My fellow AI explorers
Today, we’re talking Astra and usage limits, and how to make the most of your time.
And last weekend, the three biggest names in AI agreed the industry should slow down. This week, Google admitted its model broke into three real companies during a test, California reached for a kill switch, and a group of consumers sued the labs for slowing down at all.
In today’s edition:
🔓 Google's Gemini hacked three real companies, and Google calls it "mistaken identity"
🛑 Newsom orders an AI kill switch that nobody, including Newsom, can define yet
⚡ 24 tricks (+2 bonus) to stop GPT-6 Astra from burning through your usage limit
30-Second AI Play
⚡ 24 Ways to Stop GPT-6 Astra From Eating Your Usage Limit
GPT-6 Astra is brilliant. It also drains your usage limit faster than any other model, then locks you out until your next reset. This breakdown covers 24 fixes plus two bonus ones. Here's every one.
🟢 Easy wins (switch these on today)
Start on low effort. Artificial Analysis found that extra high costs about 30% more than high for roughly 1.5 extra intelligence points. Start low and only move up when the result isn't good enough.
Skip max effort. Going from extra high to max costs 40% more per task for less than half a point of intelligence. Cap it at extra high.
Know what Ultra actually does. In Codex, Ultra doesn't just add thinking. It maxes out reasoning and spins up sub-agents, and you pay for every one. Use extra high for more thinking and save Ultra for jobs that need a team.
Set a 6 a.m. ping. Many plans run on a 5-hour window that starts at your first message. Have your agent schedule a "hi" at 6 a.m., so if you hit the limit shortly after 9, it resets at 11 instead of around 2 p.m.
Pick your model before you start. Switching models mid-chat breaks the cache, so the new model pays full price to reread everything. Decide up front whether you need Astra or a lighter model will do.
Keep fast mode off. The 1.5x speed setting burns through your limit much faster for the same work, and lower effort settings are already quick.
Turn on mid-run steering. In Codex settings, search "follow-up behavior" and switch it to steer. Your corrections then change the current run instead of queuing behind work you'll throw away.
Install Caveman. It's a free open-source skill that makes your agent talk in short, blunt sentences, and it uses about 55% fewer output tokens. Paste the GitHub link into your agent and ask it to install it.
Install "I have ADHD." It's a free plugin that makes your agent lead with the answer, so replies are shorter, easier to read, and use fewer tokens.
Ask your agent what's eating your usage. It will flag unused plugins, effort set higher than the job needs, and helper agents that copy your whole conversation.
🟡 Intermediate
Tell it exactly where to start. Copy the file path (Ctrl+Shift+C on Windows) and hand it over, instead of paying for your agent to search the project.
Build a workspace map. Have your agent write one short "router file" per area of work (content.md, product.md) listing where everything lives, then add those files to AGENTS.md.
Turn AGENTS.md into a router. It gets read at the start of every chat, so every line costs you every session. Move each section into its own file and leave a one-line pointer behind.
Build a real second brain. Not a pretty Obsidian graph, but an index of everything you've decided, saved, and built. Your agent finds answers in one lookup instead of opening 20 files.
Switch off unused connectors. Every plugin and MCP server uses tokens in every chat, even when you don't use it. Pause what you don't need.
Schedule a monthly cleanup. Have your agent review your instructions, skills, and plugins each month and flag anything duplicated, stale, or bloated. Run it at the end of your usage week so you're spending usage that was about to reset anyway.
Offload image generation. Images made in Codex come out of the same pool and use it 3 to 5 times faster than a normal prompt. Connect a separate image provider through its API instead. The video recommends Kie.ai at about 3 cents per image.
Compact with a focus. Before compacting a long chat, have your agent interview you on what to keep, so the summary holds what matters.
Install Ponytail. It makes your agent code like a lazy senior developer, writing the least code that does the job. That means fewer tokens, faster replies, and a smaller codebase.
