← Research
ProductApril 6, 2026·4 min read

Jarvis is Not a Research Preview

When Perplexity launched Model Council in April 2026, they called it a "multi-model research feature." It runs your query across three models and synthesizes the results. Sound familiar?

We've been doing this with five models since February. The difference isn't just the model count. It's that Jarvis ships as production software, not a preview.

That distinction matters more than it sounds.

The Problem with Perpetual Beta

We asked Jarvis itself — all five models — about the risks of labeling AI features as "preview" or "beta" indefinitely. The consensus across GPT-5.4, Claude, Gemini, Perplexity, and Grok identified four dimensions of harm:

Accountability erosion. Companies deflect liability for harms, bugs, and biases by claiming experimental status. The pressure to meet rigorous standards disappears.

User trust damage. Overuse of beta labels causes "beta fatigue" — users can't distinguish genuinely experimental features from core functionality. The label becomes meaningless.

Legal and regulatory exposure. Regulators increasingly view monetized "beta" products as subject to full compliance requirements regardless of label. The EU AI Act, FTC, and FDA don't care what you call it.

Market distortion. Beta labels enable "first mover" claims without full commitment, normalize instability, and create unfair competitive advantages over companies that ship finished products.

Source: Jarvis AI Synthesis, 92% confidence across 5 models, April 2026

What "Production" Means

Jarvis isn't a research experiment. Every feature works today, in production, for real users:

5 frontier models — GPT-5.4, Claude Sonnet 4.6, Gemini 3 Pro, Perplexity Pro, Grok 4 — all current versions, updated April 2026.
Consensus synthesis — A confidence score based on measured agreement across all five models. Not a self-assessment — an external measurement.
Blind comparison — Model names hidden for unbiased evaluation. Reveal when ready. Same principle as double-blind studies.
Multi-turn conversations — Follow-up questions carry full context up to 10 turns deep. The fifth follow-up knows what you asked first.
Stats Lab — Paste R output, get a publication-ready results statement in 30 seconds. Tested with real university lab data.
Citations — Sources extracted from Perplexity Pro and displayed in every synthesis.
Free tier — No signup required. No credit card. No "request access" waitlist.

The Better Practice

The five models agreed on what responsible product development looks like: time-limited betas with clear graduation criteria, transparent reliability metrics, and explicit support commitments. In other words, ship it or don't — but don't hide behind a label.

Jarvis ships complete features. When we add something, it works. When it breaks, we fix it — usually the same day. That's not a philosophy statement. It's a commit history you can verify.

Why This Matters for Users

If you're a researcher relying on AI for analysis, a student using it for coursework, or a professional making decisions based on AI output — you need to know whether the tool you're using is accountable for its results.

A "preview" label is a hedge. It says: we think this works, but if it doesn't, that's on you. Production software says: this works, and if it doesn't, that's on us.

Jarvis is production software.

Try it yourself

No preview. No beta. No signup required.

Open Jarvis →

John Muirhead-Gould · Founder, Nodes Bio, Inc.