"Which models can I use for free?" is one of the questions we hear most — from developers prototyping before committing budget, and from teams who simply want their light workloads to cost nothing. So we went and checked, the same way we maintain our Model Library pricing: against each provider's own published pages, this week, not against a months-old listicle. Here are the fifteen best free options as of August 2026 — and, just as important, what each one costs you in the non-dollar sense.

How "Free" Actually Works in 2026

Free LLM access now comes in three flavours, and knowing which one you're using matters more than the model name:

Provider free tiers — the vendor gives you rate-limited access to real production models on your own API key (Google, Groq). Permanently free variants — models the vendor prices at $0 indefinitely, usually smaller "flash" siblings of their paid flagships (Z.ai). Aggregator free endpoints — routers like OpenRouter host community :free variants with strict daily caps and, frequently, data-collection terms. All fifteen picks below are labelled accordingly.

The Free Tier Heavyweights: Google Gemini (1–4)

The single most generous free offering in the industry remains the Gemini API free tier. Verified against Google's pricing page this week, six text models are usable at $0 — our top four picks among them:

1. Gemini 3.6 Flash — Google's newest flash model, free with rate limits. The best free model on the market, full stop.
2. Gemini 3.5 Flash — a step behind 3.6 but battle-tested and comfortably capable for agent work.
3. Gemini 3.5 Flash-Lite — the high-volume workhorse; when you need thousands of small completions, this is the free pick.
4. Gemini 2.5 Flash — the previous generation, still free and still solid (2.5 Flash-Lite and 3.1 Flash-Lite are free too, if you want spares).

The catch: Google states plainly that free-tier content is "used to improve our products." Fine for prototyping and public data; wrong for anything confidential. Paid-tier requests are excluded from training — that line item is what you're actually buying.

Permanently Free: Z.ai's Flash Models (5–6)

5. GLM-4.7-Flash — Z.ai prices this at literally $0 for input, cached input, and output. Not a trial, not a quota — the price is free. It's a genuinely capable text model from the same family as the paid GLM-4.7 flagship.
6. GLM-4.6V-Flash — the same deal for vision: free multimodal inference, which almost nobody else offers at $0.

The catch: flash-class capability, not flagship-class, and Z.ai's rate limits apply. But as ongoing free API models go, these are the least-strings-attached entries on this list.

Fast and Free: Groq's Free Plan (7–8)

7. Llama 3.3 70B Versatile — Groq's Free Plan (documented limits: 30 requests/min, 100K tokens/day) serves Meta's 70B open-weight model on hardware that streams tokens faster than you can read them.
8. Llama 3.1 8B Instant — the small sibling with a far roomier 500K tokens/day cap; ideal for classification, extraction, and other high-frequency small jobs.

The catch: the daily token caps are real ceilings — 100K tokens/day on the 70B is prototyping territory, not production. But for a free tier there's no faster inference anywhere.

The Aggregator Route: OpenRouter's :free Variants (9–14)

OpenRouter lists fourteen :free model endpoints as of this week (we pulled the live catalogue). The six standouts:

9. NVIDIA Nemotron 3 Ultra 550B — a 550-billion-parameter mixture-of-experts flagship with a 1 million token context window, free. The most raw capability available at $0 anywhere.
10. NVIDIA Nemotron 3 Super 120B — the mid-size sibling, 262K context, free.
11. Google Gemma 4 31B — Google's open-weight line; also downloadable to run on your own hardware.
12. Google Gemma 4 26B A4B — the efficient mixture-of-experts Gemma; only ~4B parameters active per token.
13. OpenAI GPT-OSS-20B — OpenAI's open-weight model, free via OpenRouter and small enough to self-host on a single consumer GPU.
14. Cohere North Mini Code — a code-specialised free endpoint with 256K context, useful as a $0 coding assistant backend.

The catch: :free endpoints carry strict daily request caps and often route through providers that may log or train on prompts — read the per-model data policy before sending anything sensitive. Treat them as evaluation endpoints, not infrastructure.

The Wildcard (15)

15. Leanstral (Mistral) — Mistral's code-agent endpoint is currently priced "Free" — explicitly for a limited period while they gather feedback. A time-limited deal rather than a policy, which is exactly why it's last on the list — but while it lasts, it's a frontier-lab code model at $0.

Honourable mention: Cerebras gives every new account $5 in credits across all its models — including the fastest GPT-OSS-120B inference available. One-time trial rather than a free tier, but $5 goes a long way at their token prices.

What Free Costs You: The Comparison That Matters

Every entry above trades one of three things: your data (Gemini free tier, most :free endpoints train on prompts), your throughput (Groq's daily caps, OpenRouter's request limits), or your certainty (Leanstral can start charging whenever the promo ends). None of that makes free access a bad deal — it makes it a deal you should enter knowingly. Prototype on free tiers; move confidential or production traffic onto paid keys where the data-use terms and rate limits are contractual.

Running Free Models Through ApiSpi

Everything on this list slots straight into ApiSpi. Connect your own Gemini, Groq, or Z.ai key as a connector and those models appear in Aria's model picker — and as of this week, you can list every model your key reaches in the connector's new "Additional Models" field and pick between them individually. Set the picker to Auto — cheapest and ApiSpi routes each message to the lowest-cost model you've connected; free-tier models, by definition, win that race. The same governance guardrails apply to a $0 model as to a frontier one — which matters precisely because free tiers are where prompts are most likely to be training data.