Is Gemini Back?
GPT-6’s 217-year code crack, OpenAI’s new work mode, and Claude Code’s modding era
In this SunBrief
Build Claude Code Expertise and Prepare for Anthropic Certification
In Claude Code Bootcamp Cohort 2, you’ll build practical Claude Code expertise through 10 new hands-on projects covering CLAUDE.md, reusable Skills, subagents, hooks, guardrails, Best-of-N workflows, testing, security, and production readiness.
Get downloadable Claude Code Skills to reuse after the workshop, plus a free full-length mock test to help you assess and prepare for Claude Certified Foundations, benchmark your knowledge, and prepare for Anthropic certification.
Learn Claude Code. Build practical workflows. Prepare for certification.
Sunday, October 18 | Starting at $189.99
Join Cohort 2 & Get Certification-Ready
01 / Model watch
Gemini 4 takes on ChatGPT and Claude
Google’s Gemini 4 Argon targets coding, research and cybersecurity work. The introductory price catches my attention, but limited access makes it hard to recommend switching.
- API rates start at $2 per million input tokens and $10 per million output tokens, before doubling.
- Arena ranks Argon first for text but tenth for agents, behind GPT-6 Astra and Claude Opus 5.5.
- Access starts with trusted cyber defenders and testers. Paid API customers and Ultra subscribers have no firm date.
Why it matters: I’d test Argon before planning a switch. A lower token price doesn’t prove it can do your work better.
Read more ↗What matters most when choosing between Gemini, ChatGPT, and Claude?
Market Snapshot
| Company | Price ($) | 5D (%) | YTD (%) | ||
|---|---|---|---|---|---|
| 340.35 | 0.81 | 7.94 | |||
| 333.69 | -1.99 | 23.13 | ||
| 728.08 | -3.02 | 11.94 | ||
| 517.53 | 2.39 | 9.43 | ||
| 251.52 | 1.95 | 11.05 | ||
| 233.95 | 1.77 | 23.88 |
02 / Research watch
GPT-6 cracks a code after 217 years
Carter Church says GPT-6 Astra helped decode a letter to one of Napoleon’s generals in six hours of model time. But this wasn’t AI working alone.
- Astra transcribed 1,300 cipher symbols, revealing troop positions ahead of the 1809 war with Austria.
- The work used a historian’s partial key, with Church also spending several evenings on the project.
- Church published the key and checking scripts. Five symbols remain uncertain but reportedly don’t change the meaning.
Why it matters: I see help for historians, not a replacement. Faster answers are useful when someone else can check the evidence.
Read more ↗Where could AI have the biggest impact on historical research?
03 / DevDay watch
OpenAI DevDay 2026: ChatGPT gets to work
OpenAI’s DevDay brought agents, a cheaper model and shared workspaces to ChatGPT. What caught my attention was the move toward handing over work, rather than asking for answers.
- Dots keep working between chats, using their own cloud computer and the apps you choose to connect.
- GPT-6.1 Sol promises performance close to Astra at one-fifth its standard token prices, according to OpenAI.
- ChatGPT Space brings files and shared pages together, letting teammates and their AI agents work on the same documents.
Why it matters: Small teams could delegate recurring research to Dots and review the findings together in Space, spending less time copying results between chats and shared documents.
Read more ↗Which DevDay update feels most important?
05 / AI highlights
AI in Brief
CODING
Claude Code Can Now Be Modded
Developers can build small mods that watch sessions, change Claude’s behavior and add custom interfaces.
SAFETY
OpenAI Cuts Ties With Three Safety Researchers
OpenAI says the researchers mishandled sensitive company information outside its approved internal procedures.
AVATARS
Nearly Half Thought This AI Avatar Was Human
In Tavus’s study, 48% of people believed Griffin was human after a one-minute video call.
VOICE AI
Microsoft’s New AI Transcribes You Mid-Sentence
Its new model starts returning text in just over 100ms, supports 60 languages and tops Artificial Analysis for accuracy.
On Our Radar
Last Week’s Poll Result
What does this incident reveal most about AI privacy?
| Agent behavior is now a privacy risk | 76.47% |
| De-identification alone is not enough | 5.88% |
| Research environments need stronger controls | 17.65% |
Does making AI hardware look fashionable make you more likely to use it?
| Yes, design matters | 11.11% |
| Somewhat | 5.56% |
| No, functionality matters more | 83.33% |
Reader voice · [email protected]
Selected: No, functionality matters more
“Cute for teens/young adults as an introduction into AI use. Academics and business professionals may want something less kitsch.”
Our take: Fair point. A playful design may lower the barrier to entry, but professional users will still judge it on utility first.
Reader voice · [email protected]
Selected: No, functionality matters more
“This Meta device is a fricken joke”
Our take: Strong reaction. If the hardware feels gimmicky, styling alone will not overcome doubts about whether it is genuinely useful.
Your turn: Vote and tell us why. We will feature your take next week.
Could typed decision models become a major layer in AI agent systems?
| Yes, they fit agents perfectly | 38.46% |
| Maybe, for narrow decisions | 38.46% |
| No, general models will dominate | 23.08% |
We’d love to hear from you!
Thanks for reading. See you next Sunday.
