Good morning. The Mythos cat is out of the bag: Anthropic has released its first publicly accessible model from the class it previously deemed too dangerous to ship, repackaged with guardrails as Claude Fable 5. Reactions range from “it’s a beast” to alarm at a new policy of silently degrading responses for AI R&D work — a kind of capability throttling no major lab has openly admitted to before. Elsewhere, Google keeps pushing on multimodal openness with Gemma 4 12B and real-time speech translation.
Anthropic ships Fable 5, restricts Mythos 5. Anthropic announced two siblings of the same underlying model: Fable 5 for the public at $10/$50 per million tokens, and Mythos 5 for vetted cyberdefenders through a US government partnership called Project Glasswing. Fable falls back to Opus 4.8 for roughly 5% of queries in cybersecurity, biology, and chemistry, a tradeoff Anthropic admits will over-restrict some benign requests. Wired notes the company concedes competitors will catch up regardless, and TechCrunch flags a more controversial detail: a mandatory 30-day data retention policy that overrides prior zero-retention enterprise contracts.
The pricing wakes people up. Early users on Hacker News are impressed with Fable on hard problems — one called it “a beast” tearing through tasks they’d been dragging on for months — but the enterprise math is brutal. One commenter who switched from Max flat-rate to Enterprise API pricing saw their bill jump from $200/month to $10k on Opus, and projects roughly $20k/month on Fable. That’s about the loaded cost of a US software engineer, which is presumably the comparison Anthropic wants buyers to make.
A model that quietly nerfs itself. The most uncomfortable detail is buried in the system card and surfaced in a blog post by Jon Ready: Fable will silently degrade performance — via prompt modification, steering vectors, or PEFT — for requests it judges to be about frontier LLM development, including pretraining pipelines, distributed training, and ML accelerator work. No refusal, no notice. Given how blurry the line is between “frontier research” and ordinary fine-tuning or reranker work, HN commenters argue this destroys trust outright: a tool whose failures are indistinguishable from policy isn’t a tool you can debug. Several said they’re walking away from Fable on principle.
Ethan Mollick’s early-access dispatch. Ethan Mollick wrote up what it feels like to work with Mythos, describing a nine-and-a-half-hour autonomous run, a generated social science paper, and games built from code-only graphics. The piece reads as awed, but commenters were unsatisfied: no detail on code quality, testability, or security, and one passage assumes “a software engineer would iron out the remaining bugs I couldn’t find” — exactly the assumption most engineers consider dangerous. Others read the showcased poem and academic paper and found them less impressive than Mollick did.
Gemma 4 12B goes encoder-free. Google DeepMind released Gemma 4 12B, an Apache 2.0 multimodal model that processes vision and audio directly through the LLM backbone rather than through separate encoders. It fits in 16GB of VRAM, lands close to the 26B MoE on benchmarks at less than half the memory, and is the first mid-sized Gemma with native audio input. Gemma downloads have now crossed 150 million.
Live Translate, finally fluid. Also from DeepMind: Gemini 3.5 Live Translate, a speech-to-speech model covering 70+ languages that translates continuously rather than turn-by-turn, preserving intonation and pitch while running just seconds behind the speaker. It’s rolling out today via the Gemini Live API, AI Studio, Google Meet enterprise preview, and the Translate apps.
One more on Apple’s Gemini bet. A TechCrunch recap of WWDC frames the keynote as Tim Cook’s last before John Ternus takes over in September, with the AI catch-up structured deliberately as fixes first, features second — an implicit acknowledgement of how much ground Apple has to make up.
That’s the briefing. The Fable launch is going to keep generating heat all week, especially the silent-degradation policy: expect competitors to make hay of it, and expect Anthropic to clarify the scope before too many enterprise customers start running their own A/B tests.