Sunday, June 14, 2026

TLDR

US issues first-ever export control on a commercial AI model (Jun 13) — Anthropic disables Claude Fable 5 and Mythos 5 worldwide after a Commerce/BIS directive bars all foreign nationals; Amazon-flagged jailbreakA form of prompt injection that makes a model ignore its safety rules.OWASP ↗ cited
Google DiffusionGemma released open-weightA model whose trained weights are published for anyone to download and run.NTIA ↗ (Jun 10) — 26B MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ text-diffusion model (3.8B active), Apache 2.0, up to 4× faster generation (1000+ tok/s on one H100)
Figure hits 1 robot/hour at BotQ (late May, still ramping) — 24× throughput gain in <120 days; 350+ Figure 03 humanoids delivered
SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ leaderboard irony (Jun 13) — the two models now blocked by the export ban (Mythos 5, Fable 5) hold the top spots on SWE-bench Verified
Final reminder: Claude Sonnet 4 + Opus 4 retire TOMORROW, June 15, 9am PT — update model strings before then or API calls will error

Things your engineering team should look at or do this week.

Audit pipelines for any Fable 5 / Mythos 5 dependency — both are now disabled globally. As of Jun 13 the US directive forced Anthropic to suspend access to claude-fable-5 and claude-mythos-5 for all customers (the only way to comply with a foreign-national ban that included staff). If you wired Fable 5 into a workflow during its free Pro/Team window (released Jun 9), it will fail now. Fall back to claude-opus-4-8 / claude-opus-4-7. (Anthropic statement)
Final call: migrate off claude-sonnet-4-20250514 and claude-opus-4-20250514 before 9am PT tomorrow (Jun 15). After retirement these strings return errors. Minimum swap: claude-sonnet-4-5 / claude-opus-4-5; better: claude-sonnet-4-6 or claude-opus-4-8. One-line change for most call sites — test before deploy. (Anthropic release notes)

Google DiffusionGemma (Jun 10) Open Source Infra — Google DeepMind's first open-weight text-diffusion model: a 26B Mixture-of-Experts (25.2B total / 3.8B active, 8 of 128 experts) that generates 256-token blocks in parallel instead of token-by-token, hitting up to 4× faster generation (700+ tok/s on an RTX 5090, 1000+ on a single H100) and fitting in ~18GB VRAM at NVFP4. Apache 2.0, on Hugging Face, Kaggle, and Vertex AI Model Garden. Trade-off: scores below standard Gemma 4 on MMLU and coding — Google explicitly labels it experimental and recommends Gemma 4 for production. Worth a look for latency-bound local inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗. (Google blog | DeepMind | Hugging Face)

US orders Anthropic to disable Fable 5 and Mythos 5 — first export control on a commercial AI model (Jun 13) Trending Policy Security — A Commerce Department directive (signed by Secretary Howard Lutnick, drafted with the Bureau of Industry and Security) ordered Anthropic to block all foreign nationals — inside or outside the US, including its own foreign-born employees — from accessing Claude Fable 5 and Mythos 5, citing national-security concerns. Because the ban reached staff on US soil, Anthropic says it had no way to comply except to suspend both models for everyone worldwide. The trigger: Amazon security research reportedly surfaced a jailbreak of Fable 5's cyber/bio guardrails. Anthropic disputes the severity, arguing the same vulnerabilities are reachable with other public models and that a narrow jailbreak shouldn't recall a model deployed to hundreds of millions. Analysts note the precedent could chill Anthropic's pending IPO. (Anthropic | Bloomberg | Fortune | CNN)

Figure reaches 1 humanoid/hour at BotQ (late May, ramp ongoing) Robotics/IoT Trending — Figure says its BotQ facility transitioned from prototype to production lines for all critical modules and now hits the one-robot-per-hour cycle time targeted for Figure 03 — a 24× throughput gain in under 120 days, with 350+ third-gen units delivered. A concrete data point that "physical AI" is moving from demos to volume manufacturing. (Figure | Robotics & Automation News)

SWE-bench Verified snapshot (Jun 13) — Claude Mythos 5 leads at 95.5%, Claude Fable 5 at 95%, Claude Opus 4.8 at 88.6%. The catch: the two leaders are exactly the models the Jun 13 export ban just pulled from the market — so for anyone shipping today, Opus 4.8 is effectively the top available model. On the harder, contamination-resistant SWE-bench Pro (Scale AI, 1,865 tasks / 41 repos), every model drops ~19–26 points; the best standardized public score sits around 59% (GPT-5.4 xHigh). Keep benchmarking agents on Pro, not Verified. (SWE-bench Pro leaderboard | Scale SWE-bench Pro)