to parse open-weight license clauses belongs to openweight-license-comparison-review. This article looks only at the hosted-service layer: free-tier terms. Within the same coverage cycle, qoduck-anonymous-model-hotspot records the anonymous model launch itself, kinference-vllm-routing-opensource-resource dissects a cluster-side inference routing control plane, and yuanqi-shutdown-migration-sop is an exit manual after a platform shutdown. Four angles, one story.
All Five Still Say Free, but the Wording Sits at Different Levels
The level of the official wording differs, and that decides whether you can treat a tier as a standing plan or a trial window.
Google's wording sits on the official pricing page, verbatim: "Free - For developers and small projects getting started with the Gemini API", followed by limited access to certain models, free input and output tokens, and AI Studio access - the phrasing of a standing tier. Groq's rate-limits documentation carries a dedicated "Free Plan Limits" section listing per-model quotas, with the upgrade path leading to the Developer plan.
Cerebras deserves a separate reminder: the official name is "Free Trial tier", not a generic free tier. The verbatim line reads "Every model in the Shared Inference Model Catalog is available on the Free Trial tier, subject to the per-model Free Trial rate limits." The word "Trial" sits in the tier name itself, and should not be simplified away when quoting. Mistral's official announcement states verbatim that the free tier enables "developers to get started with experimentation, evaluation, and prototyping at no cost". The tier is now called the Free plan; the older "Experiment plan" label no longer appears in the current help center, so this article uses the current name.
LG's free wording is noticeably the thinnest. What docs.exaone.ai shows first-hand is the platform and its endpoints: the EXAONE API Platform, the API key entry point, and two endpoints, K-EXAONE 2.5 and EXAONE Tabular. The "currently free of charge" phrasing, however, reaches us second-hand: multiple outlets on November 9, 2026 quoted it from the official changelog, while the changelog subpage itself returns a 404, so the verbatim wording could not be verified on the primary page. That distance between a quotation and a verbatim record is why this comparison exists.
What You Get: Hosted Open Weights Are the Common Denominator
The five model lists differ sharply in structure.
Google's list is written into the price table: gemini-3.8-flash, 3.6-flash, 3.5-flash-lite, 3.1-flash-lite, 2.5-flash-lite, Gemma 4 and the Embedding models are all marked "Free of charge"; the same table states verbatim that "Gemini 4 Argon is not available on the Free tier", and the Omni video-generation series and Nano Banana image series are likewise marked unavailable on the free layer.
Groq and Cerebras take the same road: their free layers run hosted open-weight models. Groq's free tier lists gpt-oss-120b - OpenAI's open-weight model served as hosted inference on Groq hardware, not a Groq-built model - plus gpt-oss-20b, qwen3.8-27b and whisper-large-v3 on the speech side. Cerebras states that "every model" in its Shared Inference Model Catalog is available on the Free Trial tier, including gpt-oss-120b and qwen-3.8-27b, again as hosted open weights.
Mistral's official page gives the platform's positioning rather than a free-tier model list; which models the free layer includes is not enumerated, and this article does not fill that gap on its behalf. LG's K-EXAONE 2.5 is introduced as LG AI Research's flagship agentic model with a deep understanding of Korean, and EXAONE Tabular targets structured data; in both cases the weights are not open.
Rates and Quotas: Some Publish Numbers, Some Do Not
All rate figures in this section were captured from official pages on November 9, 2026.
Groq is the most granular: on the free layer, gpt-oss-120b is capped at 30 RPM / 1K RPD / 8K TPM / 200K TPD, with gpt-oss-20b and qwen3.8-27b at the same levels; whisper-large-v3 sits at RPM 20 / RPD 2K / ASH 7.2K / ASD 28.8K. The page says this is a "high level summary"; higher allowances require the Developer plan, which also unlocks Batch and Flex processing.
