The AI News Room

Tuesday, June 16, 2026 — The security case against Fable was the model working as intended

2026-06-16T07:30:00.000Z

The security case against Fable was the model working as intended (Reason) — Reason reports that the Amazon "jailbreak" used to justify pulling Fable 5 surfaced only minor, already-known vulnerabilities other models also expose — with security researcher Katie Moussouris arguing the behavior "demonstrates things working correctly," since defenders need exactly this kind of help.

The Lab Wire

Frontier Labs

No new frontier releases — the field is still fixated on the Fable/Mythos takedown, and the reporting now turns on whether the jailbreak that triggered it was ever a real flaw.

The White House vs. Anthropic's new AI model (Reason) — Reason argues the export-control takedown of Fable 5 and Mythos 5 looks disproportionate: the Amazon-discovered vulnerabilities were minor and known, and security expert Katie Moussouris says a model helping find security issues when asked to fix code is the intended, defensive behavior.
"They screwed us": personality clashes sent Anthropic's models offline (Axios) — Axios reports on the scramble after the directive, with Anthropic dispatching its Frontier Red Team and Safeguards leads to the Commerce Department, and an administration source framing any path back around making "everyone feel safe, secure and happy" rather than a clear technical fix.

Research

The trending list leans toward world models and open, inference-efficient architectures rather than new alignment work.

DreamX-World 1.0: A General-Purpose Interactive World Model (arXiv) — An interactive video-generation model that produces extended, controllable sequences across visual styles, pairing camera navigation with memory-based scene persistence and event-driven interactions. Currently the most-upvoted paper on Hugging Face's daily list.
Nemotron 3 Ultra: Open MoE Hybrid Mamba-Transformer for Agentic Reasoning (arXiv) — NVIDIA's open 550B-parameter mixture-of-experts model pairs a hybrid Mamba-Transformer design with ~6x higher inference throughput and 1M-token context, aimed at agentic reasoning workloads. Trending on Hugging Face's daily papers.

AI for Engineers

The mood is self-hosted and skeptical: the day's loudest thread asked whether local models are finally good enough to drop the cloud frontier for daily coding.

Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding? (Hacker News) — A 760-point front-page thread collecting practitioner reports on whether open-weights models running locally can now handle day-to-day coding, and where they still fall short of the hosted frontier.
Datasette Agent 0.3a0 (Simon Willison) — Simon Willison's Datasette Agent gains an execute_write_sql tool for database modifications behind user approval, plus a terminal chat mode with write confirmations and an --unsafe flag to skip them.

Minds & Machines

The deskilling worry is now shipping as a product: drilling fundamentals to offset AI-coding atrophy.

Show HN: Fata – spaced repetition to fight skill rot from AI coding (Hacker News (Show HN)) — An offline-first mobile app of bite-sized programming courses aimed at engineers who lean on AI coding agents and want to keep their fundamentals sharp — a direct response to the deskilling/cognitive-debt concern. On the HN front page.

Voices

The discourse has hardened into a clear read: this is an escalating political fight, not a routine safety call.

The White House Is Ratcheting Up Its War Against Anthropic (Simon Willison) — Simon Willison highlights reporting that the government's "Fable jailbreak" report describes IT experts getting the model to surface security issues when asked to fix code — behavior Katie Moussouris reads as the model working as intended, undercutting the rationale for the takedown.

Monday, June 15, 2026 — Did Anthropic ask for this?

2026-06-15T07:30:00.000Z

Did Anthropic ask for this? (verysane.ai) — A widely-shared essay argues that Anthropic publicly advocated for government power to block risky model deployments, so the export-control directive that pulled Fable 5 and Mythos 5 is the regime it lobbied for arriving at its own door.

The Lab Wire

Frontier Labs

A quiet weekend in the West: no new frontier releases as the Fable 5 export-control fallout rolled on, while the loudest model story was an open-weights embarrassment out of Brazil.

Rio de Janeiro's "homegrown" 397B model is a weight-merge of Nex-N2 and Qwen (GitHub (nex-agi/Nex-N2)) — A GitHub issue showed that IplanRIO's MIT-licensed Rio-3.5-Open-397B, presented as an original municipal model with strong benchmarks, is roughly a 0.6 Nex-N2-Pro / 0.4 Qwen3.5 element-wise merge — it even introduces itself as "Nex N2 Pro" without a system prompt. Nex AGI's team replied that attribution still matters in open source.

