- Towards AGI
- Posts
- Google's New AI Proposed Six Fixes. Engineers Shipped Two.
Google's New AI Proposed Six Fixes. Engineers Shipped Two.
the real cost of overnight ai is the morning review
Shen Pandi
October 15, 2026
Today, we're diving into:
Gen AI: Google's Gemini 4 Argon works overnight, but humans still decide what ships
Hot Tea: 44% of ransomware victims saw attackers exploit their own AI
OpenAI: GPT-4o and GPT-4.1 leave the API on January 14
Google's New AI Proposed Six Fixes. Engineers Shipped Two.
Every frontier lab now sells the same promise: give the model a hard task, go home, and come back to finished work. Google just put real numbers behind that promise, and the most useful number is the one it did not lead with.
On October 13, Google released Gemini 4 Argon, the first model in its Gemini 4 family, built for long-running coding, finance, legal, cybersecurity and science work. It comes with a 1 million token limit shared across input and output, and a new protocol for very long outputs.
Is Your Data Ready For Agents That Work All Night?
Find out whether your metadata, lineage and access rules can keep up before an agent works through them unsupervised.
What Happened In The Overnight Run
Google's own developer team gave Argon an unattended, 11-hour task: speed up its Agent Development Kit, a Python codebase of about 205,000 lines across roughly 800 files. The agent had one job, open a single pull request with before-and-after numbers.
The first, unguided attempt found real bottlenecks. It also changed a public JSON schema, missed edge cases and ran only 216 nearby tests out of roughly 17,000.
The agent was fast. It just did not know the house rules.
The second attempt was guided by rule files stored in the repository. This time Argon ran all 16,892 unit tests and produced six optimizations across 32 files. Reviewers shipped two and rejected four, including a cache that could return stale data and a fix that only helped synthetic tests.
The two that shipped were worth it: roughly 40% less session overhead in long sessions and about 21% faster imports. Google's team summed up the division of labor neatly.
"The agent explores, the maintainer subtracts." Google Developers Blog
The Price Tag Is Not The Cost
Argon launches at $2 per million input tokens and $10 per million output tokens, rising to $4 and $20 after November 30. Artificial Analysis found it averaged $1.99 per benchmark task against $3.26 for GPT-6 Astra, but it used about 62,000 output tokens per task against Astra's 27,000. At standard pricing, Argon costs roughly 1.2 times as much per task.
In other words, the cheap model thinks out loud a lot. And every token it writes overnight is a token someone has to judge in the morning.
The Real Bottleneck Is The Morning Review
Picture your best engineer arriving at 9 a.m. to six overnight pull requests from an agent. Each one passes its tests. Four of them are wrong in ways only someone who knows the system would notice.
This is not a model capability problem. It is a review capacity problem.
The wrong reaction is to measure overnight agents by how much they produce. Output is now nearly free. The better alternative is to measure cost per accepted change: model spend plus reviewer hours, divided by what actually ships.
You should also treat the house rules as infrastructure. Google's second run worked because conventions were written down where the agent could read them. Most enterprises keep those rules in people's heads.
Write your conventions down. Put coding standards, schemas and do-not-touch areas into rule files your agents read on every run.
Track acceptance rate. Count proposed changes against merged changes for every agent and every model.
Make full test runs non-negotiable. An agent that runs 216 tests instead of 17,000 has not finished the job.
Budget reviewer time first. Size overnight agent work to what your senior people can properly review the next day.
The organizations that win with overnight agents will not be the ones generating the most code. They will be the ones whose reviewers can tell the two good changes from the four bad ones before lunch.
Ransomware Gangs Are Now Recruiting Your AI Agents
Security teams have spent two years worrying that attackers would use AI. A new report says they already do, and that many of them are using yours.
