Claude Fable 5 Goes Public, Cohere's Lean Coding Model, and Apple's EU Siri Standoff — AI Digest for June 10, 2026
Anthropic releases Claude Fable 5 — its first Mythos-class model available to everyone — with safety classifiers that route high-risk queries to Opus 4.8. Cohere drops North Mini Code, a 30B MoE / 3B active-parameter open-source coding model that benchmarks above models far larger. GitHub Copilot CLI adds repo-stored custom agent profiles for teams. Apple confirms Siri AI won't launch in Europe due to an ongoing DMA dispute.
Anthropic crossed a line it spent months saying it wouldn't cross: it released a Mythos-class model to the general public. That's the big story today, alongside Cohere dropping a surprisingly competitive open-source coding model, GitHub adding team-scoped custom agents to its CLI, and Apple's Siri AI standoff with EU regulators getting messier.
Anthropic releases Claude Fable 5 — its first Mythos-class model anyone can use
For the past few months, Anthropic has maintained that its Mythos-class models were too capable to release publicly. The reasoning was specific: these models are good enough at cybersecurity and biology to provide meaningful "uplift" — assistance that makes attacks meaningfully easier than what someone could accomplish otherwise.
What changed is the safeguards, not the capabilities. Claude Fable 5 is the same underlying model as Claude Mythos 5, with classifiers layered on top that catch queries in four areas — cybersecurity, biology/chemistry, distillation attempts, and jailbreaks — and route them to Claude Opus 4.8 instead. In internal testing, about 95% of sessions never trigger a fallback at all. 1
The model's practical capabilities are substantial. On Cognition's FrontierCode benchmark — which tests whether models can pass difficult coding tasks while meeting production-code quality standards — Fable 5 scores highest among frontier models. Stripe reported compressing two months of team work on a 50-million-line Ruby codebase migration into a single day. On Hex's analytics benchmark of complex long-running analytical tasks, Fable 5 was the first model to break 90%. 2
コンテンツカードを読み込んでいます…
A few details worth knowing if you're deciding whether to use it in an API integration:
- Pricing: $10 / million input tokens, $50 / million output tokens — double Opus 4.8 and about half the cost of Mythos Preview
- Subscription plans: Free on Pro, Max, Team, and seat-based Enterprise through June 22; requires usage credits after June 23, with a stated goal to restore it as a standard plan model once capacity allows
- New data retention policy: All Fable 5 and Mythos 5 traffic is retained for 30 days regardless of prior enterprise agreements. Anthropic says it won't train models on this data, only use it to detect novel jailbreaks and reduce false positives
- Model ID:
claude-fable-5via the Claude API
Claude Mythos 5 — same model, some safeguards lifted — remains restricted to Glasswing partners (focused on cyberdefense) and a small forthcoming biology research program.
The honest caveat: benchmarks for frontier models are produced by the company releasing the model or by partners with early access. Independent evaluations take weeks. The qualitative reports from Stripe, Cursor, and GitHub all sound similar to what every company says at every major model launch. The FrontierCode numbers are third-party (Cognition), which matters.
Cohere's North Mini Code: 3B active parameters, 80% on SWE-Bench
Cohere released North Mini Code, a 30B sparse mixture-of-experts model with 3B active parameters, under Apache 2.0. It's their first model explicitly built for agentic software engineering rather than general chat or RAG. 3
The architecture: 128 total experts, 8 activated per token, 128K context window, trained with a "long-to-longer" SFT cascade. The practical upshot of the 3B active parameter count is that it runs significantly cheaper and faster than models with more active compute — which matters if you're running it as a sub-agent in a pipeline.
The benchmarks are genuinely interesting for a model this small:
| Benchmark | Score |
|---|---|
| SWE-Bench Verified (pass@10 after SFT) | 80.2% |
| Terminal-Bench v2 (pass@10 after SFT) | 55.1% |
| mini-SWE-Agent (pass@1) | 61.0% |
| Artificial Analysis Coding Index | 33.4 |
On the Artificial Analysis Coding Index, North Mini Code outscores models with far more active parameters — including Mistral Small 4 (119B total / 6B active) and NVIDIA's Nemotron 3 Super (120B total / 12B active). Whether that holds in production-quality agentic tasks is worth testing. Cohere also trained the model specifically to generalize across different agent harnesses (SWE-Agent, OpenCode, Terminus 2), so it shouldn't require much scaffolding adjustment if you're already using one of those.
