Washington lifts the Fable 5 ban

The frontier-model export saga ended as fast as it began: Commerce lifted the two-week export controls on Anthropic's Fable 5 and Mythos 5, and the model returns worldwide today with a stricter classifier. Anthropic also had a busy launch day of its own with Sonnet 5 and the new Claude Science workbench, while Google shipped cheaper image and video models and China's open-weight pipeline kept churning.

US lifts export controls on Fable 5 and Mythos 5

Commerce Secretary Howard Lutnick lifted the June 12 export controls that had forced Anthropic to pull Fable 5 and Mythos 5 offline after Amazon researchers found a jailbreak that got Fable 5 to flag software flaws and write exploit code. Fable 5 returns worldwide today across Claude.ai, the Claude Platform, Claude Code, and Cowork; Mythos 5 stays limited to roughly 100 approved US organizations. Anthropic shipped a new classifier that blocks the specific technique in over 99% of cases (routing blocked requests to Opus 4.8) at the cost of more false positives on ordinary coding tasks.

Why it matters: There is still no binding process for shipping a frontier model in the US, only improvised export controls used as leverage. Developers get their most capable model back, but with a twitchier safety filter and a precedent that access can vanish for weeks.

Claude Sonnet 5 nearly matches Opus 4.8, but the tokenizer bites

Anthropic released Claude Sonnet 5, its most agentic mid-tier model, claiming performance close to Opus 4.8 at lower prices: 63.2% on SWE-bench Pro (Opus 4.8 is 69.2%), 80.4% on Terminal-Bench 2.1, and a slight edge over Opus on the GDPval knowledge-work benchmark. It ships with a 1M-token context, 128K max output, adaptive thinking on by default, and dropped support for temperature/top_p/top_k. Pricing is $2/$10 per million tokens through August 31, then $3/$15, but Simon Willison notes a new tokenizer produces ~30% more tokens on English text, effectively a stealth price bump.

Why it matters: Sonnet 5 makes near-flagship agentic coding cheaper per token, but the fatter tokenizer plus higher token consumption from more agentic behavior means real bills may not drop as much as the sticker price suggests.

Google ships Nano Banana 2 Lite and opens Gemini Omni Flash video to the API

Google released Nano Banana 2 Lite (Gemini 3.1 Flash Lite Image), which generates 1K images in about four seconds for $0.034 each, positioned as the drop-in replacement for the original Nano Banana. Alongside it, Gemini Omni Flash reaches developers via the Gemini API and AI Studio, generating and conversationally editing up to 10-second video clips at $0.10 per second (matching Veo 3.1 Fast). Google recommends chaining the two: draft images fast, then animate them. Caveats are real: the Lite model struggles with small text and infographic accuracy, and Omni Flash can't yet do scene extension, audio references, or reliable character consistency across cuts.

Why it matters: Cheap, fast image generation plus API-accessible video editing lowers the cost floor for media pipelines, but the quality asterisks mean this is a drafting tool, not a finishing one.

Claude Science bets on workflow, not a new model, for research

Anthropic launched Claude Science, a standalone workbench it ranks alongside Claude Code and Cowork, aimed at computational biology and drug discovery. It runs the same Opus 4.8 already available to everyone (no special model), connecting 60+ databases and toolkits for genomics, structural biology, and cheminformatics, and taps Nvidia's BioNeMo toolkit with Evo 2, Boltz-2, and OpenFold3. A project-manager agent spawns sub-agents, and a separate verification agent checks citations and calculations, though it is still the same model checking itself. It runs locally on macOS/Linux and connects to HPC clusters via SSH so data stays in the lab.

Why it matters: This is the vertical-workflow playbook applied to science: Anthropic going wide with broad subscription access while OpenAI (GPT-Rosalind) gates enterprise and Google leans on owned models like AlphaFold. The distribution strategy, not the model, is the differentiator.

Huawei open-sources OpenPangu 2.0 Flash, a 92B sparse MoE

Huawei released OpenPangu 2.0 Flash, a 92B-total / 6B-active mixture-of-experts model with a 512K context, shipping weights, inference code, and training ops. A larger Pro variant (505B total, 18B active) is slated for July, with more open-source components promised later this year.

Why it matters: Another capable Chinese open-weight release with real training artifacts, not just weights. The steady drumbeat of these launches is exactly the competitive pressure cited as a reason to loosen US model controls.

OpenAI paper leaks a three-model GPT-5.6 Pro lineup

A GPT-5.6 generation split into Sol, Terra, and Luna was announced in late June, but a new OpenAI genomics-benchmark paper is the first to list three parallel Pro variants: Sol Pro, Terra Pro, and Luna Pro. Sol Pro tops all 60 tested models at 31.5% pass rate versus 28.7% for standard Sol and 16.0% for Claude Opus 4.8. Notably, the Pro boost is largest for weaker tiers, and OpenAI omitted token-usage figures for the Pro runs that it reported for every other model.

Why it matters: If it ships, Pro stops being one top tier and becomes a speed/throughput/reasoning menu, changing how developers pick a model per task. The missing token accounting is a tell about compute cost.

AI Engineer World's Fair: everything is a loop now

Day 2 of AIEWF converged on one word, loops, with swyx's opening talk 'Loopcraft' and a main-stage track on 'software factories' where the pitch is that engineers stop writing code and instead build the system that builds the product. OpenAI's Codex team, Microsoft Foundry, Warp, Factory, and OpenClaw's Peter Steinberger all framed agent orchestration as stacked loops with deterministic gates. The other theme was the rise of Forward Deployed Engineers (aka agent engineers) who do most of their work at the orchestration layer, not in the models.

Why it matters: The industry narrative is shifting from prompting to orchestration: cron jobs, retry gates, and review loops around cheap agents. Whether 'software factory' is a real discipline or rebranded rote work is the open question.

Also worth a look

Browse previous days →