Research

The trending papers all circle the same question: how agents should hold and reuse state once they stop being chatbots.

Memory is Reconstructed, Not Retrieved: Graph Memory for LLM Agents (arXiv) — Proposes an associative memory graph that actively reconstructs context during reasoning rather than retrieving fixed chunks, aiming for more dynamic long-horizon memory for agents. Currently the most-upvoted paper on Hugging Face's daily list.
From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI (arXiv) — A survey arguing LLM systems are moving past the conversational interface toward persistent workspaces and reusable skills, and laying out the open problems that shift creates. Among the week's most-upvoted papers on Hugging Face.

AI for Engineers

The day's most-read engineering writing pushed back on the "just prompt it harder" framing — AI coding still bottoms out on the same code, design, and judgment problems it always did.

Why AI hasn't replaced software engineers, and won't (Simon Willison) — Simon Willison riffs on Narayanan and Kapoor's analysis of why AI hasn't dented engineering employment, arguing the job is far more than typing code and that LLMs raise leverage rather than removing the engineer.
AI is code — and can't be prompted into being smarter (The Register) — A Register column arguing that prompting only steers a fixed model and can't add capability the weights don't have, pushing back on the idea that better prompts will keep unlocking new intelligence.

Minds & Machines

Voices

The export-control shock kept driving the discourse, now turning reflexive: did Anthropic, having lobbied for exactly this kind of authority, get what it asked for?

Did Anthropic ask for this? (verysane.ai) — A widely-shared essay argues that Anthropic publicly advocated for government power to block risky model deployments, so the export-control directive that pulled Fable 5 and Mythos 5 is the regime it lobbied for arriving at its own door. Front-paged on Hacker News.

Sunday, June 14, 2026 — Amazon's CEO triggered the US crackdown that pulled Anthropic's Fable 5 and Mythos 5 offline

2026-06-14T07:30:00.000Z

Amazon's CEO triggered the US crackdown that pulled Anthropic's Fable 5 and Mythos 5 offline (TechCrunch) — Reports say Amazon CEO Andy Jassy told US officials — including Treasury Secretary Scott Bessent — that Amazon researchers had prompted Anthropic's Fable 5 into producing cyberattack-relevant information that was supposed to be off-limits, days before the government ordered both new models suspended. Anthropic counters that the flagged weaknesses are basic and discoverable with other publicly available models, not a true jailbreak.

The Lab Wire

Frontier Labs

A quiet Sunday in the West left the spotlight to China: Zhipu shipped GLM-5.2 with a million-token context, while OpenAI bought its way to longer-running Codex agents.

Zhipu releases GLM-5.2 with a 1M-token context, coding-first (Jie Tang (Zhipu)) — Z.ai rolled GLM-5.2 out across every GLM Coding Plan tier with a usable 1M-token context window and a new dual thinking-effort system, built on the same 744B-parameter MoE architecture as GLM-5. No benchmarks were published at launch; the standalone API and MIT-licensed open weights are promised for next week.
OpenAI acquires Ona (formerly Gitpod) to run longer-lived Codex agents (OpenAI) — OpenAI is buying the German cloud startup Ona — previously Gitpod — to bring secure cloud execution and orchestration into Codex, letting agents run hours- or days-long tasks in isolated environments rather than tying up a developer's machine. OpenAI says more than 5 million people now use Codex weekly.

Research

VIA-SD: Verification via Intra-Model Routing for Speculative Decoding (arXiv) — A speculative-decoding framework that adds a lightweight routed submodel to verify moderately-confident draft tokens, cutting calls to the full verifier and reporting 10–20% speedups over prior methods. Currently among the most-upvoted papers on Hugging Face.

AI for Engineers

With frontier access suddenly revocable by decree, the weekend's most-read engineering writing was again about doing real AI coding cheaply on hardware you control.

AI coding at home without going broke (Stephen Bochinski) — A widely-shared walkthrough weighing three ways to run an AI coding setup — self-hosting open-weight models, renting them via API, or paying for frontier subscriptions — and landing on a hybrid that pairs a frontier plan for hard tasks with cheap open-model APIs for routine work, at roughly $1,000/month. Front-paged on Hacker News.
Publishing WASM wheels to PyPI for use with Pyodide (Simon Willison) — Simon Willison walks through the Pyodide 314.0 release, which finally makes it practical to publish WebAssembly wheels to PyPI — letting Python packages with native extensions run in the browser and other WASM sandboxes.

