Photo by Emiliano Vittoriosi on Unsplash
- OpenAI launched ChatGPT Work on July 9, 2026, powered by GPT-5.6, combining ChatGPT and Codex into a single autonomous task-execution product — a direct counter to Anthropic's Claude Cowork, which debuted January 12, 2026.
- Anthropic overtook OpenAI in enterprise AI tool spending for the first time: 34.4% vs. 32.3%, per Ramp corporate card data as of July 2026.
- OpenAI earns roughly 85% of revenue from consumers; Anthropic earns roughly 85% from business contracts — making enterprise agentic tools existential for OpenAI's margin profile ahead of its anticipated IPO.
- Agentic AI operating against an over-permissioned SaaS stack is a security event waiting to happen. Audit integration grants before enabling agent features, not after the first incident.
What Happened
34.4% to 32.3%. That's the spread — Anthropic over OpenAI — in enterprise AI tool spending captured by Ramp corporate card data as of July 2026. It marks the first time Anthropic has led that metric. Keep that number in mind, because it explains the urgency behind what OpenAI shipped on July 9, 2026.
According to Google News, Reuters reported the release of ChatGPT Work as OpenAI's most direct competitive push against Anthropic to date. The product fuses ChatGPT's conversational interface with Codex — OpenAI's coding engine — into an agentic layer capable of generating documents, building presentations, and publishing functional websites through autonomous multi-step execution. It runs on GPT-5.6, OpenAI's most capable model, whose public release had been delayed weeks earlier at the U.S. government's request. Executive Order 14409, signed by President Trump on June 2, 2026, requires AI developers to provide the federal government up to 30 days of pre-release access to their most capable models — the first documented case of direct federal intervention in a major commercial AI model launch schedule.
Access launched July 9 for Pro, Enterprise, and Edu subscribers, with Plus and Business tiers receiving access over subsequent days. The announcement landed 48 hours after Anthropic expanded Claude Cowork to mobile and web on July 7, 2026 — a coordinated competitive tempo both companies are clearly setting deliberately. Microsoft, an OpenAI backer, had staked out this same territory earlier, unveiling Copilot Cowork shortly after Anthropic's January launch. The agentic AI race now has three named products from the three largest players in enterprise software, all competing for the same CIO budget line.
The Agentic Pattern: Tool-Call Orchestration at Product Scale
ChatGPT Work is not a chat upgrade. Architecturally, it's a ReAct loop (Reason + Act, a pattern where the model alternates between reasoning steps and tool invocations) built on GPT-5.6 and exposed as a product interface. In production terms: the model receives a high-level task, decomposes it into steps, selects relevant tools for each step — Codex for code generation, document APIs for slide creation, rendering pipelines for web output — calls those tools in sequence, evaluates intermediate results, and iterates when a step fails. The user sees a polished deliverable; under the hood there are multiple model calls, tool invocations, and context injections before anything surfaces.
Claude Cowork runs the same fundamental orchestration pattern — plan, tool-call, verify, loop. Where the two products diverge is architectural emphasis. Anthropic's implementation inserts safety validation at each reasoning step, consistent with its stated positioning as a safety-first enterprise AI lab. OpenAI's implementation, backed by GPT-5.6's capability ceiling, prioritizes output breadth: a single agent session can produce a document, a working codebase, and a deployable website. For enterprise teams evaluating both, AI Trends' breakdown of GPT-5.6's Sol, Terra, and Luna model tiers is directly relevant — the variant you're allocated within ChatGPT Work will meaningfully affect both output quality and per-session token costs at scale.
Why the Enterprise Revenue Gap Makes This Existential for OpenAI
ChatGPT serves more than 900 million weekly active users as of June 2026. OpenAI's annualized revenue has surpassed $25 billion, and the company was valued at $852 billion in a March 2026 funding round that secured approximately $122 billion in new capital from SoftBank, Microsoft, and Nvidia. But a Forbes analyst report identifies the structural problem beneath those headline figures: OpenAI generates roughly 85% of its revenue from individual subscriptions, while Anthropic derives roughly 85% from business contracts. Enterprise customers generate three to five times more revenue per token than consumer users — meaning OpenAI's massive user base is structurally concentrated on the lower-margin side of the market.
