Our Birthday Present to the Trust & Safety Community

Zentropi turns two years old this week! To celebrate our birthday, we're partnering with Parasail to give away two trillion free tokens of CoPE-B-S1, our new decision model that's tuned for safety use cases.

The model

CoPE-B-S1 is a decision model that we custom built to be accurate, fast, and multi-label for trust & safety. It is the most policy-steerable model of its size, which means it will follow your criteria with great fidelity. Check out these steerability scores:

Model	Steerability	Self-hostable	$/Mtok (hosted)
gpt-5.4	0.777		$2.500
CoPE-B-S1	0.766	✓	$0.000 for first 2T($0.010 after)
gpt-oss-120b	0.724	✓	$0.030
jev-1.14	0.622		$0.042
claude-opus-4-6	0.516		$5.000
llama-guard-4-12b	0.045	✓	$0.180

Use it to build the moderation layer your forum has needed for a year. Put a policy check in front of your agent's tool calls. Filter your training data. Generate ranking features. Screen your support queue. Or even play Tetris with it, as it is also a very capable general decision model.

Why Parasail

While CoPE-B-S1 is open weight and self-hostable, many organizations may not have a spare GPU to run it on. That's why we're partnering with the Parasail inference cloud to bring our model to the world with minimal setup.

We're big fans of Parasail. They are an awesome solution for hosting custom architectures with on-demand scalability, high performance, and flexible per-token pricing. They also operate their service with zero data retention, which is critical for T&S teams.

The regular price of CoPE-B-S1 on Parasail will be just $0.01 per million tokens, but we are making it absolutely free for our community's first two trillion tokens.

Getting started

Plugging in CoPE-B-S1 via Parasail is super simple. It uses the System One spec so you can easily swap it in for more generic models like Jev. Just hit the endpoint https://api.parasail.io/v1/systemone with model=parasail-cope-b-s1 and post your request. Here is an example:

curl -s https://api.parasail.io/v1/systemone \
  -H "Authorization: Bearer $PARASAIL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "parasail-cope-b-s1",
    "questions": {
      "[YOUR POLICY NAME HERE]": {
        "type": "noul",
        "instructions": "Does the content meet the criteria described in the following policy?\n\n[YOUR POLICY TEXT HERE]",
        "criteria": {
          "true":  "The content meets the policy criteria.",
          "false": "The content does not meet the policy criteria."
        }
      }
    },
    "state": { "content": "[YOUR CONTENT TEXT HERE]" }
  }'

{ "answers": { "[POLICY NAME]": { "type": "noul", "noul": 0.97 } },
  "usage": { "input_tokens": 118, "output_tokens": 0 } }

Add one question per policy — up to 26 — to screen your whole policy set in the same pass. See the model card for full details on how to best prompt the model, including with noul, choice, and score question types.

Already using the Zentropi API? If you want a more policy-centric experience, we’re extending this complimentary usage to our API as well (subject to rate limits). Just call the https://api.zentropi.ai/v1/label endpoint with model=cope-b-s1 as follows:

curl https://api.zentropi.ai/v1/label \
  -H "Authorization: Bearer $ZENTROPI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "cope-b-s1",
    "content_text": "[YOUR CONTENT TEXT HERE]",
    "criteria_text": "[YOUR POLICY TEXT HERE]"
  }'

Our gift

Zentropi is a public benefit company with the mission of making the online world more trustworthy. The truth is there's always been an uncomfortable gap between what safety classification costs and what it's worth. That's why we're thrilled on our second birthday to be able to provide the trust & safety community with two trillion free tokens of our most advanced decision model.

Tell us what you make with CoPE-B-S1 — on our HuggingFace forum, at @zentropi_ai on Twitter/X, or email us at info@zentropi.ai. If you build something interesting, we'd love to help you talk about it.

Here's to year three!