Proprietary APIs bill you whether your app is sleeping or shipping. Self-hosting takes weeks to set up. Token Factory sits between the two. Here is who it is built for.
Token Factory is for any developer or solo builder who needs open-weight LLM inference without a GPU bill, a vLLM config, or an ops team. Change one URL. Keep all your existing code. Pay per token, not per hour of idle compute.
Key takeaways
base_url and one api_key and your existing SDK calls work against open-weight models like Llama, Qwen, DeepSeek, and Mistral.Most developers do not set out to manage GPU infrastructure. They set out to build a product. At some point the API bill arrives, or the rate limit hits, or the compliance team asks where user data is going, and suddenly infrastructure is the job.
Token Factory exists to fix that. This post covers exactly who it is for, what problem it solves, and what is coming.
Token Factory sits between two options that both have real costs. Understanding the gap is the fastest way to know whether it is the right fit for you.
The first option is a proprietary API from OpenAI, Anthropic, or Google. Fast to start, but the economics break down at volume. Output tokens on frontier models cost multiples of what equivalent open-weight models cost, and that per-token rate compounds every time you add an agent call, a RAG retrieval, or a background summarisation job. Solo developers and small teams feel this hardest: no enterprise discount, no committed spend negotiation, just the rack rate.
The second option is self-hosting. Rent a GPU from packet.ai or another provider, install vLLM or SGLang, configure your serving parameters, handle scaling, watch for OOM errors, and maintain the stack as model versions change. For teams at scale, this is the right call. For a solo developer or a small team shipping their first AI feature, it is weeks of infrastructure work before a single user sees the product.
packet.ai's RTX 6000 Pro starts at $0.66/hr on Dynamic. That is 68% below RunPod's $2.09/hr for the same GPU. But even at that price, you still need to deploy and maintain the serving stack yourself.
Token Factory takes that same infrastructure and wraps it in a clean per-token API. You get the cost profile of open-weight models on packet.ai hardware, without the ops work.
The gap Token Factory fills
Proprietary APIs are easy but expensive at volume. Self-hosting on GPU cloud is affordable but requires infrastructure work. Token Factory is the option in the middle: managed inference on open-weight models, accessed through the API your code already calls, on packet.ai's hardware.
Token Factory is not a fit for every workload. It is designed for specific developer profiles where the gap between what proprietary APIs cost and what open-weight models can deliver is largest. Here are the four profiles.
This is the biggest crowd and the one Token Factory is built around. You are shipping a side project, a SaaS MVP, or a client tool. Running a 70B model yourself requires a GPU with enough VRAM, a serving framework, and ongoing maintenance. That overhead makes no sense for a solo builder.
Token Factory gives you access to Llama, Qwen, DeepSeek, and Mistral through the same OpenAI SDK call you already use, with nothing to configure or maintain. Point your existing ChatOpenAI client at Token Factory, swap the model name, and ship. The GPU infrastructure, model serving, and scaling are handled for you.
For cost-sensitive solo builders, this is the most important part: you pay per token, not per hour. If your app generates no tokens at 3am, your bill is zero. No idle GPU cost. No reserved instance fee. Scale-to-zero by default.
An AI feature that costs $50/month in development does not cost $50/month at 10,000 daily active users. Output tokens on frontier models compound fast, and most production tasks, including summarisation, classification, extraction, structured output generation, and RAG, do not require GPT-4o-level reasoning. Llama 4 Scout, Qwen3 32B, and Mistral handle them at output quality that is indistinguishable to end users.
Token Factory gives you access to those models through the exact same OpenAI SDK call. The migration is a config change, not a rewrite. Your LangChain chains, your LlamaIndex RAG pipeline, your custom tool-calling loop all carry over.
Agentic workflows are the most cost-sensitive inference pattern. A single user action can trigger 15 to 30 model calls as an agent reasons through steps, calls tools, and synthesises results. At frontier API pricing, that compounds to a number that does not fit most product budgets.
