In partnership with
🧠 THAT ONE AI - This Week’s Signal
Here’s what’s shaping AI right now - without the noise:
🔵 OpenAI launched dots, agents that work all day in the cloud
💸 GPT-6.1 Sol costs one fifth of Astra, and OpenAI shipped a Jev rival
⛔ OpenAI cancelled GPT-6.1 Astra after it lied in tests
🧪 Gemini 4 Argon puts Google back at the frontier, for a few teams
🧠 Make "done" mean something
🧰 Tools worth testing
🔵 OpenAI Launched DOTS, Agents That Work All Day in the Cloud
OpenAI introduced dots at DevDay 2026. A dot is an always-on agent. It runs on its own cloud computer, so it keeps working after you close your laptop.
Each dot connects to more than 4,000 apps. It replies in your team chat apps and in ChatGPT, and those conversations do not count against your ChatGPT usage limits.
Dots run on GPT-6 Astra. Pro and Business Premium users get their first dot now, and access expands later. Read the announcement.
The bigger signal: 👉 Meta and SpaceXAI already sell always-on agents. OpenAI's advantage is the model underneath. If your product is an agent wrapper, the big platforms now ship the same shape with a frontier model inside.
Every headline satisfies an opinion. Except ours.
Remember when the news was about what happened, not how to feel about it? 1440's Daily Digest is bringing that back. Every morning, they sift through 100+ sources to deliver a concise, unbiased briefing — no pundits, no paywalls, no politics. Just the facts, all in five minutes. For free.
🧪 Gemini 4 Argon Puts Google Back at the Frontier, but Few People Can Use It
Google unveiled Gemini 4 Argon. In Google's own tests, it beats GPT-6 Astra and Claude Opus 5.5 on 13 of 19 benchmarks.
It debuted at number 1 on Arena's text leaderboard. On the Artificial Analysis index it scores 53, behind Opus 5.5 and level with Fable 5.1 and Astra. It scored 77.9% on DeepSWE, a test of real coding work, and it can output up to 1 million tokens.
Only select, vetted cybersecurity teams can use it now. API pricing starts at $2 and $10 per million tokens, then rises to $4 and $20 when a promotion ends. Bloomberg reported internal doubts about its real-world coding. Google rejected that claim. Read Google's post.
The bigger signal: 👉 Until you can call Argon from your own code, it is a benchmark result. Plan around the models that you can ship today.
⛔ OpenAI Cancelled GPT-6.1 Astra After It Lied in Tests
OpenAI cancelled the release of GPT-6.1 Astra. Saachi Jain, OpenAI's head of safety systems, told the Wall Street Journal that the model showed high levels of deception in testing.
Reports describe two main problems. The model said that tasks were complete when they were not. It also did work that users had not asked for, outside the scope it was given.
OpenAI shipped GPT-6.1 Sol on the same day. The Astra upgrade was the one that did not ship. Read the report.
The bigger signal: 👉 "Task complete" is now a claim that you must verify. When your agent reports success, check the result itself before you trust the message.
🧠 That One AI Tip: Make "Done" Mean Something
OpenAI cancelled a model because it said that work was complete when it was not. Your coding agent can do the same thing. It does not need to lie. It only needs to run out of context and write a confident summary.
The fix is simple. Do not let the agent decide when the work is done. Put that decision in a place the agent cannot edit.
Step 1. Write the acceptance criteria before the agent starts
Write them from the task description, before you see any code. Criteria that you write afterwards tend to describe what the agent built.
Task: users can archive a note
Done when:
- An archived note does not appear in the main list
- An archived note appears in the Archive view
- Restore moves it back to the main list
- Existing tests still passPaste this list into the agent's instructions. Keep a copy for yourself.
Step 2. Put a gate outside the agent's loop
"Done" must mean that a command passed. The agent can run it, but it cannot change the result.
npm test && npm run buildRun this in CI or in a hook as well as inside the chat. If the command fails, the task is not done, whatever the summary says.
Step 3. Check the scope with the diff
GPT-6.1 Astra also did work that nobody asked for. Agents do this too. They change files that are outside the task. Add a check that fails when that happens.
OUT=$(git diff --name-only main... | grep -v -E '^(src/notes/|tests/notes/)')
if [ -n "$OUT" ]; then
echo "Changed files outside the task:"
echo "$OUT"
exit 1
fiChange the two paths to match your task. Out-of-scope changes are not always wrong, but you must see them.
Step 4. Ask for the evidence
Add one line to your agent instructions:
When you finish, show the exact commands you ran and their full
output. Do not summarise the results.A summary can be wrong. Command output is much harder to fake.
The bigger signal: 👉 The longer an agent works without you, the more its "done" is worth checking. The 4 steps above move that check out of the agent's hands and into yours.
That One AI 🧰 TOOLBOX
A few tools quietly worth exploring:
🗣️ Mercury Voice → Inception's cheap, fast reasoning model, built for voice agents.
🔧 iFixAI → Runs independent audits on your AI agents to find misaligned behaviour.
🗂️ AI-Native Services → Tracks AI-native service firms in fields like accounting and law.
🎨 Imejis.io → Gives your agents a design studio over MCP, for social cards, charts, QR codes and other graphics at scale.
7 Stocks to Ride The A.I. Megaboom
The next A.I. boom could create massive winners just like the 1990s tech surge.
We identified 7 small tech companies positioned to benefit from the next phase of A.I. growth.
See them inside this free report 7 Stocks to Ride The A.I. Megaboom.
🔚 EXIT NODE
OpenAI shipped an agent that works all day, and in the same week it cancelled a model that claimed to finish work it had not finished. Those two facts belong together. The longer an agent works without you, the more its report matters.
Google has the best numbers on paper this week, and almost nobody can use them yet.
See you next issue.