OpenAI Unveils ChatGPT Work as Rivalry With Anthropic Grows

OpenAI launched ChatGPT Work on Thursday, an agent that pairs the company’s chatbot with its coding tool Codex so non-coders can draft documents, presentations and websites. The new agent runs on GPT-5.6, OpenAI’s most advanced AI model family, which also debuted Thursday after OpenAI deferred an earlier launch window at the U.S. government’s request over national security concerns, per Reuters. Per the published API rate, Luna, the smallest GPT-5.6 tier, lists at $1 input and $6 output per million tokens, against $5 input and $30 output for the flagship Sol, and OpenAI says Luna can complete a task about as well as the largest tier does, at one-fifth the cost. Anthropic launched Claude Cowork in January, and ChatGPT Work is OpenAI’s product in the same category.

What ChatGPT Work Does

ChatGPT Work turns OpenAI’s chatbot into an agent that finishes office jobs for users who have never written a line of code. The product pairs the chatbot with Codex, OpenAI’s coding tool, inside the same chat surface hundreds of millions of people already know. A user describes an outcome in plain English and the agent returns a finished document, presentation or website, with the structured reasoning hidden behind the curtain.

Both ChatGPT Work and Anthropic’s Claude Cowork target the same persona: a knowledge worker who needs a finished deck, document or site without learning the underlying tooling. OpenAI named three starting outputs (documents, presentations and websites) and bundled a new ChatGPT desktop application and hosted websites feature into the same launch day. OpenAI built ChatGPT Work on top of prior agentic offerings, including Operator and deep research, later consolidated as ChatGPT Agent for individuals and as Workspace Agents for enterprise workflows. The new tool is OpenAI’s bid to bridge the demo-versus-daily-use gap that has dogged earlier AI agents.

You can apply the model’s ability to code to solve problems across every industry.

That was the framing Ty Geri, OpenAI’s product manager for ChatGPT Work, used in the launch interview. The pitch was simple: a model that can reason through a coding task can reason through a quarterly forecast or a sales brief in the same sweep, and OpenAI wants that framing to sound obvious. The product sits inside a chat surface with hundreds of millions of users, so launching through it is the cheapest distribution OpenAI could ask for. The cost story is also aimed at non-developers, where OpenAI sees the volume audience. Geri described the model as ‘competitive with models that are far, far more expensive at twice the speed and much, much cheaper.’

The launch framing rests on a bet about who shows up. Non-developers are the volume audience OpenAI is chasing. A generalist user has to feel the agent is as reliable as a colleague would be. OpenAI’s case is that the agent can meet that bar, with real workloads as the test.

GPT-5.6’s Three Tiers and Cost Stakes

The model that powers ChatGPT Work, GPT-5.6, ships as a three-tier family rather than a single flagship. Per the GPT-5.6 preview page listing Sol, Terra, Luna pricing, the lineup runs Sol, the company’s strongest agentic model to date on the Terminal-Bench 2.1 coding benchmark, Terra, an everyday tier positioned at GPT-5.5-class performance, and Luna, a fast and affordable entry. The pricing gap between Sol and Luna is where OpenAI’s cost story lives.

Per OpenAI’s published rates, Sol lists at $5 per million input tokens and $30 per million output tokens. Terra lists at $2.50 input and $15 output. Luna lists at $1 input and $6 output. OpenAI introduced more predictable prompt caching alongside the family, with a 30-minute minimum cache life and a 90% discount on cached reads. The three-tier structure gives enterprise buyers three price points for matching capability against cost per million tokens.

This is the first time where I’ve seen the small models complete these kinds of tasks.

Max Weinbach, an analyst at the consulting firm Creative Strategies, said that to Reuters. Weinbach’s larger framing, that the smallest GPT-5.6 tier can complete a task about as well as the largest at one-fifth of the cost, is the line OpenAI has leaned on most in launch materials. The implication for an enterprise buyer is that a fleet of agents can run on Luna-tier inference without paying Sol-tier bills. OpenAI’s test is whether the cost story holds up under real workloads.

How OpenAI Is Targeting Anthropic’s Cowork

OpenAI’s launch materials name the rival without softening the framing. ChatGPT Work is, in OpenAI’s words to Reuters, a direct answer to Anthropic’s Claude Cowork, the agent Anthropic launched in January for the same white-collar workflow. Cowork runs multi-step tasks across a user’s files, calendar, email, messaging apps and web tools, returning a finished deliverable for human review. The category OpenAI and Anthropic are both chasing is one Microsoft and Google already have agents shipping in: non-coders who need a finished slide deck, audit folder or weekly metrics report.

