Welcome back to 4IR. Here’s today’s lineup:
Google cuts AI costs up to 50% with new Flash models - Gemini 2.5 Flash-Lite slashes output tokens in half while improving coding benchmarks by 5 points
OpenAI launches ChatGPT Pulse for $200/month subscribers - New feature writes personalized morning briefings while you sleep, wants access to your calendar and email
OpenAI opens first Stargate data center in Texas - Abilene facility goes live as part of $500B infrastructure push with Oracle and SoftBank
Anthropic’s September 28 data opt-out deadline passes - Users who didn’t act now have conversations retained for 5 years for training
TOP STORY: Google cuts AI costs up to 50% with new Flash models
The story: Google DeepMind released updated Gemini 2.5 Flash and Flash-Lite models on September 25th that slash serving costs dramatically. Flash-Lite cuts output tokens by 50% while Flash reduces them by 24%, yet both models improved 5 points on the SWE-Bench Verified coding benchmark. The updates focus on better instruction following, reduced verbosity, and stronger multimodal capabilities. Developers can use new “-latest” aliases that automatically point to the newest models without code changes.
What we know:
Released September 25, 2025 to Google AI Studio and Vertex AI
Flash-Lite reduces output tokens by 50%, Flash by 24%
Both improved 5 points on SWE-Bench Verified (48.9% → 54%)
Better tool use for complex, multi-step applications
More concise answers reduce latency and costs
Improved audio transcription and image understanding
New “-latest” aliases for easier model updates
Why it matters: When the AI industry faces an $800 billion revenue shortfall by 2030 according to Bain, Google just showed you can cut costs in half without sacrificing quality. This could provide a path forward for an industry struggling with unsustainable infrastructure spending, proving efficiency gains don’t require performance sacrifices.
This is the efficiency breakthrough AI companies desperately needed. Earlier this week Bain warned the economics don’t work - companies are burning cash on infrastructure faster than they can monetize it. Google just proved you can cut serving costs in half while improving benchmarks, making those unit economics look much more sustainable.
OpenAI launches ChatGPT Pulse for $200/month subscribers
The story: OpenAI released ChatGPT Pulse on September 25th, a feature that generates personalized morning briefings while users sleep. Available only to $200/month Pro subscribers on mobile, Pulse synthesizes information from chat history, saved memories, and optionally connected apps like Gmail and Calendar. The system delivers 5-10 daily update cards covering topics from meeting agendas to birthday reminders to sports news. Updates expire after 24 hours unless saved. OpenAI plans to eventually roll it out to Plus subscribers and then all users.
What we know:
Launched September 25, 2025 for Pro users ($200/month) on mobile only
Generates 5-10 personalized briefing cards each morning
Can connect Gmail and Google Calendar for context
App integrations off by default
Updates expire after 24 hours unless saved
Uses chat history and explicit user feedback
Plans to expand to Plus users, then everyone
Why it matters: OpenAI is shifting ChatGPT from a tool that waits for questions to an assistant that anticipates your needs. The trade-off is clear: better personalization requires more access to your calendar, email, and personal data. This represents OpenAI’s vision for AI moving from reactive to proactive.
OpenAI wants ChatGPT checking your calendar and email overnight to deliver personalized morning updates. That’s genuinely useful for busy professionals who’d benefit from meeting prep, relevant news, and contextual reminders. The app integrations default to off, which respects privacy while making the feature optional for those who want it.
OpenAI opens first Stargate data center in Texas
The story: OpenAI and Oracle brought online the first Stargate data center in Abilene, Texas on September 23rd, launching the flagship site of their $500 billion infrastructure program. The facility is operational with Oracle Cloud Infrastructure and Nvidia chips, with potential to scale past one gigawatt of capacity - enough electricity to power 750,000 U.S. homes. OpenAI announced five additional sites across Texas, New Mexico, Ohio and the Midwest, bringing total investment to nearly $400 billion over three years. The project aims to build 10 gigawatts of computing capacity by end of 2025.
What we know:
First Stargate facility opened September 23, 2025 in Abilene, Texas
Part of $500 billion infrastructure partnership with Oracle and SoftBank
Can scale past 1 gigawatt - powers 750,000 homes equivalent
Five more sites announced across Texas, New Mexico, Ohio, Midwest
Total: nearly $400B investment over 3 years for 7 gigawatts
Target: 10 gigawatts by end of 2025
OpenAI pays for compute as operating expense
Nvidia invested equity to help fund construction
Why it matters: OpenAI is building data centers faster than anyone in history to solve its compute shortage problem. The scale is unprecedented - $500B committed while generating $10B in annual revenue. Whether this ambitious bet pays off depends entirely on whether AI demand grows fast enough to justify the infrastructure investment.
“No one in the history of man built data centers this fast,” OpenAI’s CFO said, and the execution is genuinely impressive. The economics are complex - Nvidia invests equity to fund data centers, then gets paid again when those centers buy chips. OpenAI commits $500B against $10B in annual revenue, betting that AI demand will grow fast enough to justify it.
Anthropic’s September 28 data opt-out deadline passes
The story: September 28th marked the deadline for Claude users to opt out of having their conversations used for AI training. Originally announced August 28th, the policy change means users who didn’t opt out will have data retained for 5 years instead of the previous 30-day automatic deletion. The change affects Claude Free, Pro, and Max users, though Enterprise customers remain unaffected. Anthropic frames it as improving model safety and capabilities, but the shift from privacy-first to opt-out data collection marks a significant reversal for the company that built its brand on responsible AI.
What we know:
September 28, 2025 was the opt-out deadline
Originally announced August 28, giving users one month
Data now retained 5 years vs. previous 30-day deletion
Affects Free, Pro, Max users and Claude Code
Enterprise, Government, Education customers exempted
Users can still change settings but already-used data stays
Pop-up defaulted to “On” for data sharing
Anthropic says it doesn’t sell data, uses it for training
Why it matters: When the company that built its reputation on privacy switches to the same data collection practices as everyone else, it reveals the competitive pressure all AI companies face. Building models that compete with OpenAI and Google requires massive datasets, and that need appears to override other priorities.
Anthropic built its brand on being the privacy-respecting alternative to OpenAI and Google, then switched to opt-out data collection with one month’s notice. The 5-year retention period and the toggle defaulted to “On” meant many users likely missed the change. Like other AI companies, Anthropic needs large datasets to stay competitive.
Note: Commentary sections are editorial interpretation, not factual claims