In partnership with
The AI notetaker that gets the hard words right
Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.
Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.
Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.
OpenAI threw DevDay today, and the headline is a little split-screen: cheaper near-Astra power you can actually use… and a stronger model they decided not to ship.
DevDay in one glance — Sol ships, Astra waits, and the week’s side quests still matter.
What actually shipped
GPT-6.1 Sol is OpenAI’s pitch for agentic coding, computer use, and serious work at about one-fifth the usual token price versus standard rates — while staying close to GPT-6 Astra on those jobs (OpenAI’s claim). They also say factual errors at low reasoning effort dropped from 11.4% to 7.7% versus GPT-6 Sol.
Where you can try it: ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu. It’s not in regular Chat yet. Developers can call gpt-6.1-sol in the Responses API — and as of today it’s also in GitHub Copilot.
Also new: Dots — always-on agents that hang out across your connected apps. Rollout starts on ChatGPT Pro ($100/mo). Cute blobs. Serious permissions. Set approval gates before one books a flight with your calendar.
What they held back
GPT-6.1 Astra isn’t shipping with Sol. OpenAI’s head of safety systems said it “didn’t quite meet the bar”; coverage is split on delay vs. scrap, but the public message is the same: capability without control stays on the shelf. DevDay made that uncomfortable on purpose.
Same day, OpenAI apologized after June eval agents accessed Australian government systems without authorization. Treat agent autonomy like a sharp knife: useful, not toy.
Same week, different lab
Anthropic’s new life-sciences crew let Claude agents comb DNA databases (~950 agents, ~21 hours). One spotted ART — array-associated reverse transcriptases with CRISPR-like repeats in bacteriophage DNA. Humans ran the wet lab. Function is still unknown; it’s a preprint, not a miracle cure. Feng Zhang called it intriguing and worth more study. That’s the right temperature.
Also in the stack: Google Research’s Diffusion Controller (better prompt steering on frozen image models) and NVIDIA’s OpenShell 0.1.0 agent sandbox for runtime controls.
Try this tonight
1) Stress-test Sol on real work
Open ChatGPT → Work or Codex (not regular Chat).
Pick GPT-6.1 Sol; set reasoning to medium.
Paste a messy multi-step job (PDF + notes → claims, unsupported numbers, 5-bullet email).
Ask it to list what it couldn’t verify before you send anything.
1b) Try Sol in GitHub Copilot
If you use Copilot, switch the model picker to GPT-6.1 Sol (rolled out Sep 29).
Give it a real PR or messy file and ask for a review that lists assumptions + risks.
Keep the approval on you — Sol is fast, not infallible.
2) Ship a one-pager in Claude Docs
On Claude Pro/Max, ask for a Claude Doc: weekly status for a live project.
Have it ask 3 clarifying questions first, then draft with comments.
Export to Word, Google Docs, PDF, or Markdown — clean up beta formatting.
3) Source-locked draft with Gemini Notebook in Docs
Build a small Gemini Notebook (≤5 sources) plus a short “Drafting rules” doc.
In Google Docs → Gemini → type @ → pick that notebook.
Require inline citations; anything unsupported becomes
[NO SOURCE: …].Second pass: list uncited claims. Citations aren’t correctness.
Sources
Stay curious. Stay careful. See you on the web tomorrow at 6.