AI Navigate
アップデート最新記事AI大全カオスマップ
2026 · 06 · 10 · Wed

Updates for 6/10

Anthropic released Claude Fable 5, the first public Mythos-class model, across API, app, and Claude Code. Google shipped 70+ language live voice translation.

A · Theme of the day

Mythos goes public: Claude Fable 5 ships for everyone

Anthropic's first public Mythos-class model, Fable 5, is now live everywhere.

Claude Fable 5 launches — Mythos-class open to everyone

Claude (Anthropic)Claude (Anthropic)
What changed

June 9, 2026 (PT): Anthropic released Claude Fable 5, the first GA Mythos-class model said to beat Opus 4.8, plus Mythos 5 for Glasswing partners.

Compared to before

Mythos-class models were limited to vetted Project Glasswing cyber-defense partners, with public access promised but undated. On June 9 (PT), Fable 5 opened to everyone.

Why it matters

Claude Pro and Max users get stronger capabilities at no extra cost. The caveat: about 5% of requests silently fall back to Opus 4.8, and cyber, bio, and chem domains are blocked.

Source: anthropic.com　·　→ View on chaos map
Read in depth →

Fable 5 is now the default in the Claude app and API

ClaudeClaude
What changed

Claude Fable 5, released June 9, 2026 as the first public Mythos-class model, is now the default in the Claude app and API; ~5% falls back to Opus 4.8.

Compared to before

Through last week, the Claude app and API defaulted to Opus 4.8, and the claude-fable-5 model ID had already surfaced in the SDK. As of June 9 (PT), Fable 5 is the new flagship.

Why it matters

API builders can specify claude-fable-5 or ride the default upgrade. Output style and safety behavior can shift, so test on a subset of traffic before full production rollout.

Source: claude.ai　·　→ View on chaos map
Read in depth →

Claude Code runs Fable 5 — long autonomous sessions get an upgrade

Claude CodeClaude Code
What changed

Claude Code now supports Fable 5, the Mythos-class flagship released June 9, 2026; Anthropic cites a Stripe migration cut from ~2 months to one day.

Compared to before

Claude Code ran on Opus 4.8, and developers running long autonomous tasks were hitting the model's ceiling. As of June 9, Fable 5 is available inside Claude Code.

Why it matters

The 'hand it a spec' parallel-coding workflow gets a real upgrade. Treat the Stripe claim with skepticism, but multi-hour autonomous sessions should feel the difference.

Source: docs.anthropic.com　·　→ View on chaos map
Read in depth →
B · Theme of the day

Google ships real-time voice translation in 70+ languages

Gemini 3.5 Live Translate landed in Meet, the Gemini app, and a developer API.

Real-time voice interpretation lands in Meet and the Gemini app

GeminiGemini
What changed

Launched Gemini 3.5 Live Translate: streaming speech-to-speech translation for 70+ languages that preserves tone, across Meet, Translate, and the Live API.

Compared to before

At Google I/O in May, voice-to-voice live translation was only a future direction. Meet and Translate had translated captions, but nothing preserved the speaker's tone.

Why it matters

Meet meetings can switch from captions to voice interpretation this week. '70 languages' is about coverage, not uniform quality — test your language pair before high-stakes use.

Source: gemini.google.com　·　→ View on chaos map
Read in depth →

Gemini Live API brings 70-language voice translation to your own app

Gemini (Google)Gemini (Google)
What changed

Released Gemini 3.5 Live Translate via the Gemini Live API: streaming speech-to-speech for 70+ languages, preserving tone, also shipping in Meet and Translate.

Compared to before

Embedding live voice translation meant stitching recognition, translation, and synthesis, each handoff adding 1-3 seconds of lag. The Live API now does it in one streaming call.

Why it matters

Conferencing, language-learning, and voice-support builders can replace a three-stage pipeline with one API call. Test the tone-preservation claim on your language pairs first.

Source: deepmind.google　·　→ View on chaos map
Read in depth →
C · Theme of the day

Cohere open-sources a coding MoE that runs on one H100

Cohere's North Mini Code 1.0, a 30B coding MoE for a single H100, is Apache 2.0.

North Mini Code 1.0: a 30B coding MoE free to self-host

Command (Cohere)Command (Cohere)
What changed

Open-sourced North Mini Code 1.0: a 30B MoE coding model, ~3B active parameters per token, running on one H100; weights on Hugging Face (fp8 included).

Compared to before

Self-hosting a commercially capable coding model meant eight or more H100s. North Mini Code 1.0, Cohere's first developer-focused model, fits on a single H100 via MoE.

Why it matters

Teams that can't send code to external APIs now have a tractable self-hosted option. Independent benchmarks are sparse, so test on your own tasks first.

Source: cohere.com　·　→ View on chaos map
Read in depth →
Archive

Past updates

A daily archive of changes actually applied to the site.