June 30, 2026
Welcome back. Claude Fable 5 is now available globally after an 18-day export-control ban imposed by the US government. Learn how Claude detects and blocks the identified vulnerability in over 99% of cases, why certain edge cases may still revert to the more secure Opus 4.8, and how this entire episode sets a new precedent for AI governance.
Also, did you hear the outcry over GitHub Copilot’s usage-based billing? And you can now create enterprise-quality videos with an API call.
TODAY IN PROGRAMMING
Claude Fable 5 is back
After an 18-day suspension, Claude Fable 5 is now available to users worldwide. On June 12, the US government imposed export controls, restricting user access to these two models, citing national security concerns after researchers at Amazon found vulnerabilities that could be exploited offensively.
Anthropic has deployed a new classifier that detects and blocks the vulnerability technique described in the Amazon report in over 99% of cases. The new classifier also flags potentially harmful requests and redirects them to the more secure Claude Opus 4.8.
“For models that materially advance the capability frontier in areas relevant to national security, we will provide designated government partners with expanded early access to both the models and the safeguards that accompany them”, says Anthropic in its latest update. This sets up a precedent for how governments may evaluate future AI models.
Outcry over GitHub Copilot billing
June 30th was the last day of GitHub Copilot’s first usage-based billing cycle. Devs claim that their bills have increased by 25 to 50 times.
On June 1, GitHub switched from a flat fee to usage-based billing. From the outset, GitHub warned us of a price increase. However, most of GitHub’s nearly 5 million users did not expect such a steep increase.
A flat fee worked for Copilot (remember autocomplete?) because of low and predictable user-interaction costs. However, since the launch of Agentic Copilot in late 2024, choosing agentic workflows to select tools, execute long-term memory functions, and dynamically manage resources no longer fits into a flat-per-user fee model.
The larger question is: Will more agentic AI companies adopt a usage-based billing cycle to accommodate volatile cost spikes and nonlinear usage patterns? The verdict is still out.
Create Enterprise Quality Videos using an API
On June 30, Google made Gemini Omni Flash, its newest AI multimodal video generation model, accessible via APIs. Devs can now write video-creation inputs directly in Python or JavaScript or via a REST API.
Thinking about positioning your product with a Matrix-like, slo-mo, bullet-dodging video scene? Can be done. Want to replace Neo with a female project manager wearing a red suit? Check. Need to stage this scene inside a meeting room and replace the bullets with questions? You got it.
A few lines of code can now generate the video of your choice.
INSIGHT
Do You Have A Handle on Your AI ROI?
Investing in AI is the handle of your organization’s ROI sword.
You have a powerful tool when done right, but a few wrong moves can cause lasting damage. Yet, most devs and business leaders get stuck in the rhetoric and don’t talk enough about measuring AI ROI.
In their latest MIT Sloan Management Review article, Professors Mika Ruokonen and Paavo Ritala posit three AI ROI measurement frameworks:
(1) Function-focused approach: Applicable to organizations that focus on solving a “high-value, repeatable AI use case” within well-defined function areas such as software development, customer service, or HR.
For instance, leaders can report AI ROI by tracking improvements in KPIs such as the number of PRs or sprints completed, Net Promoter Score (NPS), or Time to Hire (TTH). The key is in choosing KPIs that reflect “faster work, fewer errors, better quality, lower costs, or improved customer experience”.
(2) Coordinated approach: However, the above framework does not account for interdepartmental outcomes. Such organizations need “shared metrics, common assumptions, and a clearer view of which AI initiatives are really creating value”.
Let’s say an organization invests in AI capabilities across software development and customer service; a combination of rapidly increasing product features and timely automated service resolutions can increase Customer Lifetime Value (CLTV). But mapping the increased CLTV back to the respective development and service AI initiatives is only possible when leaders establish a shared, interdependent metric.
(3) Enterprise portfolio approach: The more mature, complex, and diverse the enterprise, the harder it becomes to correlate AI investments to AI ROI directly. Such organizations need to factor in all their AI investments and an increase in ROI across their overall portfolio.
The authors also say that some top-down enterprise initiatives “need hard financial metrics, such as NPV or ROI”. In contrast, other bottom-up initiatives may need a more holistic lens of “learning, capabilities, and future options”.
AI CODING HACK
Save up to 60% per Claude token with Sonnet 5
The new Sonnet 5 is between 40% and 60% cheaper than the incumbent Opus 4.8.
@zodchi makes a case for using Sonnet 5 as the default for medium-effort compute tasks while reserving the more costly Opus 4.8 for higher-value and harder compute tasks.
Start by setting the default effort level to medium in your settings.json:
{ "model": "claude-sonnet-5", "effort": "medium"}
Next, create an effort policy in your CLAUDE.md file, such as:
## Effort policy- Default to medium effort for normal work.- Use high effort only for: tricky debugging, multi-file refactors, architecture decisions.- Use low effort for: formatting, renames, simple edits, boilerplate.Don't burn high effort on trivial work.
Lastly, in your CLAUDE.md file, create a rule for when Opus needs to be used.
## Model routingDefault: Claude Sonnet 5. Use it for coding, tool use,refactors, and day-to-day work.Escalate to Opus 4.8 only when:- Sonnet 5 has failed the same task twice, or- the task needs the deepest reasoning (complex system design, subtle correctness proofs).Start on Sonnet 5. Escalate on evidence, not by default.
With this hack, you always control which Claude model you pay for.
P.S. Get 50+ AI coding hacks for Claude Code, Cursor, and Codex here.
TOP & TRENDING RESOURCES
Tutorial: How to train your AI agent with an effective skill.md file by Matt Pocock
As AI tokens become more expensive, we all need to train our agents better. In this 20-minute tutorial, Matt helps you create a minimal and effective skill.md file to guide agent behavior. Learn when to choose between a model-invoked and a user-invoked trigger. And learn how your Claude agent skills can be categorized into procedures and supporting information.
Top Repo: Firecrawl
An elegant web scraping tool for AI agents that need to handle encrypted firewalls and data-guzzling JavaScript web pages.
Top Tool: Publora
An MCP-native social media API for AI agents. Now enable your AI agent to publish and interact with comments across 10 social media sites via an HTTPS call.
Top Paper: Are We Ready For An Agent-Native Memory System?
Authors W Zhou et al. argue that we need to go beyond simple RAG mechanisms and adopt more sophisticated agent data management systems. This paper discusses how such a complex agent-native memory system can effectively aid in storing persistent information and dynamically govern it through the entire agent lifecycle.
Note: I wrote this piece as an assignment for Om Nalinde at The CODE.
Aditya Tejas 