On October 14, CBS News reported findings from Semperis's 2026 Ransomware Risk Report, a survey of 1,200 IT and security professionals across nine countries and eight industries. 84% of ransomware victims said attackers exploited AI in the attack. 44% said attackers exploited the victim's own AI systems.
How Your Own Agent Becomes The Way In
AI now helps attackers find weaknesses faster, tailor phishing to specific employees and adjust tactics when defenders react. It also lowers the skill needed to attempt intrusions that once took real expertise.
The more worrying shift is inside the network. Agents that can use several applications and make decisions for users are a new route deeper into your systems. Compromise the agent, or its credentials, and you inherit everything it can touch.
"The agent then becomes a ticking time bomb." Courtney Guss, Semperis
The Weakness Is Older Than AI
87% of respondents saw attacks aimed at their identity systems or the infrastructure that controls network access. Chris Inglis, the former White House National Cyber Director and a Semperis board member, argued that the underlying weakness is often the same one that existed before AI: poor control of who can access what.
This is not an AI security problem. It is an identity and permissions problem with a faster engine.
What You Should Do Now
The wrong reaction is to pull agents out of production. Inglis himself recommends using AI defensively rather than abandoning it. The better alternative is to give every agent the same scrutiny you give a privileged employee.
Give every agent its own identity. No shared service accounts, no borrowed human credentials.
Scope permissions to the task. An agent that summarizes invoices does not need write access to payments.
Watch for permission creep. Alert when an agent requests new access or touches systems outside its normal pattern.
Harden identity infrastructure first. If attackers own your directory, they own every agent in it.
Securing agents starts with trusted enterprise context. DataManagement.AI helps organizations create that foundation by unifying metadata, lineage, governance, business glossaries and knowledge assets into a single intelligence layer, so you can see which data each agent is allowed to reach and trace what it actually touched.
The enterprises that stay out of next year's ransomware report will not be the ones with the fewest agents. They will be the ones that know exactly what every agent is allowed to do.
OpenAI Just Put An Expiry Date On GPT-4o
Somewhere in your company, a workflow you have not thought about in a year is still calling GPT-4o. It now has three months to live.
On October 14, OpenAI announced the shutdown of GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini in its API. They will be removed on January 14, 2027, and OpenAI recommends moving to gpt-6-luna. Fine-tuned versions of these models follow the same schedule.
Why This Is Bigger Than A Version Bump
These models sit underneath a huge amount of enterprise software: support bots, document extraction, classification pipelines, internal copilots. Many were tuned prompt by prompt until the outputs looked right.
A newer model is usually smarter. It is not necessarily the same. Output formats shift, refusals change, answers get longer, and the prompt that worked perfectly starts to drift.
The model is the easy part to replace. The behavior you built around it is not.
The Fine-Tune Trap
If you trained a custom version of GPT-4o or GPT-4.1, it retires with its base model. That means rebuilding training data, retraining and revalidating before mid-January, in a quarter that includes the holidays.
This is not a model upgrade. It is a regression-testing problem with a deadline.
How To Get Ahead Of January 14
Inventory every call. Search code, configs and vendor tools for the four retiring models, including fine-tunes.
Build an evaluation set now. Collect real inputs with known good outputs so you can compare old and new models side by side.
Compare cost per task, not price per token. A model that writes twice as many tokens can cost more even at a lower rate.
Design for the next switch. Route model calls through one layer so the next retirement is a configuration change.
Know Your Inference (KYI) is becoming a core enterprise capability. By understanding how every inference affects cost, latency, accuracy and governance, you can test a replacement model against real workloads, see what it really costs per completed task, and switch with evidence instead of hope.
The enterprises that handle model retirements well will not be the ones that upgrade fastest. They will be the ones that can prove the new model behaves before the old one disappears.
Your opinion matters!
Hope you loved reading our piece of newsletter as much as we had fun writing it.
Share your experience and feedback with us below 'cause we take your critique very critically.
How's your experience? |
Thank you for reading
-Shen & Towards AGI team