Goals In, Deliverables Out

On July 9, OpenAI launched ChatGPT Work — an agent inside ChatGPT that gathers context across connected apps and files, stays on complex projects for hours, and ships finished documents, spreadsheets, presentations, and web apps — while merging Codex into a single desktop superapp, sunsetting the Atlas browser, and rolling out GPT-5.6 Sol on the same day Fidji Simo stepped down from the company's No. 2 commercial role.

Why now: IPO roadshow week needed a product story beyond government-cleared models. Washington spent 12 days gating GPT-5.6 behind roughly 20 vetted partners. Brussels adopted scraping rules. JPMorgan backtested eight agents. OpenAI answered with the thing it has been telegraphing for two years: stop asking ChatGPT questions — delegate the job.

What shipped:

• ChatGPT Work: An agent powered by GPT-5.6 that decomposes goals into steps, works across Slack, Gmail, Google Drive, SharePoint, CRMs, and calendars via a unified plugins directory, and asks for human approval before sensitive actions.
• The superapp merge: Chat, Work, and Codex now live in one desktop app on Mac and Windows — including the free tier. The standalone Codex app becomes the ChatGPT desktop app; Atlas begins sunsetting.
• GPT-5.6 family: Sol (flagship), Terra (balanced), and Luna (fast/cheap). Sol adds an ultra mode that coordinates parallel agent workstreams for the hardest tasks.
• Sites in beta: Turn work into shareable dashboards, trackers, and interactive reports by URL.
• Scheduled Tasks: Recurring agent runs — weekly Slack digests into slides, morning dashboard summaries, email-triggered deck updates.

What OpenAI claims inside its own walls:

• Finance: Month-end close and forecasting dropped from days to hours, per OpenAI's July 9 blog — teams reconcile in Sheets, build slides, and spend time explaining variances instead of hunting source data.
• Sales: A discovery call became a tailored proof of concept in 24 hours — a process that "normally takes weeks," OpenAI wrote — with the agent routing notes to solutions architects while the rep stayed customer-facing.
• Zapier: Head of enterprise marketing Angela Ferrante said ChatGPT Work traced customer touchpoints across CRM, email, and other tools to find broken follow-ups and surfaced "seven figures in potential sales" in missed pipeline.
• Codex reach: More than 5 million people use Codex weekly; over 1 million now use it for non-coding work, OpenAI said — the user base Work is meant to mainstream.
• Internal adoption: Nearly 100% of OpenAI's internal teams now use Work and Codex, per the company's blog.

What Altman said on CNBC:

• GPT-5.6 Sol is "54 percent more token efficient on agentic coding tasks" than its predecessor — Altman's direct answer to the enterprise complaint that dominated Sun Valley meetings.
• "This is the first year where A.I. spend has been a big topic," Altman told Julia Boorstin. "Everyone's asking what we can do to help reduce spend or increase value."
• On the government gate: "We made many changes through the process" after federal red-teamers found issues — collaborative, not ceremonial.
• On IPO timing, when asked if OpenAI goes public this year: "I don't know."

The governance layer — where buyers will actually live:

• Write-capable actions default to requiring human approval before emails send, CRM records update, or calendars change, per OpenAI's workspace security documentation.
• Enterprise and Edu admins control plugin access, browser use, network policies, spend limits, and Compliance API visibility into Work conversations.
• Auto-review runs frontier models against connected-tool actions before execution to block unauthorized data sharing.
• The pitch is control. The product is delegation. Those two sentences will collide in every procurement review.

The competitive and leadership context:

• The Verge framed ChatGPT Work as a direct answer to Anthropic's Claude Cowork — same week Ramp's June AI Index showed Anthropic edging OpenAI in paid U.S. business subscriptions.
• Meta shipped Muse Spark 1.1 the same day with a paid API — closed weights, metered inference, agentic coding at $1.25/$4.25 per million tokens.
• Fidji Simo — hired to scale applications and commercialization — stepped down to a part-time advisory role July 9 after a worsening POTS flare, leaving product and business responsibilities split among Greg Brockman, Sarah Friar, and Jason Kwon, per Bloomberg and the Wall Street Journal.
• Brockman posted on X: "We've brought together ChatGPT and Codex, in the form of ChatGPT Work: an agent for your most ambitious work."

What to watch:

• Usage economics: Work follows Codex's usage structure — longer tasks burn more credits. Admins can set group limits and review override requests with project rationale.
• Free-tier desktop access: Putting Work on free plans is a distribution land grab; enterprise spend controls are where margin lives.
• Agent sprawl: OpenAI's internal teams are already all-in. Your org's shadow agents are next.
• Commercial bench: Simo's exit leaves OpenAI without the Instacart IPO veteran it hired to run the applications business through listing — at the exact moment the product shifts from chat to delegation.

Convina's view: OpenAI finally named the category shift everyone has been dancing around: the unit of value is no longer the answer — it's the deliverable. ChatGPT Work is a credible superapp move, and the 54% efficiency gain is the number CFOs will quote in Q3 budget meetings. But the honest enterprise story is governance-first. Approval gates, spend caps, and auto-review are not footnotes — they are the product's admission that hours-long agents with CRM write access are liability machines wearing productivity makeup. Buy the delegate layer. Budget twice as long for the permission layer. And watch whether OpenAI can commercialize delegation without the executive who was hired to do exactly that.