Presented by
Good morning, AI enthusiasts. The US Commerce Department just lifted export controls on Anthropic's Fable 5 and Mythos 5, less than three weeks after the order that pulled Fable 5 from international users. Anthropic trained a new safety classifier to satisfy the government, and Fable 5 returns globally today.
The resolution sets the template for how Washington handles frontier model security, but it also raises a harder question: has the US just built a permanent approval gate for every major model release?
In today's recap:
Anthropic's Fable 5 returns as US lifts export controls
Claude Sonnet 5 narrows the gap with Opus 4.8
Access Claude Sonnet 5 and pick your model tier
Claude Science turns Claude into a research workbench
4 new AI tools, prompts, and more
ANTHROPIC
Fable 5 export controls lifted, returns globally
Anthropic
Recaply: Anthropic just confirmed the US government lifted export controls on Claude Fable 5 and Mythos 5. Fable 5 now returns globally across Claude.ai, Claude Code, and the Claude Platform.
Key details:
The government imposed the controls on June 12, after Amazon researchers found a way to bypass Fable 5's safeguards; Anthropic could not verify user nationality in real time, so it suspended access for everyone.
Anthropic built a new safety classifier that blocks the bypass in over 99% of cases, rerouting blocked requests to Opus 4.8, while testing showed weaker models like GPT-5.5 and Kimi K2.7 could find the same vulnerabilities too.
Commerce Secretary Howard Lutnick said Anthropic no longer needs a license, according to a letter seen by Reuters, and in return Anthropic agreed to detect security risks early and report any malicious activity to the government.
Fable 5 goes live globally on July 1 for Pro, Max, Team, and some Enterprise plans, counting toward up to 50% of weekly usage limits through July 7 before switching to usage credits.
Why it matters: This was the first real test of how the US handles frontier model security, and it sets the template going forward. Anthropic agreed to deeper government collaboration, plus a shared jailbreak severity framework with Amazon, Microsoft, and Google. But OpenAI's Sam Altman already pushed back on the government picking customers, and the new classifier's false positives will frustrate coders. Geopolitical guardrails on AI models are here to stay.
PRESENTED BY
What’s next is almost here.
On July 16th at 1PM ET, beehiiv is going live with a look at the future of publishing, audience growth, and digital business.
What started as a newsletter platform has evolved into something much bigger: a place where creators and brands can grow, monetize, and own their audiences without stitching together half the internet to make it work.
The next chapter starts live at the Summer Release Event.
Join us to see what’s coming next.
ANTHROPIC
Claude Sonnet 5 launches, narrows Opus gap
Anthropic
Recaply: Anthropic just shipped Claude Sonnet 5, its most agentic Sonnet model yet, built to plan tasks, use tools like browsers and terminals, and run autonomously at a level close to Opus 4.8, for a lot less money.
Key details:
Sonnet 5 plans, uses tools, and runs multi-step coding tasks on its own, and testers say it finishes jobs that older Sonnet models gave up on while checking its own work without being asked.
It launches at $2 per million input tokens and $10 per million output tokens through August 31, then moves to $3 and $15, compared to Opus 4.8's $5 and $25 per million.
Independent Artificial Analysis benchmarks show Sonnet 5 costs more and scores lower than Opus 4.8, though Anthropic says it closes the agentic gap at a lower price, even as it showed higher misaligned behavior rates than Opus in internal safety tests.
Sonnet 5 is live now as the default model for Free and Pro plans, alongside Max, Team, Enterprise, Claude Code, and the Claude Platform under the name claude-sonnet-5.
Why it matters: Sonnet 5 is a clear step up from Sonnet 4.6, but the real story is how Anthropic is positioning it. This new tier does not beat Opus 4.8 on independent benchmarks, and Fable 5 already sends its coding work to Opus 4.8 instead. For builders, the pitch is simple: use Sonnet 5 for daily agentic work, and escalate to Opus 4.8 only when you need maximum power. The gap between the two tiers is now the whole point of the lineup.
GUIDES
Access Claude Sonnet 5 and pick it over Opus 4.8
Recaply: In this tutorial, you will learn how to get Claude Sonnet 5, compare its cost to Opus 4.8, and pick the right effort level for your coding work.
