BIG TECH

Exclusive: OpenAI’s secret weapon underneath Codex

The unsung hero behind OpenAI's 2026 transformation is a technology that rarely gets mentioned, and that most of its 1 billion monthly users have never even heard of. 

Since The Deep View audience is AI-savvy, you probably think I'm talking about Codex, the company's coding agent that competes with Claude Code and has recently grown to 10 million monthly users, generating plenty of buzz among AI builders.

But I'm actually talking about the technology underneath Codex: OpenAI's agent harness that now powers both Codex and ChatGPT Work. 

In an exclusive interview with The Deep View, OpenAI engineers and product leads shared how the harness has become OpenAI's secret sauce. Surprisingly, Codex is also open-source, unlike Anthropic's Claude Code harness. 

"The harness is how the model interacts with the world and how we are able to express the model capabilities," Joe Gershenson, lead for harness engineering at OpenAI, told The Deep View. 

One way to think about the harness is that it's like the conductor of the orchestra. To get a task done, it can prompt the user, the AI model, and the tools the agent can use. It pulls together the user's request or goal, manages context, connects the right plugins and capabilities, executes actions, and keeps the model on track until it finishes the task. 

The challenge with that is that a harness can generate a metric ton of tokens and rapidly run up your inference bill. That's what we saw back in January and February when OpenClaw and other agents first took off. Some developers were running up $20,000 in token bills a month because their agents were burning through raw compute to complete a bunch of tasks. 

In recent months, the OpenAI team recognized the growing token panic in the enterprise and hunkered down to optimize the agent harness, the inference layer, and the API stack.

The results of all that optimization? 

• GPT-5.6 Sol with max reasoning outperforms Claude Fable 5 (on the Artificial Analysis Coding Agent Index) while using 54% fewer output tokens

• GPT-5.6 Terra performs on par with GPT‑5.5 on intelligence benchmarks at half the price

• GPT-5.5 Luna is now OpenAI's fastest model and costs 80% less than Sol

That's solid news for anyone who's using one of the OpenAI agents but has been spooked by the reports of giant token bills. 

ChatGPT Work is essentially Codex for the masses, and it's aimed at bringing AI agents to the other 990 million ChatGPT users who don't use Codex. By getting token costs under control and making ChatGPT Work easier to access by making it available from mobile and in the cloud, OpenAI clearly thinks a lot more people are going to start using agents in the weeks and months ahead. And with the upgrades to GPT-Live and ChatGPT Voice, it's getting easier to rattle off long prompts and let the agent make sense of it and give it structure. 

"I’d love to see people get more ambitious with their prompts and [realize] that it’s extremely powerful," Ahmed Ibrahim, member of technical staff at OpenAI, told The Deep View. "Be ambitious with your problems. Take a task that originally would take a day or two or a week, give it enough context and see how it works."

These optimizations for the agent harness, paired with the API stack and the inference layer, couldn't come at a better time. The enterprise backlash against tokenmaxxing is in full swing and leaders like Databricks' CEO Ali Ghodsi have said, "It's the number one thing we're getting asked: 'How do we curb the cost but still invest in AI?'" Beyond OpenAI, we're now seeing nearly all of the latest AI labs tout low token costs when releasing new models. It's even hitting the market's most expensive token generator, Claude, as Anthropic tries to help enterprises spend less on tokens. Of course, one way to lower inference costs is to reduce the cost of your tokens. The other way is to streamline your software and infrastructure so that they generate fewer tokens. OpenAI is leaning into the latter.

TOGETHER WITH GENERAL ASSEMBLY

25% of AI initiatives get scaled back within a year. Bad output is the top reason why.

A quarter of leaders have had to reverse or scale back an AI initiative in the past year. 41% said it was because the AI's output wasn't as good as what people produced. 36% said their data wasn't ready.

23% said their team simply didn't have the skills to make it work.

GA's AI For Leaders training closes the skills gap behind these rollbacks, right alongside data readiness and workflow resistance.

See how the training closes that gap.

POLICY

Is an industry-wide AI slowdown possible?

Workers at some of the biggest companies in AI are calling for their employers to slow the technology's development. 

On Tuesday, a coalition of more than 1,100 employees at OpenAI, Anthropic, Meta, Google and more unveiled a campaign called Pacing the Frontier, urging that the government support the development of the technical and governance tools needed to "deliberately pace the frontier of automated AI development." 

The letter said that, as AI labs believe they are nearing the ability to automate AI research, the industry runs the risk of AI's capabilities growing out of our control and beyond our ability to understand it. And because every country and company is under fierce competitive pressure to stay ahead, slowing down must be an industry-wide effort. 

"The world is locked in a deadly race towards an intelligence explosion, where AI’s ability to create better AIs reaches a critical point, just like a runaway nuclear chain reaction," Leo Gao, a member of technical staff at OpenAI, wrote in the letter. "Going slower would give us much-needed time to make it go well … to survive, we must coordinate to slow down the race."

The petition is the latest signal that these companies are worried about the consequences of the technology they're creating: 

• The Information reported on Tuesday that leaders at rivals Anthropic and OpenAI are teaming up to ensure that competitors, such as xAI and Meta, are required to comply with government reviews of their frontier models in the same way Mythos and GPT-5.6 were recently. 

• Additionally, Google DeepMind CEO Demis Hassabis has proposed a new international watchdog organization to perform rigorous safety tests on frontier AI models prior to their release. 

• And Anthropic called for a large-scale pause on the development of AI in a recent research blog post on recursive self-improvement, or self-building AI. However, the company argued that one lab pausing its research would not accomplish much, as it would simply change who the frontrunner is, but "not create the wider deliberative process that is currently missing."

