Inexpensive LLM API providers for open source models

Nick AntonaccioAdmin
Jul 08, 2026 at 21:09 (edited, 13 revisions)

TL;DR: ClinePass is now my primary LLM API provider - it's been a fantastic buy. Configuration prompts for using ClinePass in Pi and other agents are below.

NOTE: I'm deciding against featherless because of context size limits and other limitations on the lower cost plans see the post below NOTE 2: ClinePass has been a valuable provider. For the near future I will use it as my go-to for inference, and only pay OpenRouter when needed.

Featherless is an interesting LLM API provider. They offer truly unlimited token usage on open source models, for a low yearly fee. If you're not using a high volume of tokens daily, then a pay-per-token plan like Openrouter may end up being cheaper, but if you use a ton of tokens regularly, their offering may be hard to beat for the price: $300 per year for a premium account with 4 concurrent requests ($25/month paid yearly).

That may potentially be a great plan for a single developer who regularly uses high volumes of inference tokens for solo project work. Here's an explanation from the Featherless blog:

https://featherless.ai/blog/llm-api-pricing-comparison-2026-complete-guide-inference-costs

They offer models from:

DeepSeek 4
Qwen 3
Llama 3.1
Mistral
Gemma 3
Kimi K2
GPT OSS

Their 'scale' plan offers 8 concurrent requests for $75 per month, and they can build custom plans for production projects which require greater concurrent usage.

History

Nick AntonaccioAdmin
Jul 04, 2026 at 02:20 (edited, 9 revisions)

Cline-pass is another lower volume provider of open source model inference, which may be practical if your usage is not as high:

https://cline.bot/cline-pass

They offer an inexpensive rate-limited monthly plan at $4.99 for the first month, then $9.99/month, with access to:

GLM 5.2
Kimi K2.7 Code
Kimi K2.6
DeepSeek V4 Pro
DeepSeek V4 Flash
MiniMax M3
MiMo V2.5 Pro.
MiMo V2.5
Qwen3.7-Max
Qwen3.7-Plus

They claim 2x-5x more usage for the price, compared to 'standard API rate limits', and you can purchase a full year for 33% off the monthly price ($79.92 for the full year).

Even though it's designed for use with Cline, you can use Cline-pass with other agents.

NOTE from future Nick: ClinePass is GREAT I'll be using it as my primary LLM API for the near future. It's fast and has all the best open source models at a fraction of the cost of other providers. I'll use it up to the rate limit, and then supplement with Openrouter tokens as needed to complete big projects - or perhaps just set up some additional ClinePass accounts...

History

Nick AntonaccioAdmin
Jul 05, 2026 at 14:57 (edited, 7 revisions)

Why I'm deciding against Featherless for my current usage:

Although Pi works well with Featherless' 4 concurrent request limit, the $25 entry plan on Featherless chokes context to 32K tokens. Even their $100/month plan caps context at 256K. I actually make use of 1 million token context, constantly. Context limits are listed on the pricing page:

https://featherless.ai/#pricing

Also, because Featherless multiplexes 30,000+ open-source models dynamically across shared serverless GPU clusters, they do not keep massive weights constantly active in memory for low-tier flat-rate accounts.

So for most user's needs, I think the plain old Deepseek API still currently beats Featherless for cost per capability, especially with Pi coding agent. Because Pi preserves context via sessions, you end up getting a huge percentage of cache hits with the DeepSeek V4 APIs, and no other model family beats Deepseek for cache hit cost. With an average cached input expense of ~$0.05 per million, processing 30 million tokens per week costs something like $1.50-$3.00. You're not tied to any monthly base cost, and you're not rate-limited (it just costs more for more use). If your usage is spotty, regularly stops or dips lower, it's really hard to beat that pricing structure.

Clearly, if your open source LLM API expenses constantly exceed $200 per month, then Featherless could certainly end up being a great buy.

With most of my work now being accomplished by Deepseek V4 Flash, I'd need to consistently average 1-1.5 trillion tokens per month, to make the switch worth while.

