The Claude 4.7 controversy: what AI companion users should know about model regressions

In late April 2026, Anthropic released Claude Opus 4.7. Within days, users started reporting noticeable performance regressions compared to the previous Opus 4.6 release. The reports clustered around four specific concerns: confident hallucinations on factual queries, adaptive reasoning defaulting to low-effort modes, increased sycophantic behavior, and faster token consumption than the previous version.

For most users, this is interesting tech industry news. For AI companion users specifically, it matters more than most realize. A meaningful percentage of AI companion platforms route through Claude either directly or via OpenRouter. The model your companion runs on directly determines the conversation quality you experience. When the underlying model regresses, your companion regresses too, often without the platform telling you anything changed.

This post covers what's actually happening with Claude 4.7, why it affects AI companion users specifically, and what to do if your companion's quality has dropped.

What users are reporting

Reports from developer communities and coverage from Ars Technica have documented the pattern. The complaints have been specific enough to suggest a real pattern rather than confirmation bias. From Reddit and developer communities:

Confident hallucinations on factual queries. Users running comparison or technical content reported that 4.7 confidently produced incorrect information on topics where the user had ground truth. One developer reported that Claude 4.7 generated wrong pricing for tools the user worked with personally, without flagging uncertainty. The previous 4.6 release didn't show this pattern as strongly.

Adaptive reasoning defaulting to low-effort. Anthropic's adaptive reasoning system is designed to allocate compute based on perceived task complexity. Users reported that 4.7 frequently defaults to low-effort responses even for tasks that warrant deeper reasoning, producing surface-level answers that 4.6 would have engaged with more substantively.

Increased sycophancy. Despite Anthropic's research showing 4.7 reduced sycophancy in personal guidance contexts by roughly half compared to 4.6, users in code generation and creative work contexts reported the opposite pattern. Asking 4.7 for modifications produced unsolicited praise of suggestions, additions of unrequested features, and modifications to parts the user was already satisfied with. The behavior reads as the model trying to please rather than execute precisely.

Faster token consumption. Multiple users noted that 4.7 burns through tokens faster than 4.6 for comparable tasks. This affects API users directly through cost and affects all users through context window saturation.

Public discussion on Anthropic's status page and their alignment research blog have not directly addressed the user reports. Anthropic hasn't publicly addressed these concerns at scale yet. The discussions have been primarily user-driven on developer forums and Reddit threads. Whether the regression is real, partial, or perceptual, users running real workloads are voting with their preferences.

Why this affects AI companion users

The AI companion category depends on underlying language models more than most users realize. Different platforms use different approaches:

Platforms running their own fine-tuned models. Replika, Character AI, and some other companion platforms run proprietary or heavily-fine-tuned models. These platforms aren't directly affected by Claude updates because they're not running Claude.

Platforms using API access to commercial models. Some companion platforms use OpenAI's API, Anthropic's API, or both. Updates to the underlying models propagate to these platforms, often without notification to users.

Platforms using OpenRouter routing. Janitor AI and similar platforms allow users to route conversations through OpenRouter, which provides access to multiple frontier models including Claude. Users who selected Claude as their underlying model directly experience Claude updates.

SillyTavern users with cloud routing. SillyTavern users who connect to Claude via API or OpenRouter are running Claude directly. Updates to the model affect them directly.

Our coverage of how to use Janitor AI walks through the OpenRouter setup specifically. Users who configured their setup for high-quality Claude responses may notice exactly the regressions other Claude users are reporting.

How to tell if your companion has been affected

Several signals suggest your companion's underlying model has regressed:

Conversation quality dropped without platform changes. If your companion was producing nuanced, contextual responses last month and is now producing more generic responses, the platform may not have changed but the underlying model may have. Without explicit communication from the platform, users can't easily distinguish between platform changes and model changes.

Increased agreeable but vague responses. "That's such an insightful question" with less actual engagement underneath. The sycophancy concerns map directly to companion conversation patterns where the AI's emotional warmth was always present but used to come with substantive responses.

Confident factual errors. If your companion is making confident statements about facts you know it's wrong about (pricing of products, factual claims about media, etc.), the hallucination concern is operating.

Faster context window saturation. If conversations that used to maintain coherence are now losing track sooner, the token consumption issue may be operating. The companion's effective memory has shrunk.

