PRODUCTS

GPT-5.6 takes big step toward OpenAI's superapp

OpenAI just did a lot more than release its most advanced language model. 

While GPT-5.6 went into full public release on Thursday, as we've been expecting all week, the company also announced a ton of other developments and updates at the same time, including the next big step toward its long-awaited superapp. 

Like Anthropic's Mythos, the US government worked with OpenAI to delay the broader release of GPT-5.6 until government agencies and corporations were better prepared for the model's ability to uncover software vulnerabilities that could enable cyberattackers and other bad actors. 

The centerpiece of today's announcement remains the GPT-5.6 family of models, which now includes Sol (flagship), Terra (balanced), and Luna (fastest and least expensive). The models have been performing like champs on benchmarks such as Agent's Last Exam, outperforming Anthropic's Mythos and Fable. The OpenAI models are also less expensive. For 1M tokens, Sol costs $5 input / $30 output while Mythos and Fable cost $10 / $50. 

But OpenAI also used Thursday's announcement to roll out several other significant updates:

• Codex moves into the ChatGPT app: This has been in the works since March, as Codex transformed from a coding agent to a more general agentic tool after the meteoric rise of OpenClaw, whose founder, Peter Steinberger, joined OpenAI's Codex team. The combination of the two apps is still a little confusing, but the OpenAI team assured me that the Codex brand will remain and the agent harness will continue to be open-source. ChatGPT users will now get an AI agent inside the chatbot they already use. 

• Launch of ChatGPT Work: Perhaps the most significant new development is the launch of ChatGPT Work, an agent that brings the kinds of tasks for knowledge work that people have been doing with Codex on the desktop to the ChatGPT app across web and mobile. There will be a new toggle between Chat and Work when you load ChatGPT, so you can do agentic work from anywhere you're already using the OpenAI chatbot.

• ChatGPT Sites: From ChatGPT Work you can now describe a site you want to build in natural language with no understanding of code, and it can make it, deploy it, and give you a live link on the web that you can share with anyone. This handles the deployment work that you would have previously needed a tool like Lovable, Replit, or Vercel to do.

"We have more than 5 million people using Codex every week, but close to a billion people using ChatGPT. And so, for the largest part of the audience, they had not really seen the power of agents," Romain Huet, head of developer experience at OpenAI, told The Deep View. "[Today begins] this progressive reveal for almost a billion people that can now delegate more tasks of more complexity to something like ChatGPT Work to do more things for them."

While we expected the GPT-5.6 launch this week and the models themselves are a known quantity at this point, the first big step toward bringing Codex and ChatGPT into a single app experience was a bigger surprise. Of course, it's long overdue since OpenAI executives such as Greg Brockman have been talking about it since the end of Q1. But it's not easy, since Codex had turned into an agentic power-user app used by rabid enthusiasts, while ChatGPT is the mainstream interface used by a billion people. The first steps still look a little confusing to me, but it was time to start bringing them together to create a more unified experience, and I'm glad OpenAI has begun the process. In the short term, starting to introduce a lot more people to the time-saving power of agents will be the biggest win. And running OpenAI agents across desktop, mobile, and web will be the big upgrade for the power users who already rely on Codex.

IN PARTNERSHIP WITH LAMBDA

Cut your AI training costs by 25% or more

Most large-scale AI training runs use less than half the computing power they're paying for. Lambda's team found the root causes and built a reproducible framework that boosted efficiency by over 25%, without changing the model itself.

Lambda’s whitepaper shows you how to address:

• Memory inefficiencies silently inflating your costs

• Training configurations that aren't making full use of your hardware

• Bottlenecks that slow down GPU communication

Get the whitepaper.

PRODUCT

Can Claude's new screentime tool make AI healthier?

Wondering how reliant you've become on AI tools? Anthropic is giving you clarity.

On Thursday, Anthropic launched a new reflection dashboard in Claude's settings for the web or desktop app. The dashboard gives you a summary of how you've been using Claude, including key topics, your usage patterns, and the type of tasks you usually use it for, according to the blog post. You can access it at claude.ai/settings/reflect. 

Beyond giving you greater insight into how you use the tool, there are also options users can take to modify their behavior, such as setting up quiet hours and nudges to take breaks. Periodically, the reflection dashboard will also present users with questions prompting them to reflect on their AI usage, which they can then discuss with the chatbot. 

Soon, there will also be a view of how much time users have spent using Claude. It is now available in beta on the Free, Pro, and Max plans on web and desktop. It runs only when memory is on, as it needs to refer to past chats, so if, like me, you don't see it, it may be because you have that setting off. The reflection dashboard leaves out incognito chats and health integration content.

While the aim for a feature like this is good, helping users become aware of just how much of their lives they're offloading to AI, Anthropic isn't holding back on fueling AI usage either. On the same day, it announced a reset of the 5-hour weekly rate limits for all users of its Fable model, in response to high demand for its most advanced model and to compete with OpenAI's release of GPT-5.6.

