Best Inference Providers for Scalable Side Projects (2026)
The short answer
A side project needs three things from an inference provider: no bill when nobody is using it, a hard ceiling on spend, and room to grow without a rewrite. Per-token serverless APIs give you all three. As of October 12, 2026:
- OpenRouter — best place to start. One OpenAI-compatible key for every major model and provider, no markup on inference (5.5% fee when you buy credits), automatic fallbacks, and free model variants at 50 requests/day (1,000/day after $10 of credits).
- Cloudflare Workers AI — best free tier for apps on Cloudflare. 10,000 Neurons per day free on every plan, then $0.011 per 1,000 Neurons; gpt-oss-120b costs $0.35/$0.75 per million tokens.
- Groq — best cheap, fast default model. gpt-oss-120b at $0.15 / $0.60 per million tokens (cached input $0.075) at about 500 tokens per second.
- Fireworks AI or Together AI — when you outgrow small models. Kimi K3, GLM-5.3 and DeepSeek V4.1 Flash on the same API, plus a path to reserved capacity.
- Modal — when you must run your own model. Per-second GPU billing, scale to zero, $30 free credits a month on the $0 Starter plan.
The comparison
| Provider | Free start | Cheapest useful model (per MTok, Oct 12, 2026) | Spend control | Scales to |
|---|---|---|---|---|
| OpenRouter | Free model variants, 50 req/day (1,000 with $10+ credits) | Pass-through: same price as the provider | Prepaid credits | Any provider it routes to; BYOK fee-free to $25,000/mo list price |
| Cloudflare Workers AI | 10,000 Neurons/day | gpt-oss-20b $0.20 / $0.30 · gpt-oss-120b $0.35 / $0.75 | Hard daily cap on Workers Free | Workers Paid; frontier models (GLM-5.3, DeepSeek V4) need paid billing |
| Groq | Free tier with per-model limits | gpt-oss-120b $0.15 / $0.60 | Rate limits per org; 429 with retry-after | Developer plan: higher limits, Batch, Flex |
| Cerebras | One-time $5 credit (expires in 30 days) | gpt-oss-120b $0.35 / $0.75 | Access pauses when the credit runs out | Pay as you go; Dedicated Inference with SLAs |
| Fireworks AI | Pay-as-you-go | GLM-5.3 Flash $0.15 / $0.50 · gpt-oss-120b $0.15 / $0.60 | Prepaid balance, auto-reload, monthly spend limit | Priority tier, on-demand GPUs, batch at 50% |
| Together AI | Pay-as-you-go | GLM-5.3-Flash $0.15 / $0.50 · Qwen3.8 Flash $0.15 / $0.47 | Account limits | Provisioned throughput; dedicated H100 $5.49/h |
| Modal | $30/month credits, 10 GPU concurrency | Your own model, per second (~$3.95/h on its pricing example) | Per-second billing, scale to zero | Team plan $250/mo: 50 GPU concurrency |
Picks by stage
Weekend prototype: OpenRouter. Start on a free variant, switch the model string when you need a better one, and never sign another provider’s terms. Because prices pass through, you lose nothing by starting here except the 5.5% credit fee.
Launched, a few hundred users: Groq or Fireworks directly. At this point the credit fee and an extra network hop are worth removing for your main model. gpt-oss-120b at $0.15/$0.60 is the price floor for a capable model; a chat app sending 2,000 input and 500 output tokens per message pays about $0.0006 per message, or $6 per 10,000 messages.
An app on Cloudflare Workers: Workers AI. The model runs next to your code, with no extra keys and a free daily allowance that covers light use. Watch the frontier-model list: GLM-5.3 and DeepSeek V4 require a paid billing method.
Your own fine-tune or an unusual model: Modal. Wrap the model in a Python function, deploy, and pay per second only while requests are running. Cold starts are the trade-off; see best inference vendors for autoscaling and cold starts.
It took off: Together AI or Fireworks AI. Both serve the same frontier open models at near-identical prices and sell reserved capacity when shared serverless stops being good enough. The full comparison is in leading serverless inference providers.
Five rules that keep a side project cheap
- Prepaid credits, never open-ended billing. A leaked key or a retry loop then costs at most your balance.
- Cap
max_tokenson every call and limit requests per user in your own code. - Cache the system prompt. Groq charges half for cached input; Fireworks charges $0.015 for cached gpt-oss-120b input against $0.15 uncached.
- Keep the API OpenAI-compatible so moving providers is a base-URL change.
- Pick the smallest model that passes your own 20-example test, and re-test when a new cheap model ships. Closed-model prices are on our current API prices page.
Hub: best LLM inference providers ranked.
Last verified: October 12, 2026.