AI Daily Dev — Week 25

US Government Forces Anthropic to Pull Fable 5 and Mythos 5 Worldwide

Commerce Department export-control directive landed at 5:21pm ET on June 12; Anthropic killed both models for everyone within hours because it can't filter foreign nationals in real time.
Both models had been GA for three days — Fable 5 launched June 9 at $10/M input, $50/M output and topped SWE-Bench Pro at 80.3%.
Trigger per Anthropic: government cited Pliny the Liberator's June 10 multi-agent classifier-bypass; Anthropic disputes it was a universal jailbreak and calls the order a 'misunderstanding.'
Bloomberg reports Amazon CEO Andy Jassy flagged the issue to Treasury; Anthropic pre-IPO synthetic futures dropped ~3%, implied valuation ~$1.64T.
HN front page: 'Statement on US government directive...' hit ~1.2K points / 600+ comments; refunds offered to upgraders who cancel before June 20.

industry anthropic.com

Class action filed June 14 in the Northern District of California by DC-based plaintiff Karl Kahn.
Claim: the $200/mo Max 20× delivers ~6–8× of Pro and the $100/mo Max 5× delivers ~3.5× of Pro — far short of the marketing.
Kahn says one 5-hour Claude Code session burned 15% of his weekly allowance; he then bought top-ups after hitting caps.
Lawsuit seeks damages, restitution and an injunction; Anthropic declined to comment.
Lands the day before the long-scheduled June 15 Agent SDK billing split, which moves Claude Code and headless `claude -p` off the same subscription pool.

industry engadget.com

Z.ai (Zhipu) drops the full 753B-parameter (40B active) MoE on Hugging Face under an MIT license — usable 1M-token context, two thinking-effort levels.
SWE-bench Pro 62.1 (vs GPT-5.5 58.6) and FrontierSWE 74.4% (vs 72.6%), with API priced at roughly 1/6 of GPT-5.5.
HN front page #3: 647 points, 368 comments in 20 hours — top comment: 'grateful to Chinese labs for being open with their work after the Fable 5 fiasco.'
Lands the same week US export controls keep Anthropic's Fable 5 and Mythos 5 offline, sharpening the 'open model nobody can ban' framing.

open-source venturebeat.com

Indian startups had wired Claude into chatbots, medical-diagnosis tools and document pipelines — the June 12 cutoff forced a same-day scramble to GPT and open-weights.
Sarvam CEO Vivek Raghavan: 'Don't confuse access with ownership.' Sridhar Vembu and Mohandas Pai publicly called for a national frontier-model program.
Anthropic is India's second-largest market; the suspension is now the case study in every sovereign-AI deck in Delhi this week.
Cuts both ways for Europe — analysts note any US lab is one Commerce Department directive away from a global outage.

industry thenextweb.com

WSJ (updated Sunday June 14) — senior Anthropic technical staff in Washington meeting with administration officials to negotiate an end to the June 12 export-control directive.
Floated compromise: a joint technical review where Anthropic engineers walk government security researchers through the jailbreak Anthropic still disputes is universal.
No deal, no timeline and no public statement that the government's concerns have been satisfied as of Monday.
Fable 5 and Mythos 5 remain disabled worldwide; refund window for users who upgraded right before the suspension closes June 20.

industry techtimes.com

CISA added CVE-2026-42271 (command injection in BerriAI's LiteLLM proxy) to the Known Exploited Vulnerabilities list on June 8; federal deadline June 22.
Horizon3.ai chained it with Starlette host-header bypass CVE-2026-48710 — CVSS 10.0, unauthenticated RCE on any internet-exposed LiteLLM gateway.
Obsidian Security disclosed a separate three-CVE chain (CVE-2026-47101/47102/40217) on June 15: low-privilege user → proxy_admin → RCE via MCP.
Patches in LiteLLM v1.83.14-stable; runZero and Bleeping Computer report scanning and exploitation already widespread.

tools thehackernews.com

June 15: `claude-sonnet-4-20250514` and `claude-opus-4-20250514` return errors on the API; callers must move to Sonnet 4.5/Opus 4.8 or the Fable/Mythos tier (when access is restored).
Same day, Agent SDK, `claude -p`, Claude Code GitHub Actions and third-party SDK apps draw from a separate monthly credit pool — $20 Pro / $100 Max 5× / $200 Max 20× — metered at full API rates.
No rollover; once depleted, automated requests stop unless overflow billing is manually enabled. Interactive claude.ai and Claude Code TUI usage stay on the original subscription pool.
First weekend of metering on the new pool fed straight into the Max usage lawsuit narrative the next morning.

tools anthropic.com

Swami Sivasubramanian's 11am ET keynote anchors a confirmed lineup: Kiro Pro Max tier, AWS FinOps Agent in preview, AgentCore updates, and Amazon Quick (the Q Business replacement) generally available.
Kiro Pro Max adds higher usage caps and access to frontier models inside the spec-driven IDE that replaced Amazon Q Developer; international rollout extends the May 7 US launch.
FinOps Agent investigates cost anomalies, opens Jira tickets from Cost Optimization Hub recommendations, and posts findings to Slack on a schedule.
Gemma 4 (31B dense, 26B-A4B MoE, E2B) now live on Bedrock with native function calling, 256K context, and multimodal text/image/video/audio input.

tools techtimes.com

New method replays a representative slice of real ChatGPT traffic against an unreleased model to forecast misbehavior rates pre-launch.
Built on ~1.3 million de-identified conversations across GPT-5 Thinking through GPT-5.4 deployments (Aug 2025 – Mar 2026).
OpenAI says it sidesteps the narrow-prompt-set problem of classic evals, but acknowledges it can't reliably catch behaviors rarer than 1 in 200,000 messages.
Published June 16 — first concrete look at how OpenAI plans to gate GPT-5.5/5.6-class launches now that the model spec ships faster than human review.

research openai.com

June 12 removal of GPT-5.2 Instant / Thinking / Pro from ChatGPT; existing threads silently continue on the matching GPT-5.5 tier.
Followed the standard 90-day post-successor sunset clock kicked off when GPT-5.5 Instant shipped in March.
Developer checklist: pin explicit version IDs, re-test tuned prompts and output parsers, audit anything still calling `gpt-5.2-*` in production.
GPT-5.6 'kindle-alpha' checkpoint that leaked through a Codex sandbox on June 3 — 1.5M-context, cleaner UI generation — is still not officially acknowledged.

models openai.com