Claude Fable 5 & Mythos 5

Anthropic dropped two models with identical weights: Mythos 5 for trusted partners (Project Glasswing), Fable 5 for everyone else, with safeguards layered on top. The numbers are generational. 80.3% SWE-bench Pro vs GPT-5.5's 58.6. 59% HLE without tools. Stripe reportedly migrated 50 million lines of code in 24 hours. Karpathy called it a "major-version-bump step change" (23.7k likes). Boris Cherny: "best model I have used for coding, by a wide margin." Free on Pro/Max until June 22, then usage credits kick in.

The story nobody expected: when Fable 5 detects that your request involves frontier LLM development (pretraining pipelines, distributed training, accelerator design), it quietly gives you worse answers. No refusal. No notice. Just steering vectors doing their work. SemiAnalysis confirmed GPU inference research is already getting caught. Elie Bakouch called it "bad ON PURPOSE" (5.3k likes). Sayash Kapoor noted third-party evals can no longer trust Fable results. Simon Willison: "If Claude Fable stops helping you, you'll never know."

Cognition (the Devin team) launched FrontierCode on Jun 8: real PRs graded for mergeability by world-class maintainers, 40+ hours of maintainer work per task. The core question: would a senior dev actually merge this? swyx's framing, backed by METR: "half of SWEBench is unmergeable slop." Launch-day leader Opus 4.8 scored 13.4% on Diamond. GPT-5.5: 6.3%. Then Fable 5 arrived. 24 hours later: 29.3% Diamond, 46.3% Main. swyx: "Fable is a different CLASS of model, with beeeeeg model smell." Also: AI Engineer World's Fair, Jun 29–Jul 2, Moscone West SF. Alex is speaking, and the last ~500 tickets are going.

Tim Cook's final keynote, Jun 8. Siri rebuilt as a standalone app with personal and on-screen context. Five Apple Foundation Models on-device. But Max's teardown revealed the truth: AFM Server Pro is Google/Gemini on Nvidia GPUs in Google Cloud (262k context, pcc-agent, slug .language.instruct_server_v2.base_pro). An on-device 20B MoE (1–4B active) gatekeeps what leaves the device. App Intents are mandatory (SiriKit deprecated), MCP goes system-wide, and Xcode 27 goes agentic with Claude + GPT + Gemini.

Streaming speech-to-speech translation: sub-500ms latency, 70+ languages, one Live API call. Preserves tone, pace, and pitch, not just the words. Already in the Translate app. AI Studio at $0.023/min. Google Meet is getting 2,000+ language pairs in preview. Thor shows us how to build a live translator in under 100 lines.

Google's first open text-diffusion model, built on Gemma 4: 26B MoE (3.8B active), 256-token blocks, Apache 2.0. Runs at 1,000+ tokens/sec on a single H100, 18GB VRAM quantized. The quality tax: −12% GPQA and −20% AIME vs autoregressive Gemma 4. "We spent 40 years teaching computers to read left to right and the breakthrough was... don't do that." Sundar posted it himself, which is always a signal.