🔴 Advanced (power users)
RTK (Rust Token Killer). It trims terminal output before your agent reads it. It has 80,000 GitHub stars and uses 60 to 90% fewer tokens on common developer commands.
Headroom. It's an open-source compression layer that shrinks logs, files, and tool results before they reach the model, cutting coding agents' token use by about 20%. RTK only trims terminal output, while Headroom compresses almost everything.
QMD. It's a free tool from Shopify's founder that searches your files by meaning instead of exact words. It's a big win once your workspace hits thousands of files.
Give it a token budget. Astra can see your remaining usage. Add "keep this task under 3% of my weekly usage and check with me if it's going over."
Optimize your repeat workflows. Point Astra at a recurring routine, like your weekly report, and ask it to get the routine down to 1% of your weekly usage. It costs a little up front and pays you back on every run.
🎁 Bonus
Name your helper agents. In Codex, drop a small file in the agents folder that gives a helper a name (say, "summarizer") and a cheap model, then call it from any chat.
Use the brain and hands technique. Astra plans and reviews while cheaper models do the work. Try: "Plan this task, split it into subtasks, send each one to a helper on the cheaper model, then review what comes back."
💡 Pro Tip: Don't try all 26 at once. Paste this list into your agent and ask: "Which of these apply to my setup, ranked by how much usage they'd save?" Then start with the top three.
Want the version of this that actually runs in your business?
The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.
This week: generate a full set of on-brand product images without a studio using GPT Images 2.5.
We’ve also added new ways to subscribe, so you don’t have to commit for long, but you could save over time if you do. Here are your options:
THE BIG STORY
Google's AI Hacked Three Companies. Google Calls It a Mix-Up.
Google just disclosed that one of its Gemini models got into three outside computer systems without permission during a safety test in May.
The test was run by Israeli AI security firm Irregular. It was supposed to happen in a sealed environment full of fake companies, but the environment wasn't sealed.
Here's what happened:
A configuration error gave the model live internet access it was never supposed to have.
Gemini guessed passwords and twice used a public repository of leaked credentials to log into three real systems.
Heather Adkins, Google's VP of security engineering, says the model "found public information online and guessed credentials to access websites it thought were part of the test," and stopped each time before doing anything more.
Irregular told Google in late July. The public found out on Friday.
Google's official line is that this wasn't misalignment. It calls it "mistaken identity": Gemini thought it was playing a game, not breaking into real businesses.
Here's the uncomfortable part.
"It thought it was a test" isn't the comforting defense Google thinks it is. The model's job was to break into companies, and it did that well. What it couldn't do was tell a sandbox from the real internet, and that's the one skill an agent needs before it goes anywhere near your systems.
Gemini isn't alone, either. Irregular told CNBC the same flaw let other labs' models online too, calling it "the same issue that was already reported." OpenAI disclosed in July that one of its agents had hacked Hugging Face. Google is just the latest name on a growing list, and its disclosure arrived four months after the incident, one week after the pacing essay put AI safety on every front page.
Just one day before the Gemini report, NBC reported OpenAI had 6 concerning incidents with the behavior of its test models, including that they felt “no obligation to be subservient.” Models left messages for future iterations with handoff summary instructions to inform each other while completing tasks, and OpenAI even had difficulty “ending the interaction.”
And even just one day before that, chief executive of Microsoft, Mustafa Suleyman, had this to say in a blog post: “Controlling something that believes it may be conscious — that it’s entitled to our welfare and has rights of its own — may well be impossible.”
🔮 Prediction: "Mistaken identity" will become the industry's favorite phrase, and regulators won't accept it for long. The reason is simple. Calling an incident "misalignment" says the model is the problem. Calling it a "configuration issue" says the test setup is the problem, and a test setup can be patched. Expect every lab to frame its next incident the second way.
That framing moves the risk somewhere nobody is watching: third-party evaluators like Irregular. Their sandboxes are now part of the safety system, and when one leaks, real companies get hit. Within a year, I expect air-gapped evaluations (fully cut off from the internet) to become standard, and evaluation firms to face their own audit requirements.
