Why Skip the Aggregator Layer
Aggregator gateways (OpenRouter, Cloudflare AI Gateway and the like) are valuable because one key unlocks many models. But their free tiers have been written about endlessly, and the rules keep tightening. What is still genuinely open is the layer underneath the aggregators: the underlying inference platforms. They are competing for developers too, so each of them hands new accounts a trial credit or keeps some open-source models on a free tier.
The idea is simple: skip the aggregator, register directly with the underlying platforms, collect their free credits, then put a self-hosted gateway in front of them so you get one unified entry point while still harvesting each platform's new-user perks.
Steps
Step 1: Decide which inference platforms to register with
Platforms that in 2026 still offer new-user credits and are reasonably accessible include:
- Together AI: new accounts typically receive a trial credit, usually in the single-digit to low-double-digit USD range, varying with promotions. Several open-source models (Llama, Qwen, DeepSeek families and others) are billed per token.
- Fireworks AI: also offers new-user trial credits, focused on low-latency inference with function calling and structured output support.
- Groq: known for very fast inference, offers a free tier with rate limits (requests per minute, tokens per day) on selected models, good for prototypes and light usage.
- DeepInfra: per-token billing at low prices, with small signup credits from time to time.
- Novita AI: developer-oriented free credits plus low-cost inference across multiple open-source models.
- Cerebras Inference: markets ultra-fast inference and offers a free tier, useful for latency-sensitive work.
> Note: free-tier policies, credit amounts and available models change over time and by region. Always check each platform's current pricing and promotions page before registering. No specific amounts are promised here.
Step 2: Register with an email you can maintain long-term
Do not use disposable email addresses. Most platforms verify email, and some require secondary verification during risk checks. Recommendations:
- Use an address you can access long-term (a custom-domain mailbox or a mainstream provider both work).
- Describe your use case honestly - "personal development and testing" is fine. Do not fabricate company details.
- Some platforms require a phone number or credit card to unlock full credits. Avoid unlocking if you can - get the free tier working first.
- Do not register multiple accounts on the same platform to farm credits repeatedly. If risk control catches it, your main account may be banned along with the rest.
Step 3: Get your API key and confirm your quota
Each platform's console usually has:
- API Keys page: generate a key. It is usually shown only once - copy it into a password manager immediately.
- Usage / Billing page: remaining credit, tokens used, credit expiry.
- Rate Limits page: free-tier RPM (requests per minute) and TPM (tokens per minute).
Record each platform's remaining credit and expiry date in a single table. That table drives your routing decisions later.
Step 4: Aggregate multiple platforms behind one self-hosted gateway
This is the core of the article. Deploy an open-source LLM gateway (projects such as One API / New API that support multiple upstreams) and put each platform's key into it, exposing a single OpenAI-compatible endpoint externally.
Rough flow:
- Prepare a server that can reach these platforms' APIs (an overseas VPS is easiest; domestic machines need to solve network access on their own).
- Deploy the gateway, configure the database and admin account.
- Add each upstream as a "channel": platform name, base URL, API key, supported model list.
- Set routing policy: prefer channels that still have free credit, and fail over automatically when a channel is exhausted.
- Generate your own downstream key and hand it to clients (chat apps and IDE plugins that support a custom base URL).
The payoff: one external endpoint and one key, with automatic rotation across multiple platforms' free credits internally, so running out on any single platform does not interrupt your calls.
Step 5: Monitor quotas so you are not cut off unexpectedly
The biggest risk with free credits is thinking you still have some when you do not. Suggestions:
- Enable usage statistics in the gateway and watch daily consumption per channel.
- Set an alert threshold like "notify me when remaining credit drops below X" for each channel.
- Manually log into each platform's console periodically (say weekly) to cross-check, since gateway stats and platform billing can diverge.
- Spend credits that are about to expire first so they are not wasted.
Caveats
- Free credits expire. Most new-user credits are valid only for a limited window after registration and do not auto-renew.
- Free tiers are rate-limited. RPM and TPM are often far lower than paid tiers, so batch jobs get throttled. Build retries and backoff into the gateway side.
- Do not run production workloads on free credits. There is no SLA; platforms may change policy, retire models or tighten limits at any time.
- Compliance. Some platforms restrict free credits to personal, non-commercial use. Read the terms before commercial use.
- Network reachability. Some platforms are restricted in certain regions. Confirm your environment can actually call them before registering.
- Do not share keys. Publishing your downstream key can let others burn your credits or trigger platform risk controls.
When This Fits
- Personal development and prototyping: you want to try several open-source models quickly without paying up front.
- Learning and research: running RAG, agent and function-calling flows to validate an idea.
- Small internal tools: scripts or assistants with low volume and low stability requirements.
- Cost-sensitive model comparison: comparing models side by side in one codebase without topping up each vendor separately.
If you need stable, commercial, SLA-backed usage, free credits are only a bridge - eventually you will move to paid plans or local deployment.