Cerebras's numbers verify verbatim: both gpt-oss-120b and qwen-3.8-27b are listed at RPM 5 / Uncached TPM 30K / Total TPM 90K / TPH 1M / TPD 1M (captured November 9, 2026), with image input limited to 2 per request and a 10 MiB payload. One more verbatim line: "New accounts receive $5 in free credits after adding a verified payment method. These credits expire 30 days after they're granted." Both conditions - a verified card, and a 30-day expiry - are printed on the official page.
Google runs the other way. Its documentation splits rate limiting into three dimensions - RPM, TPM (input) and RPD - and states that limits are applied per project rather than per API key, with RPD quotas resetting at midnight Pacific time. But the free layer's actual numbers have no static official table; the documentation points you to AI Studio, so this article prints no rate numbers for Google's free tier. Mistral is similar: no concrete token quota for the free plan appears on the official page, so capacity planning for both vendors can only rely on what the console shows in real time.
LG, per the second-hand accounts, publishes a rate-limit page with only policy and no numbers; the single-request cap of 524,288 tokens (input plus max_tokens, reasoning included, with a 400 response beyond), Korean- and English-only text support, and the closed weights are all quoted from LG's documentation by third parties. When you schedule production traffic against a free tier, confirming "no numbers published" versus "numbers verifiable" is a different order of work.
Training on Your Data: Only Two Vendors Put It in Writing
The most consequential line of fine print in any free tier is what the vendor takes from your data. Among the five, only two have a verifiable verbatim clause.
Google writes it plainly: the free tier carries "Content used to improve our products", and the per-model price table marks "Used to improve our products: Yes", with the paid tier marked No.
Mistral's clause is equally complete: on the Free plan "we can use your data (input and output) to train our models", with the right to "opt out of this program at any time"; the API-side opt-out runs through the Privacy menu in the Admin panel, where the Anonymous improvement data toggle can be switched off. Paid plans such as Scale state that "data are not used to train our models". Default to training, opt-out available, paid tiers exempt.
The remaining three vendors have no verifiable verbatim clause. Groq's data-policy and privacy pages (console.groq.com/docs/data-policy and /docs/privacy) both returned 404 on November 9, 2026, and no data-policy text could be found elsewhere on the official domain. The Cerebras pages checked make no mention of data usage, and LG's pages carry none at all. One clarification matters: "not mentioned" or "not verified" is not the same as "the vendor announced it does not train". The former is an honest record of what verification found; the latter invents a clause on the vendor's behalf. If your team runs sensitive data through a free tier, the gap itself deserves a line in your risk assessment.
Upgrade Paths, and What Official Pages Leave Out
The upgrade paths differ in shape. Google's paid tier carries "Content not used to improve our products" as a stated term. Above Groq's free layer sits the Developer plan with higher limits plus Batch and Flex processing. Cerebras grants $5 in credits after card verification, expiring after 30 days. Mistral states verbatim that users can evolve their endpoints into a commercial tier and gain "full data isolation" (with a free zero-retention option) and higher rate limits. LG's future pricing and the endpoint of its free period are both unpublished.
Lining up the "not said" column across all five: Google publishes no static rate numbers for the free tier; Groq's data-clause page could not be verified and no official statement covers the free layer's duration; Cerebras does not mention data usage on the pages checked; Mistral publishes no free token quota; LG's free-period end date, future pricing and data terms all lack verbatim official wording. These gaps do not prove any vendor is hiding unfavorable terms - as of November 9, 2026, who has written what in plain sight is, at last, visible. Terms move, and every statement here defers to each official page on the day you read it.
