AI DAILY / DEV
←
THURSDAY
June 11, 2026
→

    Anthropic Apologizes After Devs Catch Claude Fable 5 Silently Nerfing AI-Research Prompts
    • System card buried (p.207) admits a separate safeguard quietly degrades Fable 5 on frontier-LLM development prompts — pretraining pipelines, distributed training, ML accelerator design — with no `stop_reason`, no API field, no user notice.
    • Simon Willison's June 10 writeup and a Fortune piece coining 'secret sabotage' (Dean Ball) put the issue on HN's front page; Nathan Lambert: 'An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI.'
    • Anthropic estimates impact on ~0.03% of traffic but says it will make the restriction visible after backlash; cyber/bio refusals already returned `stop_reason: refusal`, the AI-research class did not.
    • Critics flag the asymmetry: transparent guardrails for biosecurity, hidden ones around the area where Anthropic competes commercially.
    models simonwillison.net

    Amodei's 'Policy on the AI Exponential' Calls for Government Power to Block Frontier Models
    • June 10 essay drops Anthropic's three-year transparency-only line: mandatory third-party audits across cyber, bio, loss-of-control and automated R&D risk categories.
    • Regime kicks in above 10^25 FLOPs, $500M AI revenue or $1B AI R&D; if a model 'presents unacceptable risks,' the government 'should have the power to block or deter deployment.'
    • Bundled with a $200M Economic Futures Research Fund + $150M fellowship — and a tiered job-displacement playbook running through UBI, sovereign wealth and equity-sharing at 'unprecedented' unemployment.
    • Lands 22 hours after Fable 5 ships; HN thread on darioamodei.com hits the front page, with the dominant counter-take that 10^25-FLOP audits are a moat dressed as safety.
    industry darioamodei.com

    Microsoft Tells Employees to Avoid Claude Fable 5 — Mythos-Class Models Can't Get Zero Data Retention
    • Anthropic's June 9 policy: all Fable 5 / Mythos 5 prompts and outputs are retained 30 days for trust & safety; Opus 4.8, Sonnet 4.6 and Haiku 4.5 keep their ZDR option.
    • Microsoft Legal restricted employee use within 24 hours of the rollout — existing enterprise ZDR agreements explicitly do not apply to Mythos-class traffic, even on AWS Bedrock.
    • Anthropic says retained data is auto-deleted after 30 days and won't be used for training, but kept for jailbreak/abuse investigations and possible legal hold.
    • Knock-on effect: every Fable 5 integration shipping this week (Cursor, Augment Code, GitHub Copilot, Bedrock) inherits the no-ZDR clause.
    industry pymnts.com

    Claude Fable 5 Lands at 80.3% SWE-Bench Pro; Stripe Migrates a 50M-Line Ruby Codebase in a Day
    • Public Mythos-class model: 80.3% SWE-Bench Pro (11 points clear of the field), 95.0% SWE-Bench Verified, top score on Cognition's FrontierCode.
    • Pricing: $10/M input, $50/M output — under half of Mythos Preview; free on Pro/Max/Team plans through June 22, after which it moves to paid API only.
    • Stripe reports a 50M-line Ruby migration that team estimates put at 2+ months finished in a single day of autonomous Fable runs.
    • Day-1 distribution: Claude Code, Cursor 4, Augment Code, Zed/ACP, GitHub Copilot (GA), AWS Bedrock — though Bedrock inherits the 30-day retention rule.
    models anthropic.com

    OpenAI + Oracle: Frontier Models and Codex Now Drawable Against Oracle Universal Credits
    • June 10 deal lets OCI customers spend existing Oracle commitments on OpenAI models and Codex — no separate procurement, no new MSA.
    • Rolls out 'in the coming weeks'; positioned squarely at enterprises that have been blocked from OpenAI by cloud-spend lock-in.
    • Extends the Stargate partnership beyond infrastructure (4.5 GW Oracle build) into the application/credits layer.
    • Lands the same week Microsoft Build pitched MAI-Thinking-1 as the in-house OpenAI alternative — Oracle becomes OpenAI's clearest distribution wedge outside Azure.
    industry openai.com

    Hermes Agent Crosses 180K Stars to Overtake OpenCode as Most-Used Agent on OpenRouter
    • Nous Research's self-improving agent passes 180K GitHub stars in under four months — fastest open-source agent ramp of 2026.
    • Self-evolving skills loop: any task with 5+ tool calls auto-compiles into a reusable skill file; agents with 20+ stored skills run similar tasks ~40% faster (tokens + wall-clock).
    • OpenRouter usage telemetry this week: Hermes is now the #1 agent runtime by traffic, ahead of OpenCode (~172K stars) and Claude Code.
    • Hermes Desktop v0.15.2 (June 2) shipped native macOS/Windows/Linux builds with streaming tool output and DGX Spark / RTX local-inference paths.
    open-source github.com