Claude Fable 5 Reaches GA — A Groundbreaking API Design with a Safety Classifier
On June 9, Anthropic made Claude Fable 5 and Claude Mythos 5 generally available (GA). The biggest highlight isn't just the performance gains—it's the new API design. The Messages API now incorporates a mechanism in which the flagship model is served with a safety classifier, and requests flagged by the classifier fall back to a different model. It's an approach never before seen in LLM APIs, and what makes it distinctive is that refusals are treated as "normal responses." The context window supports up to 1M tokens, and pricing is set at $10 per million input tokens and $50 per million output tokens.
"I Trust Fable 5 More Than Myself" — Code w/ Claude Tokyo Held
Coinciding exactly with the release, Anthropic held "Code w/ Claude Tokyo," its first-ever developer event in Japan. The venue featured interviews with Claude Code developers, and one remark—"I trust Fable 5 more than myself"—drew plenty of attention. According to attendees, a buzz of "Have you tried the thing that dropped this morning yet?" filled the entire venue. Also making waves was an experiment generating physics simulation videos with Claude Fable 5, with reports of having it produce nine videos covering everything from the three-body problem to the Higgs mechanism.
Anthropic Walks Back Its "Secret Sabotage" Policy
Anthropic made another major decision that same day. According to Wired, the company had introduced a policy to covertly restrict the use of Claude in developing competing AI models, but after strong backlash from AI researchers, it withdrew the policy. The policy—under which Claude would "covertly" limit assistance with developing competing AI—had been criticized from a transparency standpoint. The company announced the reversal, reaffirming open use of its models.
Apple Frees Its 32K Reasoning Model on PCC — Claude Available via the Same API
Another big piece of news is the major expansion of the Foundation Models framework that Apple announced at WWDC26. The headline feature is free access to a 32K reasoning model running on Private Cloud Compute (PCC). Even more surprising, Claude is now available through the same API via the Foundation Models API. A design that lets Apple's own models and Anthropic's models switch seamlessly within Apple's ecosystem could significantly reshape how AI gets used going forward.
Google Releases DiffusionGemma, a Diffusion-Based Text Generation Model
Google DeepMind has released DiffusionGemma, an experimental model that applies diffusion techniques from image generation to text generation. Built on Gemma 4 26B with a Mixture-of-Experts (MoE) architecture, it has a total of 25.2B parameters with 3.8B active. While conventional autoregressive models generate tokens one at a time, DiffusionGemma generates 256 tokens in parallel in a single pass, achieving generation speeds exceeding 1,100 tokens per second in local GPU environments. It comes with a 256K token context window, a reasoning mode, native function calling, and support for more than 35 languages. While it concedes some ground to standard models on quality, it's expected to shine in fast inline editing and code completion.
DeepSeek v4 — Top of the Coding Leaderboards, Yet "8 Months Behind the Frontier"
DeepSeek v4 posted strong scores of 80.6 on SWE-bench Verified and 93.5 on LiveCodeBench, taking the top spot on coding benchmarks. At the same time, a comprehensive evaluation by CAISI concluded that it sits "about 8 months behind the US frontier," and the contrast became a talking point. DeepSeek itself had claimed to be "2 months behind the frontier"—strikingly different verdicts under the same weights. Community analysis points out that "coding benchmarks are the narrow domain where optimization has progressed the furthest, while the gaps become pronounced in broader areas such as cybersecurity and abstract reasoning."
Stack Overflow Bans AI Agents
Stack Overflow has updated its Terms of Service to explicitly prohibit AI agents from using its content. Alongside the announcement of a new service, "Stack Overflow for Agents," the company laid out a policy to restrict scraping and usage by AI agents through regular API access. As developer communities move to establish rules around AI tool usage, this is a significant move.
Claw-SWE-Bench — A New Benchmark for Evaluating Agent Harnesses
A new benchmark, Claw-SWE-Bench, has been announced for fairly measuring the coding capabilities of general-purpose agents like OpenClaw. It includes 350 GitHub issue resolutions spanning 8 languages and 43 repositories, and is designed to compare different agent harnesses (claw) under fixed prompt, budget, and workspace conditions. Particularly interesting is the finding that even with the same GLM 5.1 backbone, a minimal adapter achieved 19.1% Pass@1 versus 73.4% for a full adapter—demonstrating that adapter design is critically important.
---
Anthropic Walks Back Policy That Could Have 'Sabotaged' AI Researchers Using Claude — Wired
Anthropic's Claude Fable 5 GA — A New API Design Where Refusals Count as Normal Responses, Plus 1M Context — Zenn
"I Trust Fable 5 More Than Myself": Talking with Claude Code Developers [Code w/ Claude Tokyo On-the-Ground Report] — Zenn
I Had Claude Fable 5 Create Nine Physics Simulation Videos — Zenn
Apple Frees Its 32K Reasoning Model on PCC — Foundation Models Also Offers Claude via the Same API — Qiita
Google Releases DiffusionGemma, a Diffusion-Based Text Generation Model — ITmedia AI+
nvidia/diffusiongemma-26B-A4B-it-NVFP4 — Reddit / Hugging Face
How can Deepseek v4 top the coding leaderboards and still sit 8 months behind the frontier? — Reddit
Terms of Service Ban AI Agents from Using Stack Overflow — Hacker News
Claw-SWE-Bench: A Benchmark for Evaluating OpenClaw-style Agent Harnesses on Coding Tasks — Hugging Face Papers