Hey there! 👋
Welcome back to SavvyMonk, your daily dose of AI and tech news that actually matters.
On 9 July 2026 OpenAI turned its chatbot into something that finishes the job rather than answering questions about it. The new agent is called ChatGPT Work, and it can take action across your apps and files, stay on one project for hours, and hand back finished spreadsheets, slides and documents on its own.
It runs on GPT-5.6, the model that shipped the same day after a two-week government delay, and it points straight at Claude Cowork, the rival Anthropic has been building since January 2026.
Let's get into it.
TODAY'S DEEP DIVE
ChatGPT Work Pairs GPT-5.6 With Codex to Chase the Enterprise Market Anthropic Leads
ChatGPT Work is an agent that lives inside ChatGPT and borrows its engine from Codex, the tool OpenAI first built for developers. You hand it a goal rather than a prompt, and it breaks the work into steps, moves across the apps you connect, and keeps going until it has something finished to show you.
The pitch is that the chatbot has grown from a thing you talk to into a thing that does the work, and the launch folds three moves into a single day, a new agent, a new model family, and a new desktop app that carries all of it.
The Agent Itself
The headline capability is stamina. ChatGPT Work can stay with a complex project for hours by splitting it into smaller steps and completing them independently, then handing back sheets, slides, docs and even small web apps. It follows you across web, mobile and desktop, so you can start a task from your phone, review a draft between meetings, and pick the same work up at your desk.
The engine matters here, because Codex already has more than five million weekly users, and over a million of them now use it for work that has nothing to do with software. That existing base is the wedge OpenAI is using to push an agent at people who never open a terminal.
The Model Benchmarks
The stamina runs on GPT-5.6, which arrived in three tiers that share one generation and split on cost. Sol is the most powerful and sets the price at 5 dollars per million input tokens and 30 dollars for output, Terra balances speed and quality for everyday work at 2.50 and 15 dollars, and Luna is the fastest and cheapest at 1 and 6 dollars. Sam Altman told an interviewer that Sol is 54 percent more token efficient on agentic coding, which is the number enterprises now watch most closely as their bills climb.
On benchmarks Sol scored 53.6 on Agents' Last Exam, reached 80 on a coding agent index, and lifted its cybersecurity score to 73.5 percent from the 47.9 percent its predecessor managed.
Artificial Analysis Coding Agent Index: an independent index of coding-agent performance across implementation, terminal use, and real codebases.
Two new modes sit on top, a max setting that spends more compute on hard problems and an ultra setting that runs four agents in parallel. The whole family landed roughly two weeks late after a United States government review over cybersecurity and biology concerns, following a limited preview that opened on 26 June 2026.
Long-horizon agentic workflows across professional domains.
The Models and its Pricing
As discussed above, OpenAI rolled out three tiers of GPT-5.6:
Sol (flagship): $5 per 1M input tokens, $30 per 1M output tokens
Terra (balanced, 2× cheaper than GPT-5.5): $2.50 per 1M input tokens, $15 per 1M output tokens
Luna (fast and affordable): $1 per 1M input tokens, $6 per 1M output tokens
Sol is OpenAI’s strongest model to date, with improvements in coding, biology, and cybersecurity. Sol can inspect and refine rendered interfaces, catching visual and functional issues, not just generating underlying code.
ChatGPT Work is available immediately for Pro, Enterprise, and Edu subscribers; Plus and Business plan holders get access within days.
What It Does All Day
The worked examples are where the ambition shows. In finance, OpenAI says the agent cut month-end close and forecasting from days to hours by finding the source data, moving it into a spreadsheet, reconciling it, building the slides and checking its own results. In marketing it can take customer research, turn it into a campaign brief, generate the assets from that brief, and adapt them for different markets while carrying the context through every step.
A Scheduled Tasks feature handles the repetitive end, refreshing a meeting agenda from new messages or sending a morning summary of what changed overnight. A new Sites feature in public beta turns the output into a shareable dashboard or web app, a built-in browser lets the agent pull in websites and online files, and more than fourteen hundred plugins connect the tools where your work already lives, from Slack and Google Drive to Salesforce and GitHub.
The Price Split You Need To Read
The rollout is not as simple as free for everyone, and the detail is worth reading before you plan around it. The new desktop app that carries Chat, Work and Codex together is available on every plan including Free, on both Mac and Windows. The agent on web and mobile is not, and it reached Pro, Enterprise and Edu first on launch day, with Plus and Business following over the days after.
Billing works like Codex rather than a flat feature, so usage scales with the size and complexity of the task, and OpenAI published no per-task price at launch. That makes the first week a measurement exercise, where the sensible move is to run one real workflow and watch how much of your plan a long agentic run actually consumes before you schedule it to repeat.
The Real Target Is Anthropic
None of this is happening in a vacuum. Claude Cowork launched in January 2026 and gave Anthropic a head start of roughly six months in autonomous workplace agents, a lead one widely cited analysis showed translating into stronger enterprise usage, especially in coding.
Anthropic has since packaged Cowork with pre-built agents for finance, human resources and legal teams, and only days before this launch it brought Cowork to web and mobile, a move that looks like it saw ChatGPT Work coming.
OpenAI answered on its home turf of the consumer surface, leaning on price and reach, and it folded Codex into the main ChatGPT desktop app under co-founder Greg Brockman to build what looks like a single super app.
The clearest sign of the strategy is what it retired to get there, since the standalone ChatGPT Atlas browser is being discontinued on 9 August 2026, its job absorbed by the browser now built into the agent.
What The Builders Are Saying
Early reactions from people who had tested the model ran hot, and a couple of them are worth embedding below. Pietro Schirano, who runs MagicPath AI, called it the best model he had used after months of testing, praising it as fast, smart and creative.
Theo Browne of T3 Chat described it as world-leading at computer use and said it had made him reach for it far more often than before. The most useful take came from Dan Shipper of Every, who split the difference by likening GPT-5.6 to a Porsche and Anthropic's Fable to a warp drive, arguing each wins on different terrain rather than one model winning outright.
The Bottom Line
The throughline is a company that spent two years teaching its chatbot to answer, now betting the next phase is finishing. ChatGPT Work is less a new idea than a repackaging of Codex for people who never wanted a coding tool, and its real weapon is distribution, since it reaches the consumer surface OpenAI dominates while Anthropic owns the enterprise lead.
Treat the free desktop app as the low-risk way to test whether an agent that runs for hours actually saves you any, and watch the usage bill closely, because the same stamina that finishes your work while you step away is the meter running the whole time.
AI PROMPT OF THE DAY
Category: Task Delegation
"Act as an operations lead. I want to hand a recurring task to an AI agent that can run for hours across my apps. Here is the task [describe the task, the apps and files it touches, and the outcome you need]. Map it into a sequence the agent can follow, mark every step where it should stop and ask me before acting, flag any step that touches sensitive data such as [customer records or finances], and tell me which parts are safe to schedule unattended and which still need me in the loop."
ONE LAST THING
The two biggest agent launches of the year now say the same thing from different directions. The work is moving off the desk and into the background, closer to a colleague that keeps going while you are asleep or sitting in a meeting, and the convenience and the exposure turn out to be the same feature seen from two angles.
The habit worth building early is deciding in advance which tasks you are willing to hand to something that runs without you watching, and keeping anything sensitive on the side where you still approve every step before it ships.
Hit reply, I read every response.
See you tomorrow.
— Vivek
P.S. Know a founder or operator weighing whether to let an AI agent run their busywork for hours at a time? Forward this to them. They can subscribe at https://savvymonk.beehiiv.com/