Which AI Agents Won't Tell You What LLM Powers Them?

<a href="https://theagentsindex.com/blog/model-transparency-census-2026/" target="_blank" rel="nofollow noopener"><img src="https://theagentsindex.com/charts/model-transparency-census-2026.png" alt="How AI agent vendors disclose their underlying model — The Agents Index" title="How AI agent vendors disclose their underlying model — The Agents Index" width="640" loading="lazy" style="max-width:100%;height:auto" /></a>

Of the 93 AI agents in The Agents Index, 92 could be classified by what they disclose about the model underneath, and 23 of those 92 (25%) name no model at all. Not one of the 23 gives you a model name: their own, a third party’s, any. “Powered by GPT-5” and “Claude inside” are becoming marketing lines in this space, but a quarter of the index says nothing. That decides whether you can control cost, satisfy a compliance team, or switch providers when a model gets worse or a vendor’s pricing changes.

The pattern tracks almost exactly with how the product is sold: as software you configure, or as a managed outcome you buy. Agent frameworks name models 14 times out of 14. Support agents and sales agents, sold as an outsourced team, disclose nothing 13 times out of 16.

How we classified each tool

Every listing in our index carries a modelLLM field, researched and sourced the same way as pricing or API availability. We sorted all 93 current values into five tiers, from most to least transparent:

1. Agnostic / BYOK. You can bring or choose the model, including through an open routing layer such as LiteLLM, OpenRouter, or a built-in Ollama option. A self-hostable product that supports several LLMs sits here by construction: you supply the credentials, so the choice is yours.
2. Closed menu. The vendor runs a fixed set of models it hosts, names at least one of them, and offers no bring-your-own-key or open-routing option. Whether you pick from the menu or the vendor picks for you, the set is theirs.
3. Single-vendor locked. The product runs on exactly one named vendor’s models, full stop.
4. Proprietary, named. The vendor built and named its own in-house model. Not third-party, but not a mystery either.
5. Not disclosed. “Proprietary,” “multi-model,” or “managed,” with no model named anywhere in the vendor’s own public materials.

One listing of the 93 could not be classified at all, and it is excluded from every percentage below rather than counted as a silent vendor. SalesTouch is an agent-tools listing whose modelLLM field reads “Connects to Claude, Codex, Hermes, and OpenClaw (bring your own agent)”. It names which agents call it, not which LLM runs underneath, because nothing runs underneath. The question does not apply to it. So the denominator throughout is 92, not 93, and the one exclusion is named here rather than folded into a tier.

This census also leaves out the index’s one agency listing, Agentix Labs. It builds and operates agents for clients on the clients’ own stack rather than shipping a product of its own, and the index’s own category page states the rule plainly: agencies are “never ranked against the products in this index.” A model-disclosure question does not apply to a service engagement the way it applies to software you install or subscribe to.

The full breakdown

Tier totals across the 92 classified listings: 43 agnostic / BYOK, 15 closed menu, 9 single-vendor locked, 2 proprietary and named, 23 not disclosed. Those five figures sum to 92. SalesTouch is the 93rd and is not in any of them.

The category split is the real finding

Read as one list, transparency looks close to a coin flip. Split by category, it is not.

The eight category samples sum to 93 listings, the agnostic column sums to 43, and the not-disclosed column to 23, matching the tier totals above. Sierra and Botsify are each dual-listed under support agents and agent platforms; ElevenLabs Conversational AI is dual-listed under voice agents and agent platforms. All three are counted once, under their first category, which is why agent platforms shows n=13 against 16 category memberships. Agent tools is reported at n=2 with no percentage: two listings is not a rate, and one of them is the unclassifiable exclusion.

Agent frameworks are the only category in the index where every single listing names its models: 14 of 14. Thirteen are agnostic, which is close to the point of a framework. You install it and supply your own credentials, so the model choice was never the vendor’s to make. AgentsKit is the extreme case, with 25 native adapters spanning a catalog of 184 providers. Claude Agent SDK is the category’s one exception and deliberately so: it is Anthropic’s own harness for Claude Code, not a general-purpose framework that happens to default to one vendor, so being Claude-only is the design rather than an omission.

Agent platforms are the most transparent hosted category at 9 of 13 agnostic, though this edition’s new arrival breaks the pattern. Vellum is a personal assistant whose own documentation says only that a prompt “reaches an AI model” through an unnamed cloud provider; it never names one, in the product or in its docs, so it lands in not disclosed alongside Moveworks, the category’s other silent listing, which discloses nothing beyond an undisclosed multi-model backend behind its own Reasoning Engine.

Coding agents are the category most likely to name a model without letting you change it. Only 12 of 25 are agnostic, the lowest agnostic share of any software-shaped category, but 24 of 25 tell you something. Claude Code, Bolt.new, Muse Code and Gemini Code Assist are each explicit about the single vendor they run on rather than silent about it, and eight more publish a hosted menu. Factory is the one coding agent in 25 that discloses nothing at all.

Support agents and sales & marketing agents are the mirror image: not one of the 16 listings across both categories is model-agnostic, and 13 of the 16 name no model anywhere. Both are sold the same way, as an outsourced fully-managed team or a bundled platform rather than software you configure, and that sales model still comes with an unwritten rule for most of the group: the buyer doesn’t get to know, or ask, what is running underneath. Zendesk AI Agents is the single exception in support, and not into the agnostic column. It names its backbone (OpenAI GPT-4o primary, plus Anthropic, Google and Amazon Bedrock) without letting a customer pick between them. Unify is the equivalent exception in sales, naming GPT-5.4 or GPT-5.6 behind its GTM agent, though as a single locked vendor rather than a menu.