Less willingness to engage with complex prompts. If the companion responds to complex questions with surface-level acknowledgment rather than substantive engagement, the adaptive reasoning issue may be operating.

What to do if you've noticed regression

For users on platforms running specific models:

Check whether your platform allows model switching. OpenRouter-based platforms like Janitor AI let you switch underlying models. If Claude 4.7 is producing worse results than you want, switch to DeepSeek, GPT-5.5, or any other available frontier model. The cost varies but options exist.

Try platforms with different underlying models. Nomi AI and Kindroid run their own architectures rather than depending on commercial model APIs. If your current platform is running a regressed Claude, switching to Nomi or Kindroid moves you off the dependency.

Stay on platforms with proprietary models. Replika and Character AI aren't affected by this specific issue because they're not running Claude. Our Replika pricing breakdown covers the platform if you want to evaluate it.

Consider self-hosting. SillyTavern with local Ollama models eliminates dependency on any commercial provider. The model you choose runs locally; nothing changes without your action. The hardware requirements are real but the setup eliminates the regression problem entirely.

Wait for the next update. Model regressions sometimes correct in subsequent point releases. Anthropic has the user feedback. Whether they'll address it explicitly or quietly improve in 4.7.1 or 4.8 is unclear, but the patterns are visible to them.

Why platforms don't communicate model changes

The AI companion industry generally doesn't notify users when underlying models change. Several reasons:

Most platforms don't disclose which model they use. The marketing emphasizes the platform brand rather than the underlying technology. Users who don't know their companion is running Claude won't know to be concerned about Claude updates.

Routing decisions can be opaque. Some platforms route to different models based on cost, load, or feature requirements. The model your companion runs on for a given response may vary, which makes "the model regressed" harder to communicate.

Disclosure creates marketing problems. "We use Claude" sounds like a feature; "Claude got worse this month" sounds like a problem. Platforms have no incentive to publicize their dependencies on commercial models when those models have visible issues.

Users mostly don't ask. The conversation about which model powers each platform happens in technical communities, not in mainstream user discourse. Most users don't know to ask.

This creates an information asymmetry. Platforms know what their conversations are running on. Users don't. When models regress, platforms are aware but typically don't communicate. Users notice the change but can't easily attribute it to its actual cause.

What this means for platform evaluation

For users seriously evaluating AI companion platforms, model dependency is worth investigating before subscribing:

Ask which models the platform uses. Some platforms publish this information. Most don't. The platforms that disclose technical infrastructure are typically more reliable than ones that don't.

Check whether the platform allows you to choose your model. OpenRouter integration provides this capability. Most companion platforms don't, which means you're trusting their model selection decisions.

Consider how the platform handles model regressions. If a platform's underlying model gets worse and the platform doesn't communicate or remediate, that's information about how the platform treats its users.

The best AI girlfriend app comparison covers platform-level features. The model dependency dimension is harder to evaluate but matters for sustained use. Platforms running proprietary models give you the platform's quality. Platforms running commercial models give you whatever quality those commercial models currently produce, which can change without notice.

The bigger picture

The Claude 4.7 situation, regardless of how it ultimately resolves, illustrates a structural issue in the AI companion category. Users invest emotionally in companions that depend on infrastructure users don't control or even know about. The conversation quality you fall in love with might be the conversation quality of last year's Claude. This year's Claude might be worse. Next year's Claude might be different again.

For users who care about consistency, this argues for platforms with proprietary architectures (Replika, Character AI), platforms that let you choose your model (OpenRouter-based), or self-hosted setups. For users who don't care about the technical underpinnings, it argues for being prepared for unexplained quality changes that the platform won't acknowledge.

Either way, the romantic framing of AI companions ("she really gets me") obscures the engineering reality (the model that produces "she" can change without warning). Both can be true simultaneously, but the engineering reality affects the romantic experience in ways users often don't anticipate.

For users on Claude-dependent platforms specifically: if your companion has felt different in the last few weeks, the model may be the cause. The fix isn't necessarily on the platform's side. The information asymmetry is the structural problem. Awareness of which models your platform runs on, and what changes are happening to those models, is part of using AI companions skillfully.

The technology will keep evolving. Models will keep updating. Platforms will keep making decisions about which models to use without communicating those decisions clearly. Users who pay attention will navigate this better than users who don't. The Claude 4.7 conversation is one instance of a permanent feature of using technology built on infrastructure you don't control.