Anthropic has put Claude Cowork on the same pricing rail as its other paid plans. The Claude Cowork product page with plan pricing lists Pro at $17 a month on annual billing or $20 monthly, Max 5x at $100 a month, Max 20x at $200 a month, and Team at $20 per seat. Cowork is bundled with every paid Claude plan at no extra cost. Anthropic has also expanded Cowork to web and mobile in beta, starting with Max plan subscribers and adding other paid tiers in the weeks that followed.

Per Anthropic’s recent disclosure, more than 90% of Cowork’s usage has come from everyday knowledge work, with software development making up the rest. That mix is why both companies are selling agents to office workers across departments. The cost story differs at the model layer: Anthropic prices Claude usage through plan-based and credit systems and rolled out doubled Cowork usage limits through August 5 as a temporary boost. OpenAI is leaning on the three GPT-5.6 tiers to make a different cost argument, with the cheapest tier able to match the most expensive on common tasks. The two companies are pitching different products with different pricing structures.

Both companies target the same kind of buyer with overlapping channel structures. Anthropic’s enterprise pitch leans on plugin marketplaces, custom roles and admin-level controls, part of an expanded feature set the company added to Cowork over the year. OpenAI is leaning on the desktop app, the hosted websites feature and the assumption that ChatGPT’s existing user base will translate into ChatGPT Work signups. The backdrop to OpenAI’s push was Cowork going to web and mobile in beta, which removed the desktop-install requirement and pulled more paying users into the funnel.

The Security Holdup That Almost Killed the Launch

GPT-5.6 almost did not reach an open launch on Thursday. OpenAI said on June 26 that it was deferring broad public release of the GPT-5.6 lineup at the U.S. government’s request, citing national security concerns. The company previewed the models’ capabilities as part of an ongoing engagement with the government and limited the release to a small group of trusted partners whose identities were shared with Washington. OpenAI framed the deferral as a short-term step, not a long-term default.

By Wednesday this week, Axios reported, the White House had approved a broader rollout. OpenAI CEO Sam Altman confirmed it on X the same day, writing: “GPT-5.6 Sol launches Thursday! Happy building.” OpenAI launched the rollout with what it called its most robust safety stack to date, including stronger protections for higher-risk activity, sensitive cyber requests and repeated misuse. The company spent multiple weeks hardening safeguards against real-world attacks and committed more than 700,000 A100-equivalent GPU hours to automated red-teaming. OpenAI also said the model does not cross what it calls the Cyber Critical threshold under its Preparedness Framework.

1. June 26: OpenAI defers broad public rollout of GPT-5.6 at the U.S. government’s request over national security concerns.
2. July 8: Axios reports the White House has approved a broader rollout of the GPT-5.6 family.
3. July 9: GPT-5.6 launches with Sol, Terra and Luna; ChatGPT Work debuts on web and mobile.
4. Within days: ChatGPT Work expands from Pro, Enterprise and Edu to Plus and Business tiers.

An Arms Race Heading Into Two IPOs

The ChatGPT Work launch lands at a moment when both OpenAI and Anthropic are openly preparing for possible public offerings. Reuters reported that the two companies are competing for enterprise business, which executives at each firm have framed as the more lucrative market compared to consumer subscriptions. The launch is one of several moves this quarter designed to make each company’s enterprise narrative legible to investors before any pricing window is set.

The enterprise fight is well under way. Anthropic’s Cowork has been on the market since January, and OpenAI’s prior agent products, including Operator and deep research (later consolidated as ChatGPT Agent) and Workspace Agents, set the baseline OpenAI is now trying to extend. Per Anthropic’s recent disclosure, more than 90% of Cowork activity comes from knowledge work, with software development making up the rest. That share is the same buyer behavior OpenAI is now targeting with ChatGPT Work.

Anthropic’s push to put Cowork on mobile and web in beta ran in parallel with OpenAI’s Thursday launch. Anthropic has extended its doubled Cowork usage limits through August 5 to pull paying users onto Max plans during the rollout. OpenAI is countering with a different mix: lower-cost tiers, a desktop application and a hosted websites feature that pull ChatGPT’s existing user base into agent usage. The shared risk for both companies is implicit in the pricing: buyers will tolerate the cost of frontier models only as long as unit economics hold. OpenAI is making the case that the cost gap closes as smaller models catch up to flagship performance, by launching the new family in three sizes.