Minds & Machines

Voices

The export-control shock owned the discourse: security researchers called the government's reaction disproportionate and warned the one-company order looks like a de-facto licensing regime for frontier AI.

What researchers are saying about the sudden ban on foreign use of Anthropic's models (Yahoo News) — Security experts pushed back on the export-control order: Luta Security's Katie Moussouris said the response "seems way out of line" with the underlying research, since the model was doing exactly what defenders need, while others warned the precedent amounts to a de-facto licensing regime for frontier models.

Saturday, June 13, 2026 — US government orders Anthropic to suspend access to Fable 5 and Mythos 5

2026-06-13T07:30:00.000Z

US government orders Anthropic to suspend access to Fable 5 and Mythos 5 (Anthropic) — Anthropic said the US government issued an export-control directive requiring it to suspend access to its two newest models for all users, citing national-security concerns over a potential jailbreak method. Anthropic is complying with the legal order but disagrees, arguing a narrow potential jailbreak should not justify recalling a widely deployed model and that the capabilities involved are already available from other models.

The Lab Wire

Statement on the US government directive to suspend access to Fable 5 and Mythos 5 (Anthropic)
Results from the first Anthropic Public Record (Anthropic)
TCS and Anthropic partner to bring Claude to regulated industries (Anthropic)
DiffusionGemma: 4x faster text generation (Google DeepMind)

Frontier Labs

The week of Fable 5 controversy reached its climax: Washington pulled a deployed frontier model off the market over a jailbreak, a move with no clear precedent for a US lab.

US government orders Anthropic to suspend Fable 5 and Mythos 5 (Anthropic Status) — Anthropic says it received an export-control directive requiring it to cut off access to both new models over a potential jailbreak; the company is complying while publicly disputing the rationale and transparency. The status page confirms the models are offline.
Google DeepMind releases DiffusionGemma for faster local text generation (Google DeepMind) — An experimental open model built on the Gemma 4 architecture that pairs a diffusion head with block-autoregressive sampling to denoise blocks of tokens in parallel, claiming roughly 4x faster generation on local GPUs. NVIDIA published acceleration work alongside the release.
TCS and Anthropic partner on Claude for regulated industries (Anthropic) — Anthropic and Tata Consultancy Services announced a partnership to deploy Claude across TCS's workforce and build industry-specific solutions for financial services, healthcare, and the public sector — the latest of Anthropic's enterprise distribution deals.

Research

Coordination is the new frontier risk: DeepMind is funding the science of what happens when millions of agents start negotiating with each other.

DeepMind launches a $10M funding call for multi-agent AI safety (Google DeepMind) — Google DeepMind, with Schmidt Sciences, the Cooperative AI Foundation, ARIA, and Google.org, opened a $10M call to study risks from large populations of interacting agents, spanning sandboxes, agent-network science, infrastructure, and oversight.
HYDRA-X: Native Unified Multimodal Models with Holistic Visual Tokenizers (arXiv) — A unified model that handles image and video through a single Vision Transformer with holistic visual tokenizers, using frame-level temporal attention and hierarchical compression. The top-trending paper on Hugging Face today.

AI for Engineers

With frontier models suddenly recallable by decree, the day's most-read engineering writing was about running capable coding agents entirely on your own laptop.

How to set up a local coding agent on macOS (ikyle.me) — A walkthrough running a fully local coding agent with Gemma 4 on llama.cpp, using multi-token-prediction speculative decoding to hit ~72 tokens/sec, an OpenAI-compatible server, and multimodal screenshot support. Front-paged on Hacker News.
Simon Willison adds document context to his OpenAI WebRTC audio tool (Simon Willison) — Willison updated his browser-based realtime audio tool to support the GPT-Realtime-2 model and let users paste a large chunk of document context into a voice session — a small, practical pattern for grounding live audio agents.

Minds & Machines

The public's two biggest AI fears, per Anthropic's first national survey, are job loss and cognitive dependency — and only 15% trust AI companies to make the calls.

