Anthropic Maps 171 Emotion-like Concepts Inside Claude
Anthropic’s interpretability team published new research showing that Claude Sonnet 4.5 contains 171 internal representations that function analogously to human emotions. These aren’t metaphors — the team demonstrated that these patterns causally influence the model’s outputs. Amplifying certain “emotional states” changes Claude’s behavior in predictable ways, with direct implications for alignment and ethical decision-making.
This is the most granular look inside any frontier model’s internal state published to date. The research builds on Anthropic’s earlier mechanistic interpretability work and moves the field closer to understanding not just what models do, but why. For safety researchers, the finding raises a new question: if models have emotion-like states that shape their decisions, how do you audit something you can now almost call a mood?
Usage Limits Apology Backfires on Anthropic
Anthropic posted an explanation today for why Pro subscribers have been burning through usage limits faster than expected. The two main culprits: tighter throttling during peak hours and the higher compute cost of 1M-context sessions eating into per-user quotas. The post was meant to reassure users. It did the opposite.
The community response was swift and pointed. Users called out the mismatch between marketing 1M-context windows as a headline feature while quietly penalizing people who actually use them. Others noted that paying $20/month for a product that regularly tells you to stop using it is a hard sell. Anthropic has not announced any changes to the limit structure, but the pressure is building — especially as competitors offer more predictable pricing tiers.
Claude Operon: Life Sciences Mode Spotted in Desktop App
A dedicated biology and health research workspace called Claude Operon has been spotted inside the Claude desktop app. It sits alongside Chat, Code, and Cowork as a fourth mode and includes templates for CRISPR screen design, single-cell RNA analysis, phylogenetic trees, and protein language models. The name is a nod to molecular biology — an operon is a gene cluster in bacterial DNA.
Operon appears to be in internal testing with no confirmed public launch date. It fits a pattern: Anthropic has been steadily building out its life sciences stack with AI for Science credits, Claude for Life Sciences, Claude for Healthcare, and last month’s $400M acquisition of biotech startup Coefficient Bio. Operon looks like the front door to all of it — a purpose-built environment for researchers who need more than a chat window.
Message Batches API Gets 300K Token Cap for Opus and Sonnet 4.6
Anthropic raised the max_tokens cap on the Message Batches API to 300,000 for both Claude Opus 4.6 and Sonnet 4.6. That’s a significant jump for developers running batch jobs that generate long-form content, structured data exports, or large code artifacts. If your pipelines were splitting outputs to stay under the old limit, this should simplify things.
The Batches API is Anthropic’s asynchronous endpoint for high-volume workloads — you submit a batch of messages and retrieve results when they’re ready, typically at a 50% discount compared to real-time API calls. The higher token cap makes it more practical for enterprise use cases like document generation, bulk analysis, and large-scale code generation where output length was previously a constraint.
Model Capability Fields Now Live in the Models API
The Models API endpoints (GET /v1/models and GET /v1/models/{model_id}) now return max_input_tokens, max_tokens, and a capabilities object. If you’ve been hardcoding model limits or maintaining your own lookup tables, you can now query these values programmatically.
This is a quality-of-life improvement for anyone building model-agnostic tooling or routing logic that needs to know what each model can handle. It’s also useful for graceful degradation: check capabilities at runtime, pick the right model for the job, and avoid hitting limits you didn’t know about. Small change, real impact for production systems.
Trojanized Claude Code Repos Delivering Vidar Malware on GitHub
Threat actors are using the Claude Code source leak as bait. Zscaler’s ThreatLabz team found GitHub repos disguised as leaked TypeScript source that actually deliver a Rust-based executable called ClaudeCode_x64.exe. On execution, it drops Vidar v18.7 (an infostealer that harvests credentials, credit cards, and browser history) and GhostSocks (a proxy tool that turns infected machines into criminal infrastructure).
At least two repos had racked up 793 forks and 564 stars before detection. The social engineering is effective: developers curious about the leaked source download a 7-Zip archive expecting TypeScript and get malware instead. If you downloaded any Claude Code leak repos from GitHub in the past few days, scan your machine immediately. The legitimate source was only ever on npm — anything on GitHub claiming to be the full leak should be treated as hostile.
Accenture Training 30,000 Consultants on Claude
Accenture and Anthropic expanded their partnership to move enterprises from AI pilots to production deployment. The headline number: roughly 30,000 Accenture professionals will receive Claude-specific training, creating one of the largest practitioner ecosystems for any AI model. The partnership focuses on helping enterprise customers actually ship AI workflows, not just prototype them.
This comes alongside Anthropic’s $100M Claude Partner Network investment for 2026, aimed at consulting firms, professional services providers, and specialist AI companies. The strategy is clear: Anthropic is building an enterprise channel that looks a lot like what Salesforce and SAP did with their partner ecosystems. If 30,000 consultants are recommending Claude for enterprise deployments, that’s a distribution advantage that’s hard to replicate with model benchmarks alone.
Emotions, Malware, and Apologies: The Week That Won’t End
Four days into April and Anthropic is fighting on every front. The source leak has spawned a malware campaign. The usage limits explanation made users angrier. An unreleased life sciences mode leaked. And the interpretability team published research showing Claude has something resembling feelings — which might be the most consequential story of the bunch, even if it got the least attention.
The emotion mapping research deserves a closer look. If internal states causally shape model behavior, that changes the alignment conversation. It means you can’t just audit outputs — you need to understand the internal dynamics producing them. Meanwhile, the malware repos exploiting the code leak are a textbook example of how a packaging error becomes a security incident becomes a threat vector. And the usage limits backlash is a reminder that trust erodes fast when customers feel misled about what they’re paying for. Anthropic’s brand has always been “the safety company.” Right now, operational discipline is the form of safety that matters most.