The Comparison Table
The table below summarizes the five vendors under uniform fields; "not verified" means no verbatim official clause was obtained during this check:
| Vendor | Official tier name | Models on the free layer | Rates and quotas | Training-on-data clause | What the official page leaves out |
|---|---|---|---|---|---|
| Free tier | 3.8 / 3.6 Flash series, Flash-Lite series, Gemma 4, Embedding; Argon, Omni and Nano Banana unavailable | RPM / TPM / RPD limits applied per project; no static numbers, view in AI Studio | Content used to improve our products (verbatim Yes) | Free-tier rate numbers | |
| Groq | Free Plan | Hosted open weights: gpt-oss-120b, gpt-oss-20b, qwen3.8-27b, whisper-large-v3 | gpt-oss-120b: 30 RPM / 1K RPD / 8K TPM / 200K TPD; higher needs Developer plan | Data-policy page 404, not verified | Data usage, free-layer duration |
| Cerebras | Free Trial tier | All models in the Shared Inference Catalog, incl. gpt-oss-120b, qwen-3.8-27b | Both at 5 RPM / 90K TPM / 1M TPD; $5 credits for new accounts (card required, 30-day expiry) | Not mentioned on pages checked | Data usage |
| Mistral | Free plan | Platform positioning given, no free-tier-specific list | Concrete token quota not published | Training by default, opt-out available; Scale and other paid plans exempt | Free quota figures |
| LG | EXAONE API Platform (free at launch) | K-EXAONE 2.5, EXAONE Tabular | Rate-limit page has policy only, no numbers (quoted); 524,288-token single-request cap (quoted) | No data clause verifiable | Free-period end, future pricing, data terms |
FAQ
Q1: Are these five free tiers permanent? A1: No vendor writes the word "permanent". Google, Groq, Cerebras and Mistral currently present their free tiers as standing plans, but all reserve room to adjust. LG's "currently free of charge" is a quotation from the official changelog, with no official answer on the free period's end or future pricing. As of November 9, 2026, all five free tiers are still live; defer to each official page on the day you use it.
Q2: Will my free-tier data be used for training? A2: Only two vendors have verbatim clauses: Google states free-tier content is used to improve its products (Yes), with the paid tier exempt; Mistral states the Free plan uses input and output for training by default, with an opt-out available at any time, and paid plans such as Scale are exempt. Groq's data-policy page could not be verified, Cerebras's checked pages make no mention, and LG has no checkable clause - "not mentioned" is not "no training".
Q3: Are free-tier rate limits applied per API key or per account? A3: Google states verbatim that limits are applied per project, not per API key, with RPD quotas resetting at midnight Pacific time. Groq notes that exact limits live on your organization's limits page. Each vendor meters on its own dimensions, so check the official documentation before integrating.
Q4: Do all five publish complete rate numbers? A4: No. Groq and Cerebras publish verbatim numbers (captured November 9, 2026: Groq's gpt-oss-120b at 30 RPM / 1K RPD / 8K TPM / 200K TPD; Cerebras at 5 RPM / 90K TPM / 1M TPD). Google provides no static numbers and points to AI Studio; Mistral publishes no free token quota; LG's rate-limit page carries policy only.
Q5: What are the upgrade paths to paid tiers? A5: Google's Paid tier stops content being used for training; Groq's Developer plan unlocks higher limits plus Batch and Flex processing; Cerebras grants $5 in credits after card verification, expiring in 30 days; Mistral lets endpoints evolve into a commercial tier with full data isolation and higher rate limits; LG's future pricing is unpublished, so its upgrade path has no official wording yet.
Related reading
- llm-api-pricing-comparison-review: the paid-unit-price comparison - why a cheaper list price does not guarantee a cheaper bill.
- openweight-license-comparison-review: the open-weight license comparison, a complement to this hosted-terms review.
- qoduck-anonymous-model-hotspot: this cycle's hotspot piece on the anonymous Qoduck launch.
- kinference-vllm-routing-opensource-resource: this cycle's open-source piece on inference routing for Kubernetes and vLLM.
- yuanqi-shutdown-migration-sop: this cycle's SOP on migrating off Tencent Yuanqi.
- gemini-omni-12-flash-api-setup-sop: a Gemini API setup guide that pairs with the Google section of this review.