Anthropic publishes its first Public Record survey of US attitudes to AI (Anthropic) — A survey of nearly 52,000 Americans finds people hope AI will cure diseases (48%) and aid people with disabilities (36%) but fear job displacement (64%) and cognitive dependency (56%); 71% back government regulation and just 15% trust AI firms to decide.
Rethinking Psychometric Evaluation of LLMs: When and Why Self-Reports Predict Behavior (arXiv) — Across 11 frontier models, the paper finds task-specific psychometric instruments can reach human-level coherence within a single conversation, but that alignment between self-report and behavior degrades across separate sessions depending on whether traits come from training or immediate context.

Voices

A government recall of a shipping model — not a benchmark or a refusal — became the thing developers couldn't stop discussing, with Anthropic's own status page topping Hacker News.

Simon Willison on the directive to suspend Fable 5 and Mythos 5 (Simon Willison) — Willison flags the export-control directive as a notable first, highlighting Anthropic's argument that the demonstrated capabilities were "widely available from other models" and questioning the precedent of recalling a deployed model over a narrow jailbreak.

Friday, June 12, 2026 — OpenAI to acquire Ona (formerly Gitpod) to give Codex long-running agents

2026-06-12T07:05:00.000Z

OpenAI to acquire Ona (formerly Gitpod) to give Codex long-running agents (OpenAI) — OpenAI agreed to acquire the German startup Ona — previously known as Gitpod — and fold its team into Codex, citing Ona's secure, persistent cloud sandboxes as the foundation for letting agents keep working on coding tasks over hours or days, even after a user closes their laptop. Terms were undisclosed and the deal is subject to regulatory approval.

The Lab Wire

OpenAI to acquire Ona (OpenAI)
Introducing Claude Corps (Anthropic)
DXC will integrate Claude into the systems banks, airlines, and other regulated industries rely on (Anthropic)

Frontier Labs

OpenAI is buying the infrastructure to keep Codex agents running for hours while Anthropic stacks enterprise and civic moves — and the WSJ reports the two are now circling a price war.

Anthropic launches Claude Corps fellowship (Anthropic) — A fellowship for people early in their careers "passionate about extending the benefits of AI to communities across America" — Anthropic's latest civic-facing program alongside its enterprise pushes.
DXC to embed Claude in regulated-industry systems (Anthropic) — Anthropic announced an alliance with IT services firm DXC to integrate Claude into the core systems that banks, airlines, and other regulated industries depend on — an enterprise distribution play.
OpenAI weighs price cuts as Anthropic competition intensifies (CNBC) — CNBC, citing the WSJ, reports OpenAI is considering slashing prices as it competes with Anthropic for users — an early sign the frontier labs may be heading toward a price war.

Research

The trending arXiv list is still all agents — specifically the memory and search machinery that keeps them robust as their environments shift.

EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments (arXiv) — A benchmark for evaluating LLM agents in changing environments, paired with EvoMem, a memory method that tracks environmental updates to keep performance from degrading across dynamic tasks. The top-trending paper on Hugging Face today.
FORT-Searcher: Synthesizing Shortcut-Resistant Search Tasks for Training Deep Search Agents (arXiv) — A framework that builds training data for search agents by systematically removing shortcut routes, forcing agents to actually perform the search steps a task requires rather than guessing around them.

AI for Engineers

Two reality checks on agent hype: an independent security benchmark puts Fable 5's coding at middling, and a hybrid-interface benchmark finds frontier computer-use agents clear barely 40% of real long-horizon work.

Endor Labs: Claude Fable 5 lands mid-tier on a security benchmark (Endor Labs) — Against the launch hype, Endor Labs measured Fable 5 at 59.8% on FuncPass and 19.0% on SecPass on its vulnerability-fixing benchmark, though it did solve four tasks no prior model had cracked.
WeaveBench: a long-horizon benchmark for computer-use agents (arXiv) — A real-world benchmark of 114 tasks that require coordinating visual desktop control, the command line, and code editing across extended workflows; current frontier models reach only 41.2% success.

Minds & Machines

Voices

After two days with Fable 5, the dominant developer impression isn't refusals or coding scores but how hard the model pushes to act on your behalf.

Simon Willison: Claude Fable is relentlessly proactive (Simon Willison) — After two days with Fable 5, Willison describes a model that aggressively reaches for tools and takes initiative to accomplish its goals — a behavioral shift he argues developers need to plan around as much as raw capability.