People have been over-reliant on their devices for decades, seduced by the entertainment and convenience sitting right in their pocket. AI's hook is different: it doesn't just entertain; it promises to make your life better and your work more productive. Whether that's genuine ambition or laziness in disguise, AI can now help with many tasks tied to knowledge work or human interaction. Just yesterday, I wrote about how many adults are turning to it for emotional support. The first step in recognizing a problem is seeing it quantified, but we've been here before. Screen time trackers proved that awareness alone rarely changes behavior. Still, even if this doesn't move the needle on healthy usage, there's value in knowing exactly what you're leaning on AI for, if you're willing to evaluate it honestly.

IN PARTNERSHIP WITH DATALAB

Turn docs into records that fit your schema. Open source

Datalab released Lift, an open source 9B model that turns messy documents into clean, structured data:

• Pulls user-defined fields into your schema

• Works with any schema - invoices, scientific papers, 10-Ks, etc.

• Fast - 9.5s median, 8x quicker than Azure

• Run it locally or via Datalab's API for scale

Try Lift on your own documents - select 'Extract' to start.

RESEARCH

CEOs think AI is working. The data disagrees

AI promises to automate workflows and save employees time. But a lack of alignment among company leaders may keep those productivity gains out of reach.

While 85% of C-suite executives report deploying AI across their organizations, only 54% of managers say AI is a top priority, according to new data from Nitro, a document software company, which surveyed more than 1,300 executives, managers, and directors across the US, UK, and Canada. Barely half of managers say AI has reached at least some of their own workflows, revealing a disconnect between executive aspirations and day-to-day realities.

"At an executive level, it can be about the message: 'Yes, we are using AI,'" Cormac Whelan, Nitro's CEO, told The Deep View. "At the manager level, it is often about solving for the next level of detail on how we are using it."

Among managers whose teams have adopted AI, 37% primarily rely on standalone AI tools, such as copying and pasting content into ChatGPT. Only 12% say AI is fully embedded into their document workflows.

Much of the work remains manual. More than half of managers say employees spend the most time extracting data from documents into spreadsheets, followed by editing and formatting files, managing documents, summarizing content, and redacting sensitive information. Because of that, 62% of managers say employees spend six or more hours each week on manual document tasks, with nearly a third estimating those workloads consume 11 or more hours.

"If AI isn't built into the workflows, work surfaces, and systems people use every day, with the outcome in mind, it doesn't matter how many tools you buy or how many tokens [you burn] or how much budget you commit," Whelan said. "The work stays manual."

The findings underscore a growing challenge in enterprise AI adoption. Businesses are betting big on AI to boost productivity and cut costs, but many managers appear to be taking a more measured approach to rolling it out across their teams.

Managers cite security and trust as their biggest barriers to AI adoption, followed by integration complexity and implementation costs. More than half also say sensitive documents are being uploaded to public AI tools, while only 43% report having clear, actively enforced AI policies.

To close the gap, the findings suggest that companies need AI embedded into the software and workflows employees already use every day, with security and governance baked in from the start. Otherwise, AI risks becoming another set of tools rather than the productivity engine executives expect it to be.

AI adoption is as much an organizational challenge as it is a technological one. While executives focus on transforming the business to appease stakeholders, managers are responsible for making AI work day-to-day. As layoffs shrink middle management and spread teams thin, many are juggling a mounting list of responsibilities. They're not only expected to lead their teams, but also figure out how to integrate AI into existing workflows and systems. Closing that gap goes beyond creating an AI strategy. It requires building a culture where employees have the time, training, and trust to experiment with AI out of curiosity rather than purely top-down pressure.

LINKS

• Meta introduces Muse Spark 1.1, a low-cost multimodal reasoning model

• SpaceXAI rolls out Grok 4.5 model aimed at coding, agents, and cost savings

• Open-source AI tool Ollama raised a $65 million Series B

• OpenAI published a detailed audit that found task issues in SWE-Bench Pro 

• Cerebras, Flex expand partnership to scale CS-3 AI chip production capacity

• AI companies are spending millions of dollars on elections, report 

• Starbucks uses AI for tools to cut reliance on Microsoft and IBM software

• Notion: Introduced Ship OS "the agent-native way to ship software"

• Claude: Has a new reflection dashboard to see your usage patterns 

• Muse Spark 1: Meta's new  multimodal reasoning model 

• Manus: Introduced branch, a feature that lets you split any conversation into a parallel session that inherits your full context

• Google Photos: Video Remix turns clips into stylized videos, powered by Gemini Omni

• National Hockey League (NHL): AI Engineer

• BlackRock: Artificial Intelligence & Machine Learning Engineer, Associate

• Capgemini: Gen AI / Agentic AI Developer

• Brooklyn Sports & Entertainment: AI Engineer

GAMES

POLL RESULTS

Would you use AI for relationship advice?

No (54%)
Yes (25%%)
Maybe (21%)