anthropic news, stories and analysis

Slope chart showing Claude voice mode expansion from 2025 to 2026. Models with voice support rose from 1 to 3, languages from 1 to 10, and app integrations from 0 to 3. Anthropic added Opus, Sonnet, Gmail, Slack, Canva, and nine new languages.

Claude voice mode expands to Opus and Sonnet models

Claude voice mode is Anthropic's spoken AI interface. It now supports Opus and Sonnet for deep reasoning, plus Gmail, Slack, and nine new languages.

Lars Cornelissen · Jul 24, 2026

Slope chart showing interpretability visibility before and after the J-space discovery. Five behavior categories improve: task progress from 20 to 70, recognition from 10 to 60, commentary from 5 to 80, bias detection from 15 to 55, and cheating signals from 0 to 70. Illustrative values.

Research

Anthropic J-space reveals LLM thoughts, but not fully

Anthropic's J-space exposes hidden words shaping LLM reasoning. Mechanistic interpretability advances, but misuse detection stays unproven.

Lars Cornelissen · Jul 14, 2026

Abstract timeline data art of the drug development pipeline with the headline figure 13.5 years: preclinical research at roughly 3 to 6 years, Phase 1 at 1 to 2 years, Phase 2 at 2 to 3 years, Phase 3 at 3 to 4 years, and FDA review at 1 to 2 years, illustrating the decade-plus wet-lab gap AI drug discovery still faces.

Business

Anthropic drug discovery hits the wet-lab wall

Anthropic drug discovery pushes Claude Science into pharma R&D, but wet-lab costs and data gaps mean no AI-designed drug has cleared FDA approval yet.

Lars Cornelissen · Jul 4, 2026

Abstract waterfall data art of AI detection coverage on AO3: starting from 100 percent of AI-assisted works, roughly 60 percent leaks away to models other than Claude and another 30 percent to text routed through Google Docs or Word, leaving about 10 percent detectable, shown with the headline figure 10 percent.

AI

AO3 Claude detector is a narrow signal treated as a verdict

The AO3 Claude detector is a fan-made skin that catches one specific paste path from Claude into Archive of Our Own, and fandom communities are already treating its red screen as a verdict. The tool's false negative rate is enormous by design.

Lars Cornelissen · Jul 4, 2026

Claude on GB300 in Azure, with each GB300 NVL72 rack carrying 72 GPUs, 36 Grace CPUs, 20 TB HBM, and up to 1.44 exaFLOPS FP4.

AI

Claude on GB300 puts Azure agents on serious new iron

Claude on GB300 is now generally available in Microsoft Foundry, giving Azure teams more inference headroom but a tighter cloud bet.

Lars Cornelissen · Jun 30, 2026

Mythos 5 access reopened to at least 100 approved US agencies and companies, while GPT-5.6 started with around 20 partners and Fable 5 public access remained 0.

AI

Mythos 5 access returns as a government whitelist

Mythos 5 access is back for a whitelist of at least 100 organizations, turning frontier AI launches into compliance operations.

Lars Cornelissen · Jun 27, 2026

Claude Fable 5 guardrails session split showing more than 95 percent of sessions with no fallback and fewer than 5 percent routed to Opus 4.8.

AI

Claude Fable 5 guardrails get a visibility reset

Claude Fable 5 guardrails are now visible after backlash. Builders should log fallback events, cost, and retention before trusting runs.

Lars Cornelissen · Jun 11, 2026

Donut chart showing where a Claude Fable 5 call runs: about 95 percent of sessions get the full Mythos-class model and under 5 percent are silently routed to Claude Opus 4.8.

AI

Fable 5 and Mythos 5 are one model with a bouncer

Claude Fable 5 and Claude Mythos 5 are the same Anthropic weights; a runtime classifier, not the model you call, decides which capability you actually get.

Lars Cornelissen · Jun 9, 2026