History

Nick AntonaccioAdmin
Jul 05, 2026 at 15:05 (edited, 9 revisions)

Cline-pass has been absolutely killer in all my initial tests.

I've now got it set up on all my machines.

I torture tested it with a massive pile of vibe coding prompts and agentic tasks all day yesterday and into the night. I built a lot of applications, and had the LLMs perform all sorts of system configurations, app installations, file management tasks, etc., for more than 12 hours straight.

I used every available model to complete multiple tasks which required lots of reasoning, including a large collection of vibe coded games, CRUD apps, and a collection of demos made with the old Rebol 2 programming language. Even the biggest models don't know how to code in Rebol well, so they burn lots of tokens looping around debug error output iterations. See https://aibynick.com/thread/52#post-139 - that was just a small piece of everything I completed in my torture test of cline-pass.

In all that work, I never got above 7% of cline-pass's 5 hour limit, and all that activity burned only 4% of my weekly limit, and only 2% of my monthly limit. That included use of GLM 5.2, Kimi 2.7, and all the other huge models in the cline-pass stable.

The massive volume of work completed with Deepseek V4 Flash seemed to be nearly free - I only ever saw the Cline dashboard show 1% of my 5 hour limit reached, while absolutely thrashing that model constantly for hours. Performance of V4 Flash was ridiculously fast on cline-pass.

I should point out that one significant difference between Cline-pass and OpenRouter is that OpenRouter often simply runs requests through the APIs of the original providers of each LLM, such as Deepseek.com (not always, but for a good percentage of all their requests). Because providers like Cline-pass can run the open source models on their own servers, potentially in the US, any ban of foreign services would not necessarily shut them down.

Diversification across providers reassures my worries a bit, because I think there is a likelihood of potential collapse in the very volatile LLM industry. With open source models that are hostable on GPU infrastructure anywhere in the world, there will likely always be some way to use capable LLMs, without necessarily having to resort to only using self-hosted inference on local hardware. This is really important because the GPUs/VRAM required to self-host large models is extremely expensive - even if you spend 10s of thousands of dollars, performance won't be nearly as fast or as accessible as the APIs we've all gotten accustomed to.

History

Nick AntonaccioAdmin
Jul 04, 2026 at 22:08

To set up cline-pass as a model provider in Pi (in addition to OpenRouter or any other LLM API provider you already have set up), first create an API key at:

https://app.cline.bot/dashboard

Click:

Account > API Keys > Create API Key

Then, run the prompt below in Pi, replacing <your_api_key> with the API key you created above:

Please set up the cline-pass provider in pi. Here's what to do:

1. Update `~/.pi/agent/models.json` with a provider called `cline-pass` pointing at `https://api.cline.bot/api/v1` using `api: "openai-completions"`, with these 10 models (all using `cline-pass/` prefix):

   | Model | Thinking? |
   |---|---|
   | `cline-pass/glm-5.2` | no |
   | `cline-pass/kimi-k2.7-code` | no |
   | `cline-pass/kimi-k2.6` | no |
   | `cline-pass/deepseek-v4-pro` | yes (thinkingFormat: "deepseek") |
   | `cline-pass/deepseek-v4-flash` | yes (thinkingFormat: "deepseek") |
   | `cline-pass/mimo-v2.5` | no |
   | `cline-pass/mimo-v2.5-pro` | no |
   | `cline-pass/minimax-m3` | no |
   | `cline-pass/qwen3.7-max` | yes (thinkingFormat: "qwen") |
   | `cline-pass/qwen3.7-plus` | yes (thinkingFormat: "qwen") |

2. Then store the API key. **Check if `~/.pi/agent/auth.json` already exists first using `cat ~/.pi/agent/auth.json`** — if it has other provider keys, add the cline-pass key alongside them; don't overwrite the file. Store the key in this format:
   ```json
   {
     "cline-pass": {
       "type": "api_key",
       "key": "<your_api_key>"
     }
   }
   ```

3. Run `pi --list-models` to verify the 10 cline-pass models and all existing models still show.

4. Tell me when it's done.

Nick AntonaccioAdmin
Jul 05, 2026 at 15:08 (edited, 2 revisions)

I used the prompt below to add cline-pass as a provider for Picoclaw on a few PCs, and on my Android phone:

Add the cline-pass API as a provider to Picoclaw by editing your configuration files.