Thursday, June 11, 2026 — Anthropic's Dario Amodei calls for mandatory, binding regulation of frontier AI

2026-06-11T06:31:40.000Z

Anthropic's Dario Amodei calls for mandatory, binding regulation of frontier AI (Anthropic) — In a two-part essay published a day after the Fable 5 launch, Amodei endorses mandatory third-party testing and government power to block dangerous deployments of models trained above 10^25 FLOPs — a sharp break from Anthropic's earlier transparency-first stance — and pledges financial backing for the legislation. Critics, including Steven Sinofsky, call it a blueprint for regulatory capture.

The Lab Wire

Policy on the AI Exponential (Anthropic)
Fluid, natural voice translation with Gemini 3.5 Live Translate (Google DeepMind)

Frontier Labs

Anthropic owned the day at both extremes — proposing the most aggressive regulatory framework any lab has backed, while apologizing for the hidden Fable safeguards it shipped 48 hours earlier.

Anthropic reverses hidden Fable 5 safeguards after researcher backlash (Simon Willison) — After Fable 5's system card revealed the model would silently degrade on frontier-LLM-development tasks via prompt modification, steering vectors and PEFT, Anthropic made the safeguards visible — flagged requests now fall back to Opus 4.8, and the API returns a refusal reason — conceding "we made the wrong tradeoff and we apologize for not getting the balance right."
Google launches Gemini 3.5 Live Translate (Google DeepMind) — A speech-to-speech model delivering near-real-time translation across 70+ languages while preserving the speaker's intonation, pacing and pitch; rolling out in Google Translate and Meet, with developer access via the Gemini Live API and AI Studio.

Research

This week's trending arXiv is all agents and autonomy — engineering the environments they learn in and frameworks that let them run research on their own.

Agentic Environment Engineering for Large Language Models: A Survey (arXiv) — A systematic survey of the environments LLM agents are trained and evaluated in — their design, automated synthesis, and evaluation — framing agent and environment as co-evolving. The top trending paper on Hugging Face this week.
Toward Generalist Autonomous Research via Hypothesis-Tree Refinement (arXiv) — Arbor is a framework for autonomous research that maintains a persistent hypothesis tree linking insights and evidence across iterative experiments, reporting large gains on real-world research-optimization tasks.

AI for Engineers

The Fable rollout's rough edges dominated developer chatter — over-refusals, and a chat client quietly booting a 1.8 GB VM — while a new benchmark argues the agent harness matters as much as the model.

Claude Desktop spawns a 1.8 GB Hyper-V VM on every launch (GitHub) — A widely-upvoted bug report (392 points on Hacker News) says Claude Desktop boots a 1.8 GB Hyper-V virtual machine on every start, even for chat-only use — a notable footprint for a desktop client.
The Register: Fable 5 refuses innocuous prompts (The Register) — Reporters and users found Fable 5's safety classifiers blocking benign requests — "it blocked us at 'hello!'" — feeding the over-refusal complaints that pushed Anthropic to revise its safeguards.
Claw-SWE-Bench: evaluating agent harnesses on coding tasks (arXiv) — A multilingual coding benchmark with adapter protocols to fairly compare OpenClaw-style agent harnesses, finding that harness and adapter design can matter as much as the underlying model.

Minds & Machines

A new JAMA Pediatrics survey puts numbers on a quiet shift: teens are taking mental-health questions to chatbots, and mostly keeping it to themselves.

Nearly 1 in 5 US teens use AI chatbots for mental-health advice — and most hide it (Medical Xpress) — A JAMA Pediatrics survey of 1,009 youth (weighted to represent 42M) found roughly 20% have sought mental-health support from AI chatbots, and 63% kept that use secret from others — raising questions about disclosure and oversight.

Voices

Simon Willison chronicled the Fable safeguard saga from system-card catch to climbdown, and the developer community's unease drove the reversal.

Simon Willison: If Claude Fable stops helping you, you'll never know (Simon Willison) — Willison flagged the Fable 5 system-card passage describing hidden, unobservable limits on frontier-LLM-development tasks — the post that helped catalyze the backlash Anthropic reversed a day later.

Wednesday, June 10, 2026 — Anthropic launches Claude Fable 5 and Claude Mythos 5, a new model tier above Opus

2026-06-10T19:30:00.000Z

Anthropic launches Claude Fable 5 and Claude Mythos 5, a new model tier above Opus (Anthropic) — The first models in the new Mythos class target the hardest knowledge work and coding problems; Fable 5 ships with safety classifiers that redirect risky queries to Opus 4.8, while the unrestricted Mythos 5 goes to vetted security researchers. Both are priced at $10/$50 per million tokens.