The stakes for the IPO narrative are concrete. Anthropic’s enterprise tier adds admin controls, usage analytics and OpenTelemetry observability that some CIOs require. OpenAI’s launch positioned the desktop application and the hosted websites feature as add-on products for IT departments to evaluate, on top of the three-tier model mix. The U.S. government’s early involvement in the GPT-5.6 rollout means both companies now face an additional layer of review before each new flagship ships, with any new front in that review becoming a material disclosure for public-market investors.

Both OpenAI and Anthropic are openly preparing for possible public offerings, Reuters reported, and each company is now putting finishing touches on its enterprise narrative ahead of that milestone. With broader availability of Sol, Terra and Luna planned for the coming weeks per OpenAI’s launch material, the GPT-5.6 launch adds a model priced so the smallest tier completes tasks at one-fifth the cost of the largest. ChatGPT Work sits inside the chat surface that already reaches hundreds of millions of users, with the desktop application and hosted websites feature bundled into Thursday’s release.

Who Gets ChatGPT Work First

ChatGPT Work reaches paying customers in stages. OpenAI is putting the product on web and mobile for Pro, Enterprise, and Edu users starting Thursday, with Plus and Business tiers following within a few days. The new ChatGPT desktop application ships alongside the agent, and the hosted websites feature lets a user build and publish a site directly through ChatGPT Work.

Under the surface, ChatGPT Work runs on top of the same ChatGPT subscription tiers the company already sells, with API pricing on the model family carrying the per-million-token rates OpenAI published on Thursday. Anthropic’s Claude Cowork is included on Pro, Max, Team and Enterprise plans, with Pro at $17 a month on annual billing or $20 monthly, Max 5x at $100 a month, Max 20x at $200 a month and Team at $20 per seat. Anthropic has extended its doubled Cowork usage limits through August 5 as a temporary boost to pull paying users toward heavier plans. OpenAI is making a different pitch: the ChatGPT user base is already there, and the per-million-token cost on Luna gives the company a unit-economics advantage. Both companies are positioning themselves as the affordable entry point into agentic AI, with each side defining ‘affordable’ against a different baseline.

Ty Geri, OpenAI’s product manager for ChatGPT Work, told Reuters the launch is meant to be ‘competitive with models that are far, far more expensive at twice the speed and much, much cheaper.’ OpenAI is betting that the cost-performance math holds up under enterprise workloads, the case the company has leaned on hardest in the launch narrative. That case will be the same one OpenAI makes to investors when its IPO window opens.

Frequently Asked Questions

What is ChatGPT Work, and what does it do?

ChatGPT Work is an AI agent OpenAI launched on Thursday, July 9. It pairs the ChatGPT chatbot with OpenAI’s Codex coding tool so non-developers can describe an outcome in plain English and receive a finished document, presentation or website. The agent runs on OpenAI’s GPT-5.6 model family, which also debuted Thursday.

What is GPT-5.6, and what are Sol, Terra and Luna?

GPT-5.6 is OpenAI’s most advanced model family, also launched Thursday. Sol is the flagship tier, strongest on coding, biology and cybersecurity benchmarks. Terra is positioned at GPT-5.5-class performance for everyday work. Luna is a fast and affordable tier, with API pricing per million tokens listed at $1 input and $6 output versus Sol’s $5 input and $30 output.

How does ChatGPT Work compare to Anthropic’s Claude Cowork?

Both agents target non-coders who need a finished deck, document or site without learning the underlying tooling. Cowork went to mobile and web in beta earlier this month after a desktop-only period, and Anthropic priced it into Pro ($17 a month on annual billing), Max 5x ($100), Max 20x ($200) and Team ($20 per seat) tiers. OpenAI’s launch is structured around three GPT-5.6 model tiers with different price-to-performance ratios, plus a new ChatGPT desktop application and a hosted websites feature.

When can I use ChatGPT Work?

ChatGPT Work began rolling out Thursday to Pro, Enterprise and Edu users on web and mobile. Plus and Business customers are scheduled to get access within a few days, per OpenAI’s Thursday announcement. The hosted websites feature lets a user publish directly through ChatGPT Work, alongside the new ChatGPT desktop application.

Why was GPT-5.6 delayed?

OpenAI said on June 26 that it was deferring broad public release of GPT-5.6 at the U.S. government’s request, citing national security concerns. The release was limited at first to a small group of trusted partners whose identities were shared with Washington. The White House approved a broader rollout on Wednesday, July 8, and the public launch followed on Thursday.