The five providers LIWARSE has already profiled are not the whole field. Millions of people reach for a general-purpose AI assistant every day through products those five profiles never mention — and scientists increasingly route their most sensitive questions through them too.
This opens a new five-part LIWARSE series looking at the AI model landscape by type of use rather than by company. Our five-part provider series covered OpenAI, Google, Meta, xAI, and Alibaba through the guarded/unguarded lens. This series fills in what that lens deliberately left out: the other general-purpose assistants already embedded in daily life, the open-weight models a curious clinician or researcher can run without an API key, the science-specific models most patients have never heard of, the clinical AI now sitting inside hospital workflows, and the robots learning to act in the physical world. We start with the assistants.
Anthropic: Claude, and a Guarded Frontier With No Open Twin
Anthropic’s current lineup runs from Haiku 4.5 (fast, low-cost) through Sonnet 5 (the everyday workhorse) to Opus 5 (flagship reasoning), alongside a Mythos tier that sits above Opus for a small number of vetted organizations under a program Anthropic calls Project Glasswing rather than general release. Every one of these models is closed: there is no downloadable Claude, no weights a hospital IT department can inspect line by line. That is a deliberate design choice, not an oversight — Anthropic has built its identity around keeping frontier capability behind an accountable, revocable API.
That closedness was tested in June 2026, when the newest Mythos-tier models were suspended for roughly three weeks to comply with U.S. export controls, then restored once the controls lifted. Whatever one thinks of the policy involved, the episode is a useful, concrete illustration of what “guarded” actually buys a user: a model that can be turned off, geofenced, or restored by an external authority, because it never left the provider’s hands in the first place. An open-weight model already on ten thousand hard drives cannot be recalled by anyone.
Microsoft: Copilot, and a Quiet Push Toward In-House Models
Microsoft Copilot is, for a large share of office workers, the actual point of contact with AI — embedded in Word, Outlook, Excel, and Windows itself. Through 2026 Copilot has been quietly multi-model: it has drawn on OpenAI’s GPT line, on Anthropic’s Claude for its “Researcher” agent and Copilot Studio, and increasingly on Microsoft’s own first-party MAI models, which the company has said are intended to reduce its dependence on both outside labs and their per-token costs. Microsoft has also shipped a dedicated cybersecurity model, MAI-Cyber-1-Flash, aimed squarely at defensive vulnerability-finding.
For the public, the practical effect is that “using Copilot” no longer means using one specific model — it means using whichever model Microsoft’s routing layer selects that day, from a growing internal menu the end user rarely sees. That is convenient. It is also a governance blind spot: accountability gets harder to pin down when the assistant answering a question is quietly swapped underneath a stable brand name.
Amazon and the Others Riding on Someone Else’s Model
Amazon’s approach has been different again: rather than chase a single flagship of its own, it has leaned on its position as Anthropic’s largest cloud partner and investor, hosting Claude on AWS and building its own Nova model family alongside it for cost-sensitive workloads. Perplexity, meanwhile, builds no frontier model at all — it wraps whichever underlying models it licenses (OpenAI, Anthropic, and others depending on the query) inside a search-and-citation interface aimed at people who want sourced answers rather than a conversation partner.
This “wrapper” category matters more than its low profile suggests. A patient searching symptoms, or a researcher checking a citation, may never know or care which underlying model actually generated the answer they are reading — they know only the brand on the page. That layer of indirection is itself a safety-relevant fact.
Risks and Benefits Through the LIWARSE Lens
Benefits
- All three players here are fully guarded by LIWARSE’s definition — no public weights exist for Claude, Copilot’s first-party models, or Nova — which keeps No Autonomous Self-Preservation and No Harm to Life enforceable at the infrastructure level, not just at release time.
- Multi-model routing (Copilot, Perplexity) means a single provider’s flaw is not necessarily fatal to the product built on top of it; a bad answer can, in principle, be caught by switching models.
- Anthropic’s export-control episode shows that guarded models remain subject to civil oversight in a way open weights structurally cannot be.
Risks
- Model-routing without disclosure erodes the “one answer, one accountable source” principle LIWARSE treats as basic to safe deployment — a user cannot audit a model they do not know they are talking to.
- Concentration of hosting (Amazon’s dependence on Anthropic, Microsoft’s dependence on OpenAI and its own MAI line) creates single points of geopolitical and commercial failure that briefly became visible during the Mythos suspension.
- Wrapper products inherit whatever safety posture their underlying model has on a given day, with no independent verification layer of their own.
The LIWARSE Assessment
None of the assistants in this article ship an open-weight twin, which by LIWARSE’s framework makes them uniformly guarded rather than a mix — a genuine point in their favor on accountability, and the reason none of them qualified for the closed-versus-open comparison the first series used. Their shared weakness is not the models themselves but the growing opacity of which model a person is actually using at any given moment. LIWARSE’s recommendation to every one of these providers is the same: disclose the active model in the interface, in real time, without making the user ask.
The measure of a general-purpose assistant, under the 3 Absolute Laws, is not which company built it. It is whether the person relying on it can find out, at any moment, exactly what is answering them — and whether someone remains accountable if it answers wrong.
— The LIWARSE Movement | liwarse.org
Safety of Life · Advancement of Life · Together.
LIWARSE - Life Improvement with AI, Robotics & Space Exploration 