You are running on a phone. Edit your config.json and .security.yml files in ~/.picoclaw/ to add these 10 OpenAI-compatible models pointing at https://api.cline.bot/api/v1 using the API key: <your_api_key>

In config.json, append these entries to the model_list array:

| model_name | provider | model | api_base |
|---|---|---|---|
| cp-glm-5.2 | openai | z-ai/glm-5.2 | https://api.cline.bot/api/v1 |
| cp-kimi-k2.7-code | openai | moonshotai/kimi-k2.7-code | https://api.cline.bot/api/v1 |
| cp-kimi-k2.6 | openai | moonshotai/kimi-k2.6 | https://api.cline.bot/api/v1 |
| cp-deepseek-v4-pro | openai | deepseek/deepseek-v4-pro | https://api.cline.bot/api/v1 |
| cp-deepseek-v4-flash | openai | deepseek/deepseek-v4-flash | https://api.cline.bot/api/v1 |
| cp-mimo-v2.5 | openai | xiaomi/mimo-v2.5 | https://api.cline.bot/api/v1 |
| cp-mimo-v2.5-pro | openai | xiaomi/mimo-v2.5-pro | https://api.cline.bot/api/v1 |
| cp-minimax-m3 | openai | minimax/minimax-m3 | https://api.cline.bot/api/v1 |
| cp-qwen3.7-max | openai | qwen/qwen3.7-max | https://api.cline.bot/api/v1 |
| cp-qwen3.7-plus | openai | qwen/qwen3.7-plus | https://api.cline.bot/api/v1 |

In .security.yml, under the model_list section, add each model name followed by :0: with the API key:

  cp-glm-5.2:0:
    api_keys:
      - <your_api_key>
  cp-kimi-k2.7-code:0:
    api_keys:
      - <your_api_key>
  cp-kimi-k2.6:0:
    api_keys:
      - <your_api_key>
  cp-deepseek-v4-pro:0:
    api_keys:
      - <your_api_key>
  cp-deepseek-v4-flash:0:
    api_keys:
      - <your_api_key>
  cp-mimo-v2.5:0:
    api_keys:
      - <your_api_key>
  cp-mimo-v2.5-pro:0:
    api_keys:
      - <your_api_key>
  cp-minimax-m3:0:
    api_keys:
      - <your_api_key>
  cp-qwen3.7-max:0:
    api_keys:
      - <your_api_key>
  cp-qwen3.7-plus:0:
    api_keys:
      - <your_api_key>

Insert these before the "web:" line. Do NOT overwrite any existing config — only add the new entries. Preserve all existing models and their API keys.

When done, confirm the models are added by listing them.

History

Nick AntonaccioAdmin
Jul 08, 2026 at 21:14 (edited, 4 revisions)

Cline-pass is now serving as my primary open source LLM API provider, but it may be worth pointing out that I still use OpenRouter constantly for all sorts of tasks. Its killer features are: support for hundreds of well known models (commercial and open source), and very importantly, built-in support for the OpenRouter API in just about every AI environment that needs LLM access.

I use it constantly to try out new models - https://openrouter.ai/models is my favorite daily news source about new models, generally the instant they appear.

I use it to test the capability of various models in new applications I develop, and to get models hooked up in new agents & software that I'm testing from other developers.

OpenRouter will also be trusted as a backup for any other LLM API provider I hook up.

OpenRouter is also typically one of the quickest APIs to set up when I want to get a client going with a new app. Most of my clients have already learned how to use it, so they often don't even need my help. The instructions 'just select OpenRouter as the provider and paste in your API key' becomes familiar to everyone.

OpenRouter also has a fantastic reporting system, so it's great for testing cost comparisons between LLMs, for tracking cache hits, and to get all sorts of other feedback about model performance.

Other features like LLM Fusion are also useful.

History

Please login to post a reply.