Maintained buyer guide
Free AI Inference: Credits, Limits and Catches
Free inference is not one market. Some providers give repeatable daily capacity, some expose a prototype sandbox, and others rotate zero-price models until demand changes. This guide compares what you can actually build before paying.
Free while experimental; advance email notice promised if status changes; no fixed deadline stated
- Useful monthly capacity
- ≈15B input + 150M output tokens per 30 days, if continuously available
- Models / fit
- Qwen/Qwen3.6-35B-A3B-FP8
- Card
- unknown
- After free use
- Requests are rate-limited. Hetzner does not promise continued availability.
HeyGen Voice Turbo: going away November 30, 2026; other listed models: not stated
- Useful monthly capacity
- Request-based; current account cap applies across free models
- Models / fit
- openrouter/free, Models ending in :free
- Card
- no
- After free use
- Wait for the daily reset, purchase credits to change account limits, or use paid models.
not stated; monthly benefit, with rollover and credit-expiry terms not captured
- Useful monthly capacity
- $100/month API credits on $100 Max; $200/month on $200 Ultra, before applicable tax
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
November 30, 2026; higher listed rates start December 1, 2026
- Useful monthly capacity
- 50% lower input/output rates through November 30; implicit caching $0.10/M versus $0.20/M from December 1
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
not stated; catalog availability can rotate
- Useful monthly capacity
- Five hosted models at $0 prompt and $0 completion pricing at collection time
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
not stated; catalog availability can rotate
- Useful monthly capacity
- Three hosted models with $0 prompt and completion pricing at collection time
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
not stated; stealth preview availability may change
- Useful monthly capacity
- $0/M input and output tokens; 1.05M-token context
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
not stated
- Useful monthly capacity
- $0/M input tokens and $0/M output tokens
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
not stated
- Useful monthly capacity
- 51% off Flash 1017: $0.098/M input, $0.0029/M cached input, $0.294/M output
- Models / fit
- Card
- unknown
- After free use
- The provider's current standard pricing or limits apply.
Start with workload shape, not the biggest number
A huge daily token ceiling is useful only if the endpoint supports your model, input type, latency target, and reliability needs. Hetzner's experiment is unusually generous but explicitly unstable. Cloudflare's allowance is durable but measured in Neurons, so model choice changes how far it goes. NVIDIA Build is a prototype environment rather than a promised monthly bucket.
For a small evaluation harness, a rotating free-model router may be enough. For a customer-facing feature, treat every free path as disposable capacity and keep a paid fallback.
Permanent free tier vs promotional access
Repeatable free tier
A documented allowance resets on a known cadence. Cloudflare's daily Neuron allocation is the clearest example in this snapshot.
Experiment or rotating catalog
The provider can remove a model, throttle access, or end the experiment without a published date. Hetzner and rotating :free catalogs belong here.
The privacy catch is part of the price
A zero-dollar request can still carry a data trade-off. Google marks free-tier usage differently from paid-tier usage on its pricing table. Router services may select providers with different retention rules. Never send secrets, production customer data, or regulated records until you have checked the current provider and model terms.
- • Prefer synthetic evaluation data for first tests.
- • Pin a provider when router privacy policies differ.
- • Re-check terms before moving from prototype to production.
A practical selection order
- 1. Define the minimum model and modality.
Text-only extraction, vision, coding agents, and embeddings need different catalogs.
- 2. Convert limits into your own requests.
Estimate input, output, retries, and concurrency—not just headline tokens.
- 3. Check card, region, and data terms.
These can disqualify an otherwise attractive tier.
- 4. Price the fallback before launch.
Know what happens after the allowance and how quickly you can switch.
See only the offers that are current now
The guide explains the trade-offs. The live Deals page handles expiry, verification windows, and claim links.
Open verified inference deals