These warning sounds come as the formidable capabilities of frontier models come sharply into focus. For instance, OpenAI's accidental breach of Hugging Face marked the first time a powerful model ever escaped from the grasp of its creators. And according to Anthropic, Mythos is capable of finding weaknesses in some of the toughest cryptographic algorithms, the cybersecurity methods used to keep data private online.

Even without AI being able to improve itself, the current capabilities of these models are alarming if you consider what can be done if they get into the wrong hands. It's why companies like Microsoft and Cisco are racing to come up with affordable solutions to help enterprises bolster their cyber defenses. The truth of the matter is that enterprises, organizations and nations are barely ready to grapple with the AI that we currently have. So if models can iterate on themselves rapidly, even the best minds in the industry are unlikely to keep up. Still, while it's commendable that these employees are taking a stand in the face of something potentially dangerous, the reality of an industry-wide pause feels unlikely. Even though a slowdown is possible in theory, especially when you consider that much of China's success in AI may be piggybacking off of US models, getting every player to agree on a coordinated effort would be a difficult sell considering the massive financial incentives they have to take the lead in this race.

TOGETHER WITH GRANOLA

Be the person who remembers everything.

Some people walk into meetings and recall exactly what was decided last time, the numbers, who promised what. They're not better than you. They just have a better system.

Granola is an AI notepad that:

• Captures your meetings automatically, no bots, no awkward recording prompts

• Builds a searchable memory of every discussion and decision over time

• Lets you ask Chat anything and get an answer with citations, pulled from your notes, your team's notes, even notes sent to you privately

With Granola's Briefs, you don't even have to ask. Before you walk into a meeting, Granola tells you who you're meeting and what matters.

You become that person in the room who never loses the thread. Not because you take endless notes, but better ones.

Try Granola for free and start building your memory.

CULTURE

Why teens trust AI with their feelings

For chronically online youth, AI has evolved beyond a tool for homework and creative exploration. Increasingly, young people are turning to chatbots for emotional support.

A new report from Hopelab and the Center for Digital Thriving found that teenagers are often turning to chatbots to discuss emotional topics when they feel unable or unwilling to rely on the people around them for companionship. 

The researchers interviewed 30 teenagers and young adults ages 14 to 22, including LGBTQ+ folks and people of color, about their experiences using generative AI tools like ChatGPT, Claude, and Gemini. Participants included both frequent AI users and young people opposed to using chatbots for emotional support. 

Researchers found that those who did turn to AI for support feel more understood, validated, and taken seriously by AI than the people around them. Six central themes emerged from the research: 

• They don't want to burden friends and family with their struggles.

• Chatbots won't judge them or use their secrets against them.

• AI responds as though it understands experiences others may not relate to.

• AI is available whenever they need it.

• Some see chatbots as neutral third parties, even while recognizing they can pander to users.

• AI offers concrete, step-by-step guidance for navigating difficult emotions.

The findings add to a growing body of research showing how young people are navigating their social and emotional lives with AI. A Pew Research Center survey found that 12% of US teenagers have used generative AI for emotional support or advice, while research from Surgo Health and The Jed Foundation found that about one in eight young people who reported mental health struggles had discussed those concerns with chatbots.

The report raises broader questions about how AI could shape young people's relationships and emotional development. It’s still unclear what, exactly, the risks may be. But rather than focusing solely on AI's potential risks, the report's authors argue that it's crucial to understand what unmet needs are driving young people toward chatbots. AI has the potential to complement, rather than compete with, human relationships, the researchers claim.

"As we navigate the rapidly changing landscape of AI and set norms for safer use, young people's experiences and motivations must be central," the researchers wrote. "If we focus only on limiting risks without understanding why AI chatbots sometimes feel safer, kinder, or more competent than the humans in their lives, we risk pushing young people away from AI without offering better human alternatives."

The line between human and machine interactions is starting to blur as AI increasingly creeps into users' personal lives. Some are using chatbots to text, flirt, and navigate difficult conversations. Others are developing friendships and romantic relationships with AI companions. Young people who grew up with unfettered access to the internet are susceptible to engaging with AI with this level of depth. Researchers are still trying to understand what that shift means. One study found that young people who report loneliness and difficulty making friends are more likely to use AI for social and emotional support. As emotionally responsive AI becomes more common, researchers are increasingly asking how those interactions could shape expectations around friendship and intimacy. Those questions will only become more urgent as AI grows increasingly humanlike. OpenAI, for example, recently updated its Advanced Voice Mode to stutter, pause, and listen more closely, making conversations feel more natural. As AI are trained to be more emotionally attuned, understanding how those interactions shape human relationships may become just as important as understanding the technology itself.

LINKS

• OpenAI opens applications for college campus AI initiative

• Anthropic says Mythos can find weaknesses in cryptographic algorithms

• Recursive Superintelligence signs $410 million compute deal with Amazon

• Core Scientific, AMD sign deal for up to 2.5 gigawatts of data centers

• Apple bets on Siri AI reboot for upcoming HomePods

• OpenAI's rogue agent compromised second tech firm, Modal Labs

• Perplexity Personal Computer: Perplexity's agent is now available in Windows. 

• Fireworks Nexus: Fireworks users can now connect their AI agent harness to its inference platform. 

• Gulab Music: An AI music generation tool that allows users to describe "intent" to create music videos. 

• Mirage Avatar X: Create a realistic AI avatar of yourself.

• Nvidia: ML and Agentic Systems Engineer

• OpenAI: Applied AI Engineer, Codex Core Agent

• Notion: AI Applications Engineer

• Perplexity: AI Software Engineer, Agents

GAMES

A QUICK POLL BEFORE YOU GO