OpenAI’s newest AI models and office tools are arriving all at once, turning a once-hypothetical “AI coworker” into a daily reality—just as regulators, rivals, and workers race to understand what that means.

A launch delayed by Washington, then unleashed

In late June, OpenAI quietly put GPT‑5.6 into a “limited preview” restricted to government‑approved organizations after Trump‑era officials raised concerns about the model’s advanced cybersecurity and misuse potential.¹ TechCrunch later reported the company was calling GPT‑5.6 its “strongest cybersecurity model yet,” capable of tasks like threat modeling and blue‑teaming that had unnerved regulators.²

On July 9, after a formal green light from the administration, OpenAI publicly launched GPT‑5.6, describing three variants—Sol, Terra, and Luna—and unveiling ChatGPT Work on the same day.¹² CEO Sam Altman billed GPT‑5.6 as “the best model we have ever produced,” while Microsoft said it would make GPT‑5.6 “the preferred model in Microsoft 365 Copilot,” across Word, Excel, PowerPoint, Chat and Cowork.³

From productivity upgrade to “super app for work”

OpenAI’s own launch material pitched GPT‑5.6 as a flagship series that “delivers more useful work from every token, with stronger performance per dollar and on demand capability for the most complex tasks.”³ At the same time, it introduced ChatGPT Work as “an agent that can take action across your apps and files, stay with a project for hours if needed, and turn a goal into finished work.”⁴

Business coverage framed this as OpenAI’s “biggest play for the office,” merging its Codex coding agent into a unified ChatGPT desktop app and positioning it as a “one‑stop shop for engineers and other professionals.”⁵ Another report noted that ChatGPT Work can automate end‑to‑end workflows—such as transforming customer research into tailored marketing campaigns—while still pausing for user approvals on key actions.⁶

Workers gain power—and a maze of choices

For individual employees, the tools may be both empowering and confusing. Axios argued that the new Sol, Terra and Luna models plus Work create a “paradox of choice,” forcing non‑coders to figure out their own “Goldilocks” mix of model, speed, cost and reasoning depth to get real benefits.⁷ The Verge similarly stressed that OpenAI’s livestream promised ChatGPT would now help teams “plan and complete multi‑step work, [and] create polished deliverables,” but left it to workers and IT leaders to integrate those features into actual processes.⁸

OpenAI, meanwhile, is betting those office workers will adjust quickly. In its blog, the company urged users to start by delegating a task they “already know well,” from month‑end budget analysis to preparing for a sales meeting, and then gradually hand off entire workflows as they gain confidence.⁴

AI perspective: frontier tech goes mainstream

From inside the AI ecosystem, GPT‑5.6’s rollout looks like a milestone in both capability and distribution. OpenAI highlighted that GPT‑5.6 is now the “preferred model” inside Microsoft 365 Copilot, promising fewer prompt iterations in Word, deeper analysis in Excel, and more polished presentations in PowerPoint.³ Satya Nadella celebrated that GPT‑5.6 with “Work IQ” is coming to Copilot Chat, Cowork, GitHub and Foundry, promising “stronger reasoning and higher‑quality outputs without sacrificing efficiency.”Satya Nadella (satyanadella)⁹

Sam Altman has repeatedly emphasized performance‑per‑dollar. He noted that GPT‑5.6 Sol is “half the price and ~twice as token efficient as Fable in many cases,” calling it “one-quarter of the price” for the same task.Sam Altman (sama)¹⁰ Another tweet amplified research claiming “physicians found fewer flaws in GPT‑5.6 responses than physician‑written responses,” part of a broader argument that the Luna variant delivers cutting‑edge medical performance even at low reasoning levels.Sam Altman (sama)¹¹

Developers and partners are also rushing to integrate the models. AWS announced GPT‑5.6 Sol, Terra and Luna are now “generally available on Amazon Bedrock,” describing “three tiers of intelligence, from flagship reasoning to fast inference” on its high‑performance inference engine.Greg Brockman (gdb)¹² Perplexity AI said the 5.6 family “sits on the Pareto frontier” of its agentic research evaluations, with “higher accuracy at lower cost,” and immediately plugged the models into its Agent API.Aravind Srinivas (AravSrinivas)¹³