Token Factory supports function calling and streaming natively. These are the two capabilities agentic applications depend on most. If your agent loop uses tool use and streamed output, it works with Token Factory without changes to the application logic.
Building a product on one proprietary API is a business risk most founders only notice after the first price increase or model deprecation. Token Factory is an OpenAI-compatible endpoint backed by packet.ai's GPU scheduling infrastructure, which means it plugs into any routing layer or fallback chain that already speaks the OpenAI API format. The switching cost, if you ever need to move, drops to a config change.
Token Factory is right for you if
Token Factory is not right for you if
Most developers land on Token Factory after trying one of the other two options and hitting a wall. Here is the honest comparison.
OpenAI-compatible means Token Factory exposes the same HTTP endpoint structure as OpenAI's API. The route is POST /v1/chat/completions. The request body follows the same schema: model, messages, temperature, max_tokens, stream, tools. The response format matches what the OpenAI SDK expects.
Switching to Token Factory requires changing three values: the base_url, the api_key, and the model name. Three lines in your config. Everything else stays the same.
Change base_url
Point your OpenAI client at Token Factory's endpoint instead of api.openai.com. LangChain's ChatOpenAI, LlamaIndex's OpenAI LLM class, and any direct API call all accept this as a single parameter.
Swap your api_key
Use your packet.ai Token Factory API key. Your OpenAI key stays with OpenAI. No changes to how you manage or store credentials.
Update the model name
Change gpt-4o to the open-weight model Token Factory serves, such as a Llama, Qwen, or DeepSeek variant. Run your evals on this model before sending production traffic.
Check before you migrate
OpenAI compatibility covers the API surface, not model behaviour. Llama and Qwen do not produce identical outputs to GPT-4o on every task. Run your eval suite on the target model before sending production traffic. For summarisation, classification, and extraction, output quality is comparable. For complex multi-step reasoning, frontier models still lead.
If you outgrow Token Factory and want to run your own serving stack, packet.ai's GPU cloud is the most affordable place to do it. These are the current live rates, verified from packet.ai/pricing.
All prices in USD per GPU-hour, published starting rates. Verified from packet.ai/pricing, September 2026. Competitor rates from their public pricing pages.
The RTX 6000 Pro at $0.66/hr is the standout for solo developers running 30B to 70B models. It has 96 GB GDDR7 on a Blackwell chip, at 68% below RunPod's rate for the same workload. For anyone self-hosting Qwen3 32B or Llama 4 Maverick, that is the card.
packet.ai started as a GPU cloud. The core product is giving developers access to NVIDIA hardware at prices that reflect what compute actually costs, without hyperscaler markup. Token Factory is the layer above that: the same hardware, packaged as a managed inference API for teams who do not want to run the serving stack themselves.
The two products solve different problems and work together on the same account.
Token Factory is launching soon. Here is what packet.ai has published publicly on the Token Factory product page:
Pricing and specific model availability are not published yet. Those decisions are still being finalised.
For background on why packet.ai built Token Factory and what problem it was designed to fix, read Why We Built Token Factory. For the technical deep dive on how it works under the hood, see Token Factory: How We Built a 98% Cheaper OpenAI Alternative.
On pricing and models
No Token Factory pricing or model tier details are confirmed at the time of writing. Everything above is drawn from what packet.ai has published on the Token Factory product page. For the latest, check packet.ai/token-factory directly.
Token Factory is launching soon
Join the waitlist. Get early access.
Open-weight models. Drop-in OpenAI compatibility. No GPU to manage. Built on packet.ai's infrastructure.
Last reviewed: September 1, 2026. Token Factory is launching soon. All product details above are drawn from packet.ai/token-factory. GPU prices verified from packet.ai/pricing. No Token Factory pricing or model tier details are confirmed. Check the product page for the latest.
Same models. Same API. Fraction of the cost. Start free — no credit card required.
Start Building →
