How to use SolCard Inference

Is SolCard Inference right for me?

How do I make my first request?

Create an account, add credit, create a key and send one request. New accounts get $1 of free credit, per the install runbook.

1. Sign up at app.solcardinference.ai. Try the chat app with a real question before you add money.
2. Open API keys and create one. Keys start with sk-si- and are shown once, so store yours in a secrets manager.
3. Give the key a monthly spend limit if you'll share it or use it in an agent. A key at its limit returns 402 key_limit_exceeded.
4. Open Top up to add credit, or Plans to choose a plan. Check the network and token on the page before you send funds, because a crypto payment is hard to reverse.
5. Send a request and log usage from the reply, so you can see how much was cached.

export SOLCARD_INFERENCE_KEY="sk-si-..."

curl https://api.solcardinference.ai/v1/chat/completions \
  -H "Authorization: Bearer $SOLCARD_INFERENCE_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "solcard-inference-fast",
    "max_tokens": 1000,
    "messages": [{ "role": "user", "content": "Say hello in five words." }]
  }'

# OpenAI SDK: base URL includes /v1
from openai import OpenAI
client = OpenAI(base_url="https://api.solcardinference.ai/v1", api_key="sk-si-...")

# Anthropic SDK: base URL has no /v1, and max_tokens is required
from anthropic import Anthropic
client = Anthropic(base_url="https://api.solcardinference.ai", api_key="sk-si-...")

Older IDs still work. flash, large-v2, base and large map onto the new names, and replies report the new ID. SolCard's own docs introduction still uses flash. Note that the old large now means the legacy model, so use the full solcard-inference-large.

Is SolCard Inference expensive?

Yes, against the same base models elsewhere. You're paying for SolCard's modified models and its privacy claims, not for lower token prices. The largest gap is on cached input.

A worked session, by simple arithmetic. Take 2,000,000 cached tokens, 200,000 fresh tokens and 50,000 output tokens, which is a fair shape for a coding-agent session. On Fast that costs about $0.63. On DeepSeek's own API it costs about $0.07 off-peak and $0.13 at peak. So the same session costs roughly 5 to 9 times as much. The session size is an illustration, not a vendor figure.

SolCard's own comparison is fair on some points. It says OpenRouter adds 5.5% when you buy credits and that Venice plans don't cover API use. It also concedes that some Venice uncensored models are cheaper per token. Its pages don't say how far above DeepSeek's own rates its prices sit, and that's the number you should work out first.

Small inconsistency. Max is listed as 10× Starter's usage and as 25% more than pay as you go. Those work out to $240 and $250 of usage, so check the dashboard rather than either figure. SolCard's own estimate of a Fast coding turn is about $0.02, which would put Starter's window at around 60 turns.

Why did a request fail or cost more than I expected?

Usually because of the hold. SolCard reserves your maximum possible cost before each request, then refunds the difference, so a small balance can fail or get its reply shortened.

Streams can end without a bill breakdown. If an error occurs after a stream starts, it ends with no final usage event, so your own cost log may have gaps. Reconcile against the usage page. That reconciliation step is general advice.

Does it really never refuse?

Not quite. The homepage says "Answers everything", but the API documents a policy error, and SolCard says every message passes an automated content check at its model host.

No usage policy could be found. Searches of the docs sitemap and likely addresses found no published policy, even though an API error cites one. So you can't tell in advance what triggers a 403 or whether repeats risk an account suspension, which is its own error, 403 account_suspended. Ask support before depending on it.

What it's sold for. SolCard's examples are reproducing a disclosed CVE in a sandboxed lab, writing detections for a malware sample, and generating adversarial test data for a safety classifier. A model that doesn't refuse also won't warn you when you're wrong, so review output as carefully as you'd review a contractor's work. The law and your own contracts still apply. That's general advice, not a SolCard statement.

The quality question. SolCard shows an Artificial Analysis Intelligence Index score of 32 for both Large and Fast on its homepage, and links out for each. Those scores appear to relate to the base models. Removing refusal training is a modification, and the pages read for this guide don't show benchmarks of the modified models. Test your own task. SolCard doesn't publish what Base is built on.

Is SolCard Inference private?