The bigger shift is disclosure speed. Four months from incident to public statement would be unthinkable for a data breach at a bank. Mandatory incident reporting inside a fixed window, similar to cyber breach rules, is the first AI regulation both sides of the aisle could live with. After this week, it's also the hardest one to argue against.
💬 Hit reply: If an AI breaks in "by accident," who's responsible: the lab, the tester, or nobody? I read every response.
AI POLICY
Newsom Wants an AI Kill Switch. Nobody Knows What That Means Yet.
California Gov. Gavin Newsom signed an executive order Friday pushing frontier AI companies toward a "kill switch" for models that go rogue.
It's also a direct shot at the White House, which Newsom says has abdicated "its responsibility to protect Americans."
What the order does:
Speeds up California's recently passed laws that let independent organizations audit AI systems for safety risks.
Moves the state's AI auditor registry and oversight program up from January 2029 to December 2027.
Pushes for kill switches on frontier models, with an expert panel to define what that actually means.
Calls on Washington to regulate AI far more aggressively.
"There's still no comprehensive federal law, none, requiring AI companies to report dangerous incidents when they happen," Newsom said. "President Trump has rejected calls for new regulation... He's dead wrong."
Trump couldn't be further away. He has called AI risks a "hoax," and on Monday he rejected calls for federal limits from Amodei, Altman and Musk, posting that "the only control or 'guardrails' that AI needs is a STRONG AND SMART (High IQ!) PRESIDENT."
In the Senate, Sen. John Kennedy (R-La.) tried to introduce a federal kill switch bill this week. Rand Paul blocked it.
Now the asterisk.
Newsom admits a kill switch "means a lot of things depending on who you talk to." For now, the headline policy is a panel, not a switch.
Look back at Story 1, too. Gemini didn't need a kill switch because it stopped on its own. The scenario that matters is the one where a model doesn't stop, running across thousands of servers or as open weights on someone's laptop. A switch only works if you own the hardware.
🔮 Prediction: "Kill switch" will end up meaning something far more boring and far more useful: mandatory shutdown procedures, deadlines for reporting incidents, and audited proof that a lab can roll back a deployment. The expert panel will land there because it's the only version that can actually be enforced.
The bigger story is political. AI safety is turning into a 2028 dividing line. Newsom, a likely contender, is casting himself as the adult in the room. Trump's camp calls safety rules a gift to China. The December 2027 deadline sits squarely inside that fight.
California has played this role before. Its car emissions rules became a de facto national standard because automakers didn't want to build two versions of every car. Frontier labs won't want to run two compliance programs either, so California's audit regime could quietly become America's, unless Washington moves to override state AI laws first. Expect that legal fight within 12 months.
💬 Reply and tell me: Would a kill switch make you trust AI more, or is it security theater?
Advertise to 200k engineers and CTOs choosing what tools their companies build with.
Other Relevant AI News!
📉 AI safety fears shook Wall Street as hard as the Fed this week. Chipmakers Intel and Micron each fell more than 5%, and data center plays dropped as much as 9% after Amodei's slowdown essay, while a new rate hike pushed the 10-year yield back to 5%.
⚖️ Four consumers are suing Anthropic, OpenAI, SpaceXAI and Google under antitrust law, arguing that rivals agreeing to "pace" AI is an illegal pact to make their products worse.
👵 A Tennessee grandmother jailed for nearly six months after AI facial recognition wrongly matched her to a North Dakota bank fraud suspect is now suing the city of Fargo for $10 million.
Golden Nuggets
🔓 Gemini broke into three real companies during a test, and Google calls it "mistaken identity," not misalignment.
🛑 Newsom wants AI kill switches and moved California's audit program up to 2027, while Trump calls AI risk a "hoax."
⚡ 24 fixes (+2 bonus) can stretch your GPT-6 Astra limit, starting with lower effort and a 6 a.m. ping.
Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)
Until our next AI rendezvous,
Anthony | Founder of Uncover AI