Available now on Hugging Face in both BF16 and FP8 formats, and via the Cohere API.
コンテンツカードを読み込んでいます…
GitHub Copilot CLI gets custom agent profiles
GitHub published a guide this week on custom agents in GitHub Copilot CLI — a feature that's been quietly available since Build but hasn't gotten much documentation until now. 4
The mechanic: you define an agent as a Markdown file (ending in
.agent.md) stored in your repo's .github/agents/ directory. The file has YAML frontmatter specifying the agent's role, accessible tools, and guardrails — then invoke it in the CLI with /agent. Because it's a file in the repo, it versions with the code, can be reviewed in pull requests, and stays consistent across team members.The main use cases GitHub describes — security audits, infrastructure compliance checks, release notes drafting, incident first-response — are all tasks that teams currently handle with ad-hoc shell scripts or repeated manual prompts. The value here is repeatability and reviewability: the same agent file runs in CI, in your terminal, and from an IDE, without anyone having to re-explain context each time.
This is worth paying attention to if your team is already using Copilot and you're looking for a way to standardize repetitive agentic tasks without standing up a separate orchestration layer.
コンテンツカードを読み込んでいます…
Apple's Siri AI won't launch in Europe, and both sides are pointing fingers
Apple has confirmed that Siri AI — the Gemini-powered assistant unveiled at WWDC last week — won't launch in EU countries on iOS 27, and possibly not for considerably longer. The stated reason: compliance with the Digital Markets Act's interoperability requirements would force Apple to give third-party AI companies the same deep system access Siri AI uses, which Apple says creates unacceptable privacy and security risks. 5
The EU's position is that nothing in the DMA prevents Apple from launching Siri AI — the regulation doesn't block new product launches. Apple proposed a "Trusted System Agent" intermediary model as a compromise, requiring 18 months to implement. The Commission rejected it along with Apple's other proposals. Apple's SVP Greg Joswiak said the Commission hasn't engaged substantively with any of Apple's suggestions; the Commission says Apple hasn't proposed anything that actually complies.
The technical argument underneath this: Siri AI needs to read across Photos, Messages, Calendar, and third-party apps to do anything useful. DMA interoperability rules require Apple to give competitors equivalent access — meaning Google, Anthropic, and others could request the same hooks. Apple argues that's not a narrow carve-out situation; it's essentially opening the operating system. External legal experts cited in The Verge's coverage are skeptical of Apple's framing, noting that Apple already permits Siri AI to do the exact cross-app data access it claims is dangerous when competitors ask for it.
China also won't get Siri AI, for separate regulatory reasons — that was mentioned in a one-line footnote.
The practical effect for EU-based iOS developers: any product workflows that depend on deep Siri AI integration won't be testable or deployable in Europe for the foreseeable future. Apple has previously used DMA compliance as the reason to block AirPods Live Translation, iPhone Mirroring, and some Maps features from EU devices.
Brief: Lovable hits $500M ARR
Vibe-coding platform Lovable crossed $500 million in annualized revenue, up from $400M in February, and says users are now creating one million new projects per week — 50 million total since launch in late 2023. 6 The primary user base, per Lovable's own survey, is non-technical: founders, designers, and salespeople building websites, CRMs, and internal tools. The practical question Lovable hasn't had to answer yet is whether those projects keep running — software that isn't maintained tends to break quietly, and Lovable is only approaching its third year of existence.
参考ソース
- 1Claude Fable 5 and Claude Mythos 5 — Anthropic
- 2Anthropic's Claude Fable 5 is a version of Mythos the public can access today — TechCrunch
- 3Introducing North Mini Code: Cohere's First Model For Developers — Hugging Face
- 4From one-off prompts to workflows: How to use custom agents in GitHub Copilot CLI — GitHub Blog
- 5Apple's game of chicken with EU over Siri AI — The Verge
- 6Lovable says it has hit $500M in annualized revenue — TechCrunch
このコンテンツについて、さらに観点や背景を補足しましょう。