This AI agents 2026 exchange compares the LLMs companies admit to using against what they actually run. Shadow model usage is the gap between AI policy and AI reality in 2026.
The shadow model gap is the difference between the AI models a company officially approves and the ones employees actually use — often newer, cheaper, or more capable models reached through personal accounts. In 2026 the gap keeps widening as models proliferate, and every unapproved model is data leaving through an unmonitored door.
Companies admit to one set of AI models in public. Inside, teams run another. The Shadow Model Exchange will be the anonymous, aggregated view of what LLMs companies actually run in production versus what they admit to — the shadow-AI gap nobody publishes.
If this were real data, these bars would show the share of surveyed companies that publicly acknowledge each model family vs. the share that actually run it in production. These numbers are invented for illustration.
Illustrative pattern to watch for: open-weight models are the ones teams rarely admit to — which is exactly what makes them invisible to security teams.
A published version of this matrix would map each vendor's models to where they actually show up: officially sanctioned, team-level shadow use, or embedded inside third-party tools.
| Model family | Vendor | Illustrative status | Why it matters |
|---|---|---|---|
| GPT-5 | OpenAI | Sanctioned · actually run | The one on the invoice — visible, budgeted, monitored. |
| Claude | Anthropic | Sanctioned at some, shadow at others | Popular with engineering teams before official approval lands. |
| Gemini | Embedded via Workspace tools | Often running without a formal "AI purchase" decision ever being made. | |
| Llama | Meta | Shadow self-hosted | Deployed on internal GPU boxes — no API logs, no vendor bill. |
| DeepSeek | DeepSeek | Shadow individual use | Free tier pulls developers in under the radar. |
| Mistral / Qwen | Mistral / Alibaba | Embedded in third-party apps | Your vendor's "AI feature" is a model you never approved. |
All statuses above are illustrative examples of the format, not real findings.
The published exchange will aggregate fully anonymous responses: no names, no emails, no company identifiers. Respondents report (1) which models their company publicly acknowledges and (2) which they actually run in production — including shadow use. Results are published only in aggregate once enough responses arrive; per-company data is never disclosed. The live survey is the Shadow Model Census.
Context: GoodFirms' AI-tool tracking of 14,022 active companies finds ChatGPT used by 11,359 companies, Claude by 10,103, and Gemini by 4,294 (visible adoption, not shadow use). Stack Overflow's 2026 developer survey found only 24% of companies have published an approved AI toolset or formal AI policy — the reason the gap exists at all.
The Shadow Model Exchange is a planned anonymous, aggregated view of which LLMs companies actually run in production, compared against what they admit to.
No. Everything on this page is an illustrative sample, clearly badged as such, shown only to demonstrate the concept until real census data arrives.
Respond anonymously to the Shadow Model Census with the LLMs you run in production and the exchange aggregates the results.
Much enterprise AI use is shadow IT: Stack Overflow's 2026 developer survey found only 24% of developers say their company has an official AI policy.
GoodFirms' AI-tool tracking of 14,022 active companies finds ChatGPT used by 11,359 companies — the most widely used by far.
The Shadow Model Exchange is a planned anonymous, aggregated view of which LLMs companies actually run in production versus what they publicly admit to — the gap known as shadow AI. It launches publicly once enough real, anonymous responses arrive.
No. Everything on this page is an illustrative sample, clearly badged as such, shown only to demonstrate the exchange's format. The published exchange will launch after enough real, anonymous responses come in through the Shadow Model Census.
Respond anonymously to the Shadow Model Census with the LLMs you run in production and the ones your company publicly acknowledges. Once enough responses arrive, we publish aggregated, anonymized deltas — no names, no emails, no company identifiers.
Much enterprise AI use is shadow IT: Stack Overflow's 2026 developer survey found only 24% of companies have published an approved AI toolset or formal AI policy, so most model choices are invisible to security and compliance teams — creating unknown attack surface.
GoodFirms' AI-tool tracking of 14,022 active companies finds ChatGPT used by 11,359 companies, Claude by 10,103, and Gemini by 4,294 — but that tracks visible adoption, not the shadow usage this exchange is designed to reveal.
ModelMinder watches the AI models your stack depends on — quiet deprecations, silent version swaps, behavior drift — so shadow models never blindside you.
Get ModelMinder early access