Anthropic adds Claude Fable 5 to Premium Plans

#1 𝕏

OpenAI’s GPT-5.6 Sol set a new state-of-the-art on “The Last Ones” cyber range and, through Codex Security, is already helping teams find, validate, and fix real-world code vulnerabilities.

#2 📝 Simon Willison

Claude make Fable 5 permanent - Anthropic announced that Claude Fable 5 will be included in all Max and Team Premium plans starting July 20 at 50% of limits, with Pro and Team Standard users retaining access via usage credits and receiving a one-time $100 credit. Simon suggests competition from OpenAI's GPT-5.6 Sol (and Kimi 3) pressured Anthropic to reverse plans to make Fable 5 API-only.

#3 𝕏

NVIDIA AI’s Nemotron 3 Embed models swept the LMEB leaderboard, with the 8B and 1B variants ranking #1 and #2. This demonstrates their top-tier performance on memory-intensive, long-context conversational embedding tasks.

#4 𝕏

NVIDIA AI NeMo AutoModel now supports @huggingface Diffusers, extending its open-source fine-tuning library from Transformers to image and video models. It offers ready-to-run full and LoRA recipes to streamline distributed training.

#5 𝕏

OpenAI released the Codex Security plugin, which installs directly in Codex and adds a “Try in chat” button to launch a prebuilt security scan prompt. Simply select your code folder and press “Send” to run the scan.

#6 𝕏

Philipp Schmid demos automated PR triage with Gemini Managed Agents by sandboxing GitHub CLI using a dummy GH_TOKEN, intercepting GitHub API calls at the network layer to swap in your real token via an egress proxy, and reusing the container to keep gh installed—all while kee...

#7 𝕏

Jason Zhou launched Treg, an open-source, self-hostable skill & secret registry that bundles skills with secrets/CLI/endpoints, injects auth server-side (so agents never hold keys), and logs every agent/user call—eliminating the need to paste keys into Claude Code or Slack DM...

#8 📝 Ampcode Chronicle

From Agent to Agent - Amp agents can now spawn other agents on orbs, your local machine, or any other machine, and they can send messages and files to each other. Example workflows include spinning up orb threads for side quests, running four parallel low‑mode threads to test a flow in Chrome at four screen sizes, pulling files from an abandoned prototype into a current workspace, or starting a thread on cloud-dev-box to run each row of a test matrix at least 10 times.

#9 📝 Simon Willison

LLM cliché highlighter - Simon created a small web tool (built by Fable 5 vibe code) that highlights common clichés found in LLM-generated writing, showing patterns and flagged sentences. The tool highlights ten common patterns, and the post includes a screenshot of the interface.

#10 𝕏

Guillermo Rauch launched ▲ AI Gateway, Vercel’s new unified API for LLMs and vector databases, offering built-in routing, caching, and real-time observability.

#11 𝕏

Aravind Srinivas announced that the Perplexity Agent API now supports custom skills, complete with detailed documentation and a research PDF cookbook to help developers build and integrate new Skills.

#12 𝕏

Cognition launched the FrontierCode leaderboard, a live page tracking merge-ready code performance of models like Grok 4.5 and Inkling, complete with full methodology and sample tasks.

#13 📝 Claude Code Blog

Zero risk isn't the job: a CISO's guide to agentic AI - A CISO-focused guide that argues eliminating all risk from agentic AI is unrealistic and provides principled advice for managing and mitigating risks when deploying agentic systems in enterprises.

#14 𝕏

Andrew Ng launched a short course with Cerebras on building LLM applications for fast inference using the Wafer-Scale Engine.

#15 📝 Surge AI Blog

Cross-Benchmark Generalization for Long-Horizon Agentic Tasks - Explains how post-training on Surge AI's agentic RL environments yields generalization to external tool-use benchmarks like Toolathlon, τ²-Bench, and BFCL-V4. Argues that agentic RL training improves long-horizon tool-use capabilities across benchmarks.

#16 𝕏

Madhu Guru warns that enterprises stall at basic chatbots because they lack the specialist talent to build rigorous eval frameworks and model-agnostic harnesses for routing, multi-agent orchestration, context management and tool integration.

#17 in

Dharmesh Shah observes that a crowded LLM landscape—OpenAI, Anthropic, Google, xAI and open-source models like GLM and Kimi—is driving down costs and making platform switches easier.

#18 𝕏

Santiago launched TapNow Creative OS, an AI-native video creative system where teams can craft briefs, import reference images and generate clips to direct and produce projects.

#19 𝕏

Thariq advises building quick prototypes—mockups, schemas, data models, PoCs—to catch misalignments early and avoid wasting tons of tokens on unwanted AI outputs.

#20 𝕏

Aravind Srinivas shows how Sun Microsystems plunged from a $205 B peak valuation (394 B inflation-adjusted) to a $7.4 B Oracle sale—a 96% loss. He warns that open-source AI models on local commodity hardware could trigger a similar upheaval for today’s enterprise incumbents.

#21 𝕏

Guillermo Rauch open-sourced the Next Evals OSS benchmark suite on GitHub (github.com/vercel/next-evals-oss), giving developers full access to run and contribute performance and feature evaluations for Next.js.

#22 𝕏

Ali Ghodsi is raising funding at a $188 billion valuation to turbocharge Databricks’ AI stack with Unity AI Gateway (multi-AI cost governance), Genie (AI coworkers fluent in your business data), and Lakebase (serverless Postgres for AI agents).

#23 ▶️

OpenAI is being sued for stealing, again…

Fireship

#24 📝 OpenAI News

A scorecard for the AI age - OpenAI defines a "Useful Intelligence per Dollar" scorecard that measures useful work completed, the full cost per successful task (including compute, employee time, retries, and human review), dependability categories (ready to use / needs correction / needs escalation), and whether AI value improves at scale. It also announces GPT‑5.6 (released last week) in three tiers—Sol, Terra, Luna—and claims GPT‑5.6 Sol set a new state of the art on DeepSWE v1.1 (Artificial Analysis Coding Agent Index) at 72.7% versus Claude Fable 5’s 69.9%, while using 54% fewer output tokens and delivering a 36.2% lower estimated API cost.

#25 𝕏

Madhu Guru argues that enterprises won’t buy Kimi directly but will consume it via Google Cloud to ensure security, compliance, data residency and chip access—keeping spend within Google’s ecosystem.