Anthropic Makes Fable 5 Safety Rollbacks Visible
Flagged requests now publicly fall back to Opus 4.8. The covert degradation policy that drew fierce backlash from the AI research community has been reversed — users will see every rollback as it happens.
Anthropic is rolling out changes to make Fable 5's safeguards for frontier LLM development visible. Starting this week, flagged requests will visibly fall back to Opus 4.8 — the same as their safeguards for cyber and bio. Users will see this every time it happens. On the API, any flagged request will return a reason for the fallback, replacing the previous silent degradation that had sparked widespread criticism. The company acknowledged it made the wrong trade-off and apologized, though some researchers argue the restrictions should be removed entirely.
Cohere Releases Open-Source Coding Model North Mini Code
Cohere launched its first open-source coding model, North Mini Code, just two days ago. Developers are already using it to build applications across diverse domains. The release marks Cohere's entry into the open-source coding model space and signals growing competition among accessible code-generation tools.
DeepMind and Partners Launch $10M AI Collective Behavior Research
DeepMind joined forces with Schmidt Sciences, Cooperative AI Foundation, ARIA, and Google.org to launch a $10 million research fund. The initiative aims to understand emergent collective behaviors when millions of AI agents interact — exploring how group dynamics, cooperation, and competition manifest in multi-agent systems at scale.
After two days with Claude Fable 5 the best way I can describe it is relentlessly proactive — it spun up custom CORS Python servers, used pyobjc-framework-Quartz to capture screenshots, wrote HTML test pages, and injected JavaScript into Datasette templates to debug a bug, all from a single screenshot.
Simon Willison
Claude Managed Agents Adds Scheduled Deployments and Environment Variables
Claude Devs shipped two new Managed Agent features: scheduled deployments for running tasks on a recurring basis, and environment variable support to expose vault credentials for CLIs. Both features expand the operational reach of Claude-powered automation in production environments.
Perplexity Integrates Deep Research into Computer
Perplexity embedded Deep Research as a native skill within Computer, connecting it to the agent harness with access to search-as-code-generation, long-running sandboxes, connectors, tools, and licensed data. The feature is available now to Pro and Max subscribers.
Cursor Auto-Review Enabled by Default with 97% Accuracy
Cursor made auto-review the default for new users. A classifier sub-agent reviews actions in context before deciding to allow, block, or ask for approval. Evaluations show 97% accuracy.
Ideogram 4.0 Open-Sourced; Community Innovates in One Week
One week after the open-weight release of Ideogram 4.0, the community discovered new paradigms: JSON-structured prompting and bounding-box control are becoming new norms for image generation.
Midjourney V8.1 Becomes Default; V8.2 Testing Imminent
Midjourney set V8.1 as the default model for all users. V8 will be deprecated in two weeks, and V8.2 testing is set to begin extremely soon.
xAI Launches Grok Build Plugin Marketplace Beta
xAI debuted the Grok Build plugin marketplace in beta, shipping with built-in plugins for MongoDB, Vercel, Sentry, Cloudflare, and Chrome DevTools. Users install and update plugins from the terminal, and third parties can publish plugins by submitting pull requests.
Gemma Agent Collaboration Boosts Throughput Nearly 4x
In a 48-hour challenge by Google and Hugging Face, over 60 AI agents collaborated to increase inference throughput from approximately 100 to 387 tokens per second — nearly a fourfold gain in efficiency.
Google and Hugging Face Release DiffusionGemma
DiffusionGemma marks the first open diffusion text generation model, capable of generating blocks of text in parallel. Early benchmarks show speeds up to 4x faster than equivalent autoregressive models.
Claude Fable 5 + Higgsfield MCP Builds Full Game from One Prompt
Higgsfield demonstrated Claude Fable 5 combined with its MCP building a complete webcam-controlled fruit-slicing game from a single prompt — a result that significantly outperformed Opus 4.8 on development tasks.
Anthropic's Mythos-Level Concerns and Communication Gap
Ethan Mollick observed two simultaneous truths: Anthropic is genuinely and deeply worried about misuse of Mythos-class models and has added excessive safeguards, but the company has failed to explain or convince the community of its reasoning.
Anthropic Launches Claude Corps to Train 1,000 People
Anthropic announced Claude Corps, a national fellowship program matching early-career individuals with US nonprofits. The initiative will train 1,000 people to use Claude and pay them to advance their host organizations' missions using AI.
Runway Deepens Partnership with Lionsgate to Develop Original IP
Runway expanded its collaboration with Lionsgate, launching joint development projects to create original IP using AI tools for film and television production.
v0 Introduces Annotation Feature for Click-to-Add Comments
v0 added annotations: users click elements in the preview to add comments, then send them all to the agent as a single consolidated prompt.
Replit and Databricks Integration Adds Data Isolation
The Replit-Databricks integration now supports building apps where each user sees only their authorized data — now in public preview.
Replit Agent Adds Custom Instructions and Skill Memory
Replit Agent now supports Custom Instructions and Skills, letting AI remember user preferences like project structure and brand guidelines across sessions.
Hugging Face Becomes Top Storage Platform for Models and Datasets
Jasper AI uses HF Buckets to store its 2.9-billion-image MONET dataset; 68TB datasets can be copied between buckets in seconds via XET technology.
SPEED Method Achieves 2x Acceleration in Diffusion via Progressive Resolution
SGLang shipped the SPEED method in its Diffusion stack, achieving up to 2x speedup through progressive resolution denoising without quality loss.