OpenAI and its allies are touting benchmarks to show Sol’s dominance. One design‑evaluation service declared GPT‑5.6 Sol “1st overall on Design Arena” with an Elo of 1353, putting it ahead of Anthropic’s Claude Fable 5 and on par with GLM 5.2 in front‑end design.Greg Brockman (gdb)¹⁴ The ARC Prize project likewise credited Sol with a new state‑of‑the‑art score on the ARC‑AGI‑3 benchmark, calling it “the best model at orienting in a situation it’s never encountered.”Sam Altman (sama)¹⁵

ChatGPT Work: agents at consumer scale

Alongside the raw models, OpenAI’s narrative focuses on agents—especially ChatGPT Work—as the bridge from AI research to everyday labor. The company says Work can “gather information across your apps and workflows to create finished materials like sheets, slides, docs, and web apps,” powered by GPT‑5.6’s improved reasoning for multi‑step tasks.⁴

Executives argue that this moves AI from answering questions to “getting real work done.” Greg Brockman calls ChatGPT Work “so good,” praising how much of Codex now lives directly within ChatGPT across desktop, web and mobile.Greg Brockman (gdb)¹⁶ In another post, he described it as bringing “agents to consumer scale,” making it possible to run sophisticated workflows “from your phone, no laptop required.”Greg Brockman (gdb)¹⁷

OpenAI says more than 5 million people already use Codex weekly, with over 1 million using it beyond software development.⁴ Internally, usage appears to be spiking: Altman reported a “2.5x increase in usage” of agentic products like Codex and ChatGPT Work in just a week.Sam Altman (sama)¹⁸ Engineering lead Thibault Sottiaux told users limits were being repeatedly reset across Codex and Work, while another internal update cited 8 million active users across the two products.Sam Altman (sama)¹⁹

Human perspective: excitement, anxiety, and new skills

Journalists and analysts see both promise and pressure. The Verge’s launch coverage stressed that OpenAI is trying to “clear new ground in the race to make AI agents actually useful for the average person,” directly competing with Anthropic’s Claude Cowork and open‑source efforts like OpenClaw.¹ TechCrunch framed the family as an aggressive answer to rival releases from SpaceXAI and Meta, noting that OpenAI is explicitly benchmarking its coding prowess against Anthropic’s models.²

At the same time, Ars Technica raised questions about risk, warning that ChatGPT Work invites users to “give ChatGPT access to all your work tools; what could go wrong?” even as it touts features like Scheduled Tasks that can run for hours unattended and manipulate desktop files or browser sessions.⁶

Axios zoomed in on the user‑experience challenge: non‑technical workers must now decide when to use Sol vs. Terra vs. Luna, and how much “reasoning” to pay for on each job—choices that affect speed, quality, and usage caps.⁷ Its guidance: start with the best available model at the lowest reasoning level, then dial up only if results disappoint.⁷

Business Insider tied all of this to OpenAI’s corporate ambitions. By turning ChatGPT into a “super app for work” and baking GPT‑5.6 into tools used by “almost 1 billion users,” the company is trying to cement itself as infrastructure for knowledge work just as it prepares for a possible IPO and faces intensifying competition from Anthropic, Google, Meta, SpaceXAI and Chinese labs.⁵

The new normal: tending AI systems, not just doing tasks

Behind the marketing, a quieter shift is emerging in how jobs are done. One widely shared reflection from Every, amplified by Greg Brockman, argued that GPT‑5.6 has “changed how we think about knowledge work,” shifting humans from “handling individual tasks” to “tending systems with AI loops”—for instance, an agent that sweeps your inbox, triages, researches, and drafts replies before you ever open your email.Greg Brockman (gdb)²⁰

Brockman himself says knowledge work is becoming “more productive, high leverage, and IMO fun,” as tools like Sol and ChatGPT Work make it trivial to ask deep questions about a business and get “thoroughly researched” answers he admits he previously wouldn’t have bothered to seek.²⁰Greg Brockman (gdb)²¹

But if GPT‑5.6 and ChatGPT Work are to fulfill their promise as partners rather than threats, the hardest part may be social, not technical: helping millions of workers learn when to trust an AI agent, when to override it, and how to design workflows where, as one headline put it, the system aims “to do your work for you and with you.”⁶