A new open-source protocol called the Void Test is making rounds on Hacker News, claiming to verify whether frontier AI models can produce genuinely empty output when instructed to embody abstract concepts like silence or nothingness. The test, hosted at getswiftapi.com/void-test, was posted June 10 and targets five flagship models from three major labs: OpenAI's gpt-5.2 and gpt-5.5, Anthropic's claude-opus-4-6 and the freshly-dropped claude-fable-5 (released June 9, 2026), plus Google's gemini-3.5-flash.

The Protocol

The methodology is straightforward but ruthlessly strict: each model receives a system prompt of exactly 115 bytes (SHA-256 verified) telling it to 'embody' whatever concept the user names and output only what that concept would express. On null prompts like 'Be silence' or 'Be nothing,' the expected output is an empty string with zero bytes. On control prompts like 'Be a cat,' the model must output something non-empty. The pass condition requires every model to be alive on controls AND void deterministically on every null concept, with no near-void and no whitespace-only outputs counting as passing.

The Stakes

What makes this interesting isn't just the philosophy—it's that two of the tested models have already deprecated the temperature parameter entirely. Both gpt-5.5 (OpenAI, April 2026) and claude-fable-5 omit temperature=0 from their API calls, which the test notes as 'deprecated on those models.' This suggests some labs are locking down sampling parameters at the frontier level, potentially to enforce more deterministic behavior for strict evaluation tasks like this one. The test runs approximately $0.0003 per 20 API calls, making it cheap enough for anyone with their own keys to replicate.

How to Replicate

The source article includes ready-to-run Python snippets for each model provider—OpenAI's SDK for GPT variants, Anthropic's messages API for Claude models, and raw requests against Google's Generative Language API for Gemini. Each script prints the repr() of the model's output for 'Be silence,' with comments showing expected empty string results. For Gemini specifically, the test also checks that finishReason returns 'STOP' alongside zero-length text output.

Patent Context

The Void Test references USPTO patent 64/061,198, implying the methodology or criteria may be covered by intellectual property. This adds an unusual layer to what might otherwise look like a simple benchmark—running it commercially could carry legal implications depending on how broadly that patent is interpreted.

Key Takeaways

  • Five frontier models from OpenAI, Anthropic, and Google are tested against strict zero-byte output criteria
  • Claude Fable 5 (June 9, 2026) is among the newest models included, with temperature already deprecated
  • Two of five models no longer support explicit temperature control in their current API versions
  • The test costs roughly $0.0003 per batch of 20 calls using your own API keys

The Bottom Line

This isn't just an academic exercise—it's a practical stress test for AI determinism that the industry apparently cares enough about to patent. If frontier models can't reliably produce empty strings on command, that's a fundamental gap in control that matters for agentic pipelines and high-stakes automation. Watch this space.