Step-by-step:
Open Claude.ai, where Sonnet 5 is now the default model on Free and Pro plans, or call it through the Claude API or Claude Code using the model name "claude-sonnet-5"
Compare the price: Sonnet 5 costs $2 per million input tokens and $10 per million output tokens through August 31, while Opus 4.8 costs $5 and $25, and Sonnet 5 rises to $3 and $15 after that date
Set your effort level in Claude Code or the API, starting at medium for daily work, since Anthropic's own charts show medium effort covers most tasks at a fraction of what Opus costs
Use Sonnet 5 for multi-step jobs like coding, tool use, and debugging, where testers say it finishes tasks that older Sonnet models used to give up on halfway through
Save Opus 4.8 for the hardest tasks only, pushing Sonnet 5 to high effort first since it can match Opus on some evaluations there, and only escalate if it still falls short
Pro tip: Push Sonnet 5 to high effort before you reach for Opus 4.8. On agentic search and computer use tests, high-effort Sonnet 5 already matches Opus, so test your own task first.
TOGETHER WITH CATCH
AI Executive Assistant for Normies
Catch is a smart, proactive AI admin agent that handles your scheduling, inbox, emails, follow-ups, and day-to-day admin across Gmail, Outlook, Slack, WhatsApp, and phone.
No setup. No learning curve.
From swamped to sorted, instantly.
ANTHROPIC
Claude Science turns Claude into a research workbench
Anthropic
Recaply: Anthropic just unveiled Claude Science, an AI workbench built for scientists. It brings together the tools researchers already use, like PubMed, Jupyter, and R, and produces figures and manuscripts you can trace back to the code that made them.
Key details:
A coordinating agent taps into more than 60 skills built for genomics, proteomics, structural biology, and chemistry, spinning up specialist agents for specific jobs while a reviewer agent checks citations and catches errors along the way.
At the Allen Institute, researcher Jerome Lecoq built a pipeline of 20 custom skills that has already produced about 10 reviews, some over 100 pages long, in work that used to take his team up to two years.
At UCSF, Stephen Francis said Claude Science cut his glioma research time down to about a tenth of what it used to take, and his team checked the results themselves and confirmed they held up.
Claude Science is in beta now on macOS and Linux for Pro, Max, Team, and Enterprise plans, and Anthropic is funding up to 50 science projects with grants up to $30,000 each.
Why it matters: This is Anthropic building agents for a specific field, not just another chatbot. The workbench comes pre-loaded for life sciences, with native views for protein structures and genome tracks, a bet on locking in how scientists work day to day. The reproducible, auditable angle also solves AI science's biggest trust problem: can you actually check the work? If this holds up past beta, expect Anthropic to build the same kind of tool for other research fields next.
TOOLS
Trending AI Tools
💨 Claude Sonnet 5 - Anthropic's most agentic Sonnet model, approaching Opus 4.8 performance at lower introductory pricing
⚙️ Ornith-1.0 - DeepReinforce AI's open-source self-improving agentic coding models, spanning 9B to 397B and matching Claude Opus 4.7 on SWE-Bench
💨 Acti - The world's first agentic keyboard, letting you type an intent anywhere and long-press the Acti Bar to invoke AI actions across apps
🔊 Seed Audio 1.0 - BytePlus's TTS model that generates voice, music, and sound effects in a single pass
NEWS
What Matters in AI Right Now?
OpenAI engineers reportedly cut inference costs in half for some of its models, The Information reports, a real win as labs race to shrink the cost of running frontier models.
Google released Nano Banana 2 Lite, its fastest and cheapest Gemini image model yet, alongside Gemini Omni Flash for video generation and editing, both live now in Google AI Studio, the Gemini API, and Google Flow.
Google also added a Short Video Overview feature to NotebookLM, joining the audio overviews and mind maps the tool already offers.
Godot will no longer accept AI-authored code, with maintainers saying they cannot trust heavy AI users to understand their own code well enough to fix it, a sharp pushback from a major open-source project.
Bloome launched a human-AI teammate system where Claude, ChatGPT, and DeepSeek work together in one shared context, routing work through drafting, challenge, and refinement stages until it reaches a final answer.
Anthropic launched Claude Desktop for Linux, bringing its desktop app to a third operating system alongside macOS and Windows for anyone who wants Claude outside the browser.
Nvidia's AI chip sales in China are stalling, AP News reports, with local chipmakers like Huawei taking the lead in a market Nvidia once dominated as export restrictions tighten.
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