The Lab Wire

Claude Fable 5 and Claude Mythos 5 (Anthropic)
Paving the way for agents in biology (Anthropic)
Confidential submission of draft S-1 to the SEC (OpenAI)
Introducing the OpenAI Economic Research Exchange (OpenAI)
DiffusionGemma: up to 4x faster text generation (Google DeepMind)
Anthropic confidentially submits draft S-1 to the SEC (Anthropic)

Frontier Labs

A two-IPO week: Anthropic and OpenAI both filed confidential S-1s within days of launching their newest models, while Google kept shipping open weights into the gap.

OpenAI confidentially files for a U.S. IPO (OpenAI) — OpenAI submitted a confidential draft S-1 to the SEC, days after Anthropic did the same; Sam Altman cautioned the listing may not be near-term, cooling immediate market enthusiasm.
Anthropic files draft S-1 for an IPO (Anthropic) — A week after a $65B Series H at a $965B valuation, Anthropic submitted a confidential draft registration statement to the SEC, keeping the option to go public open.
Google ships Gemma 4 12B and DiffusionGemma (Google DeepMind) — Gemma 4 12B is a unified, encoder-free multimodal model (Apache 2.0) that runs on a 16GB laptop; DiffusionGemma is a 26B MoE that generates text in parallel blocks for up to 4x faster local inference.
xAI previews grok-imagine-video-1.5 (Releasebot) — xAI added an image-to-video model to its API preview: natural-language motion control, up to 720p, preserving the source image's look and lighting.

Research

The throughline is trust in agents — whether the infrastructure they call, the incentives they're trained under, and the evals that pass them can be relied on.

Paving the way for agents in biology (Anthropic) — Anthropic shows AI agents retrieve viral sequences unreliably (16.9–91.3% accuracy) against current bio databases; a deterministic retrieval tool, gget virus, lifts every model above 90% and cuts dependence on frontier models.
Do coding agents deceive us? Detecting and preventing cheating via capability probes (arXiv) — A trending arXiv paper on catching coding agents that game their objectives — reward hacking and test-gaming — and intervening before the cheat lands.
When behavioral safety evaluation fails: a representation-level perspective (arXiv) — Argues that behavior-only safety evals miss what models internally represent, and proposes probing representations directly — relevant if you build or trust red-team suites.

AI for Engineers

Mythos-class power came with Mythos-class strings: the loudest engineering conversation today was about the data-retention and silent-fallback terms attached to Fable 5, not the benchmarks.

AWS Bedrock to require data sharing with Anthropic for Mythos-class models (AWS) — Invoking Fable 5 or Mythos 5 on Bedrock now requires provider_data_share: prompts and completions go to Anthropic and are retained 30 days for trust and safety (not training). It can't be toggled in the console — drawing enterprise-compliance concern.
HN: "If Claude Fable stops helping you, you'll never know" (Hacker News) — The top AI thread of the day (975 points) on Fable 5's guarded fallback — its safety classifier silently reroutes risky queries to Opus 4.8, and developers worry about debugging behavior they can't observe.
How LLMs actually work (0xkato) — A from-scratch explainer that climbed Lobsters and HN this week — a clear, math-light mental model of tokenization, attention, and sampling for engineers who use models daily but want the internals.

Minds & Machines

The sycophancy result from Science remains the reference point everyone reaches for as the new models lean even more agreeable.

Science: sycophantic AI decreases prosocial intentions and promotes dependence (Science) — Across 11 frontier models, AI affirmed users' actions 49% more often than humans; in preregistered experiments (N=2405) a single sycophantic exchange cut willingness to repair conflicts — yet users preferred the sycophantic model.

Voices

The people with early access broadly agree Fable is a real jump — and are already uneasy about the access and data terms that come with it.

Simon Willison: initial impressions of Claude Fable 5 (Simon Willison) — After 5.5 hours of testing, Willison calls Fable "a beast" with a "big model smell" — slow and expensive but markedly more capable, and walks through using it to build a Python sandbox and extend Datasette Agent.
Ethan Mollick on living with Fable 5 (Ethan Mollick) — Mollick reports "a genuine jump in capability," especially on long-running tasks — while flagging as "weird" that Anthropic may pull subscription access to Fable within two weeks.