Good morning, and welcome to today’s Techzip.

At the center of today’s issue is Astra: OpenAI says the model can find previously unknown security flaws, raising new questions about how powerful AI should be released. Anthropic is making AI systems that work on their own cheaper, while Google’s next model may close the coding gap.

Meanwhile, AI infrastructure spending could reach $32 trillion by 2050, smaller models are catching up fast, and governments are racing to secure their place in the AI era from cybersecurity to bioweapons and sovereign AI

OpenAI says Astra can find previously unknown vulnerabilities and build exploit paths across hardened systems without a person guiding every step. The company has therefore classified Astra at the Critical threshold—the first model it has placed at that level and applied stronger controls.

The evidence OpenAI published goes beyond a single impressive benchmark. Astra scored 100% on ExploitBench, was then tested on an internal set of 20 newer high-severity vulnerabilities, and discovered two zero-days during evaluation. These results reflect a configuration with Daybreak Blue access rather than the default production configuration, and OpenAI says it is disclosing the vulnerabilities to maintainers.

The important shift is not only what the model can do, but the fact that a capability threshold is becoming part of the release decision. OpenAI says Astra refused 91.5% of cyber-jailbreak requests, compared with 59% for GPT-5.6 Sol, and has added monitoring that can stop unauthorized actions. Access to its advanced cyber capabilities will initially be limited to a small tester group before expanding through Daybreak. Once an agent can find and exploit vulnerabilities on its own, safety controls have to be part of the release decision from the start.

Anthropic says Fable 5.1 cuts cache-read pricing by 75%, translating into estimated savings of about 25% for typical workloads and up to roughly 45% for highly agentic ones. The model costs $10 per million input tokens, $50 per million output tokens and $0.25 per million cached tokens, and is available through the Claude Platform, AWS, Google Cloud and Microsoft Foundry. In one real-world example, Simon Willison said Fable created and animated his best SVG pelican yet for a reported $3.30 using Max thinking.

Anthropic positions Fable for work that runs for hours or days: clearing a Cowork backlog, picking up tasks from Slack, operating web interfaces, running unattended as a managed agent, writing tests and checking its own output with vision. The company also cites Cognition saying Fable 5.1 could replace some Opus 5 traffic in Devin at a lower cost per task, while Square says it was more token-efficient than Opus 5 in a simulated 30-day business-operations test.

For agents, the same context is often read repeatedly across many steps. Cache pricing can therefore determine whether a long-running workflow is affordable enough to run continuously. Fable 5.1 still includes cyber and biology safeguards, and flagged queries may be routed to other models; default data retention is 30 days for safety monitoring, while some eligible enterprise customers can use zero data retention. The competition is moving from “which model is smarter?” to “which model produces more useful work per dollar?”

The Wall Street Journal reports that internal tests show a significant improvement in Gemini 3.8 Flash's coding ability. The model is reportedly known internally as “Skimaki” and could be released as soon as this week. That remains an internal-test report, not a public benchmark or a launch Google has confirmed.

The report says Google is evaluating the model in its internal coding tools, where engineers reportedly preferred Gemini 3.8 Flash to some Anthropic Opus models. If accurate, Google is trying to close the gap through both model quality and a faster release cadence for a commercially important workload.

If released as described, Gemini 3.8 Flash would show how quickly the frontier can move when capability improvements are paired with shipping speed. For now, the things to watch are the actual release, pricing and results on verifiable coding tasks—not a premature conclusion that Gemini has already won.

With a Cash App Card, teens can take their first steps toward independence with a secure debit card. They’ll learn how to spend, save, and manage money, all with your guidance and oversight to help them get started. 


1. Learn to spend responsibly A debit card gives them a safe way to practice managing money under your supervision. It gives you the opportunity to teach them how to make smart spending choices.

2. Start saving for their goals Setting goals can help them see how saving a little at a time can help them reach their short-term and long-term goals.

3. Manage their own money Whether they get paid with direct deposit or use Cash App to get allowance or gifts, they get real experience with money.

Get started with Cash App

Cash App is a financial services platform, not a bank. Banking services provided by Cash App’s  bank partner(s). Prepaid debit cards issued by Sutton Bank, Member FDIC. Cash App Visa®  Debit Flex Cards issued by Sutton Bank, Member FDIC, and The Bancorp Bank, N.A., pursuant  to a license from Visa U.S.A. Inc. See terms and conditions for the Sutton prepaid card (https://cash.app/legal/us/en-us/card-agreement), Sutton debit flex card (https://cash.app/legal/us/en-us/debit-flex-card-agreement-sutton), and Bancorp debit flex card (https:// cash.app/legal/us/en-us/debit-flex-card-agreement-bancorp).

An agent can perform impressively on a few short tasks and still degrade when it has to remember history, handle changing conditions and make repeated decisions over time.

Qwen's E-Commerce Bench puts that problem into a 365-day simulation in which an agent runs several online stores, negotiates with suppliers and manages orders, returns and cash flow.

The practical adaptation is straightforward: do not give an agent one task and judge only the final answer. Turn a repeated process into a multi-round test with a clear starting state—data, budget, goals and action limits.

You can start with five steps:

A strong final result is not enough if the agent keeps repeating the same mistake. And an agent that learns but cannot turn that learning into better action is not ready for a long-running process.

A small test like this helps you see whether the real problem is the model or the way the work is structured.

The Verge essay looks at agent societies and the accountability gap as increasingly autonomous systems coordinate with one another. It offers a useful governance lens: responsibility cannot be assigned only to individual model responses.

Giovanni separates bounded from unbounded tasks and argues that AI research, software engineering and trading could create reflexive sources of token demand. The figures are estimates, but the framework adds a useful question to compute forecasts: is AI spending creating more demand for AI?

Fireside Alpha quotes Sam Altman saying he is confident in OpenAI's committed compute but skeptical of large neocloud projects without clear revenue or buyers. The useful distinction is between aggregate demand and capacity with real customers, financing and a utilization plan.