Research agents are the least transparent category with a software shape: 5 of 11 name nothing, against 1 of 11 agnostic. Four of the five silent ones, Elicit, GC AI, Undermind and Webhound, carry the identical “multi-model backend, not publicly disclosed” phrasing. Microsoft Fabric Data Agent sits at the opposite pole: Microsoft’s own documentation states the underlying model, Azure OpenAI’s Assistant APIs on a GPT-4-series model, cannot be changed at all. That is a stricter fact than a closed menu, but a named and disclosed one.

Why this is a real buying question, not trivia

Four concrete reasons the model behind an agent matters even if you never touch a system prompt:

• Cost control. An agnostic or BYOK tool lets you route to a cheaper model for high-volume, low-stakes tasks and a frontier model only where it earns its keep. A vendor charging a flat per-seat or per-resolution price on an undisclosed model gives you no way to reason about their margin or your own unit economics.
• Compliance and data residency. If you can’t find out which model, or whose infrastructure, processes your data, you can’t answer a security questionnaire about it. Bland AI is the one exception worth naming among the 23: on its own site it won’t say which model it runs, but it is explicit that the model runs on its own infrastructure rather than routing through a third-party provider, which is itself a meaningful disclosed compliance fact even without a model name.
• Quality and behavior drift. A vendor can swap the model behind a managed product at any time with no changelog entry a buyer would see. You would only notice output quality or behavior changing. A tool with a documented model, or a choice of one, at least tells you what changed when a vendor announces an upgrade.
• Switching leverage. If a vendor doubles its price or a model’s outputs regress, an agnostic tool lets you point at a different model tomorrow. A locked or undisclosed one leaves you with no lever but to renegotiate or migrate the entire product.

None of this makes “not disclosed” a red flag by itself. An enterprise buyer evaluating Crescendo or Decagon is typically buying an outcome, resolved tickets or booked meetings, and delegating the infrastructure decision on purpose. It does mean that buyer is trusting the vendor’s model choices sight unseen, which is worth naming explicitly during procurement rather than discovering later.

Proprietary doesn’t have to mean silent

Two of the 92 classified listings run their own in-house model and name it: Intercom Fin and TinyFish. Intercom Fin names Apex 1.0 for general resolution and Apex Flash tuned for low-latency voice. TinyFish names Mako, its own web-native model, and sells it as one of five products under a single API key rather than hiding it behind a generic “multi-model backend” line.

That is the useful contrast with the 23 in the not-disclosed tier, which give a category (coding, voice, agent platform, research, support, sales) but never a model name, their own or a third party’s. Being closed-source and proprietary is a legitimate product choice. Declining to name what you built is a separate and stricter kind of opacity, and only two vendors of the 92 draw that distinction in their own favor.

This tier has held at exactly two members across the last two editions of this census. Neither of this edition’s two new entrants, VoiceAgent or Vellum, joined it: one names its providers outright, the other names none at all.

What this census doesn’t tell you

“Not disclosed” here means not stated anywhere in the vendor’s own public materials as of this research pass. It does not mean the tool’s output is worse, or that the vendor is hiding something disreputable. Several of these products (Decagon, Crescendo, Cresta) are well-funded, enterprise-vetted platforms with strong compliance credentials elsewhere. Model disclosure is simply a fact they have chosen not to publish, possibly for competitive reasons.

This is also a census of a category, not a survey of a market. It describes the 93 agents in this index on the date below. We do not claim it describes every AI agent that exists, and a percentage here is a percentage of our index.

And it is a snapshot in a fast-moving field: a vendor could publish its model stack the day after this runs. Perplexity’s Sonar and Windsurf’s SWE models already show that a company can name its own model and offer frontier alternatives. The two are not mutually exclusive.

Methodology

Every row above comes from our own listings’ sourced, gated research, and the “what the vendor says” column quotes each listing’s modelLLM field directly rather than paraphrasing it. Category shares are computed against the current Published corpus: 93 listings, 8 categories, three dual-tagged, re-derived on 2026-08-26, and will shift as we add or re-verify listings, so treat that date as when this was last checked against the live index.

Four rows draw on more than the modelLLM field, and all four are stated here rather than left for a reader to reverse-engineer. CrewAI is classified agnostic on the strength of the LiteLLM and Ollama routing recorded in its integrations, which its modelLLM field no longer repeats. Sim is classified agnostic because it is open-source and self-hostable, so a buyer running it supplies their own model credentials, even though its field claims broad support without naming a model. Glean stays on the closed menu despite a field that reads “model-agnostic,” because what it describes is an administrator switching between three named models in a console rather than any LLM, and a menu of three is a menu. Vellum is classified not disclosed because its own documentation states only that a prompt “reaches an AI model” through an unnamed cloud provider, never naming one anywhere on the product site or in its docs.

Two listings entered the index since the previous edition, and one left. VoiceAgent joined voice agents in the agnostic tier, bringing its own OpenAI, Anthropic and Google model stack plus a Spitch integration for African-language calls. Vellum joined agent platforms in the not-disclosed tier. Clay is no longer listed; its closed-menu row is removed rather than reclassified. ElevenLabs Conversational AI also picked up a second category tag, agent platforms, alongside its original voice agents listing, though its own tier, agnostic, did not change. No listing already in the census changed tier this edition.

If you are evaluating a specific tool, verify directly with the vendor before relying on a classification here: model support is one of the faster-moving facts in this space. To cite these figures, attribute them to The Agents Index model transparency census, 2026-08-26.