It makes specific zero-retention claims and describes them in detail, which is more than many do. But they're unverified vendor statements, in a blog post rather than a signed document, and it isn't anonymous.

Don't call this anonymous. You sign in with an email or wallet, SolCard's server logs your IP address, and Cloudflare sees it too. A wallet's payments are public on its blockchain, so a top-up can usually be linked to the sending wallet. That's general reasoning about public blockchains. Paying from a SolCard card can involve identity checks, since its privacy policy describes KYC "depending on your card type", naming Platinum.

The legal pages don't appear to cover this product. solcardinference.ai/terms opens the "SolCard Card Agreement", about prepaid cards, and /privacy is a policy last updated on 26 January 2025 that focuses on card customers. The "Data handling" link leads to a blog post. So no contract promising zero retention could be found. SolCard's card terms also bar people in Hong Kong and sanctioned territories from "any SolCard service", and they allow forfeiting a balance over location misrepresentation. Read them yourself.

Treat the "Check it yourself" section with care. SolCard says it ran a test message through and found zero matches, and it describes checks of the live code. Those are its own tests. Its Network tab advice only shows where your browser sends data, not what happens after it arrives.

What does the Claude Code or Codex install change?

It runs a local proxy all the time, and your own Claude or ChatGPT login passes through it. That's more than the one-sentence prompt suggests, so read the runbook first.

Read these pages as text. Don't pipe them to a shell. The prompt "Add SolCard Inference to my models: solcardinference.ai/install" tells your agent to fetch instructions from a web page and carry them out, with your files and shell. That page can change at any time. Read the runbook yourself first: curl -fsSL https://api.solcardinference.ai/install/claude-code. Then decide. SolCard's docs do say it's written for an agent and isn't a script.

1. To undo it, ask the agent to "remove SolCard Inference from my models".
2. Check that the solcard-inference rows and the hook are gone from your settings file.
3. Check that the line is gone from your shell profile, and the service is stopped.
4. Delete the key at app.solcardinference.ai/keys, since the runbook says it still exists afterward.

Docs versus runbook. SolCard's coding-agents page matched the current runbooks on the main points when checked, including the browser approval. The homepage still says the agent "asks for your key", while the runbook says you approve in the browser and only use a pasted key if your message includes one. Follow the runbook, and don't paste a key into an agent chat if you can avoid it.

Which model should I pick?

Start on Fast, and move to Large only for planning or hard debugging. Large costs about three times as much and always reasons, which adds output tokens.

One odd choice. Base has the same input price as Fast and a higher output price, with a smaller context, and SolCard doesn't say what it's built on. Without a clear reason, there's little case for it. That's a judgment from the published table, not a SolCard statement.

Try before you pay for an agent. Paste the prompt that another assistant refused into the chat app on the free credit, on Fast then on Large. If it doesn't help, you've saved yourself the setup. If you're only after a lower bill, look at the DeepSeek guide or OpenRouter guide instead.

If privacy is the real goal, a model running on your own hardware needs no one's promise. It's the only option here where you can verify where your prompts go, at the cost of hardware and setup.

SolCard Inference FAQ

Sources

Facts checked 5 November 2026. Products change, so confirm anything important on the official pages.

• SolCard Inference docs: Quickstart
• SolCard Inference docs: Pricing
• SolCard Inference docs: Models
• SolCard Inference docs: Rate limits
• SolCard Inference docs: Errors
• SolCard Inference docs: Claude Code and Codex
• Claude Code runbook, read as text
• Codex runbook, read as text
• SolCard Inference homepage
• Keeps nothing: follow one message
• How billing works
• SolCard vs OpenRouter, by SolCard
• SolCard vs Venice, by SolCard
• Add to Claude Code and Codex
• SolCard terms of use
• SolCard privacy policy
• Anthropic: Other LLM gateways
• DeepSeek pricing
• OpenRouter: DeepSeek V4.1 Flash
• OpenRouter: GLM-5.3
• Venice pricing
• SiliconFlow pricing
• CLIProxyAPI, upstream
• CLIProxyAPI, SolCard fork

Keep reading

• How to use DeepSeek
• How to use OpenRouter
• How to use Claude Code
• How to use Codex
• How to use SiliconFlow
• How to use Ollama