The 2028 forward projections, drawn from a Deep Research Global company analysis, make the gap stark. Anthropic projects $70 billion in revenue with gross margins approaching 77% and $17 billion in positive cash flow. OpenAI's computing expenditure alone is projected at $121 billion in 2028 — with a $74 billion projected loss that year. Those trajectories don't converge without a meaningful shift in OpenAI's enterprise customer base.
Chart: Anthropic vs. OpenAI enterprise AI tool spending share, based on Ramp corporate card data as of July 2026 — the first recorded instance of Anthropic leading this metric.
Gartner adds urgency to the timeline: as of 2026, Gartner forecasts that 40% of enterprise applications will embed task-specific AI agents by year-end, up from under 5% in 2025. The AI agents market is estimated at $10.9–12.1 billion in 2026, growing at a 44–46% CAGR through 2030, per Gartner. The platform that becomes the default enterprise agent runtime during this adoption window will accumulate compounding advantages in integration depth, proprietary workflow data, and organizational lock-in that become extremely difficult to displace later.
The Andreessen Horowitz enterprise AI report offers useful calibration on OpenAI's actual position: 78% of surveyed enterprise CIOs report running OpenAI models in production today. ChatGPT Work doesn't need to open new accounts — it needs to deepen usage within accounts it already holds, expanding from passive text generation into agentic workflow execution. That's a shorter sales motion than Anthropic faces, which needs to convert existing CIO interest into signed enterprise contracts at scale. In my analysis, the product launch is almost secondary to the IPO narrative: OpenAI filed a confidential S-1 on June 8, 2026, targeting an $850 billion valuation, with CFO Sarah Friar reportedly eyeing a 2027 market debut. Demonstrating credible enterprise margin expansion between now and that listing is the real objective behind ChatGPT Work's release date.
Where This Breaks in Production
The demos look clean. The production failure modes are not.
The permission scope problem. Agentic AI doesn't read your SaaS ecosystem — it acts on it. The Noma Security CISO guide states the risk directly: "The biggest risk isn't the AI itself; it's that your SaaS ecosystem is over-permissioned. If your employees are using ChatGPT in agentic mode, your attack surface just expanded dramatically." An OAuth grant that felt harmless for read-only summarization — Google Drive, Slack, Notion, Jira — becomes a write-access risk the moment the agent moves from suggesting content to executing tasks. The agent that can draft a document can also move a file. The agent with Slack credentials can send a message to a channel, or an external contact, without a human review step. Audit your integration permission grants before enabling agentic features, not after the first incident report.
The orchestration cost problem. A five-step autonomous workflow generates multiple GPT-5.6 calls, with tool outputs injected back into the context window at each step. Context window blowups — where accumulated intermediate results push the total token count past the model's effective reasoning threshold — cause silent quality degradation, not hard failures. The model continues to produce output; the output quality quietly degrades. ChatGPT Business is priced at $20 per user per month on annual billing ($25 monthly); Enterprise starts at approximately $60 per user per month. Neither rate reflects the token burn of multi-step agentic chains running at volume. Model that cost explicitly before organization-wide deployment — the per-seat subscription is the predictable line item; the token cost of agentic chains is where budgets get surprised.
The silent completion failure. The agent reports success; the output is plausible but wrong. Without eval-driven development — automated checks on agent output before it reaches a downstream system or gets delivered to a stakeholder — this failure mode is invisible until something breaks publicly. Agent demos are optimized to hide the retry logic. Build your evaluation harness before you build your agent workflow, not as a phase-two afterthought.
Frequently Asked Questions
What is ChatGPT Work and how does it work?
ChatGPT Work, launched July 9, 2026, is an agentic AI product from OpenAI that combines ChatGPT with the Codex coding engine. Powered by GPT-5.6, it autonomously executes multi-step tasks — drafting documents, generating presentations, building websites — without requiring user input at each intermediate step. Unlike standard ChatGPT, which responds to individual prompts, ChatGPT Work plans an entire task, calls relevant tools in sequence, evaluates each intermediate output, and iterates until the task is complete. Initial access was limited to Pro, Enterprise, and Edu accounts on launch day, with broader tier rollout following over subsequent days.
How does ChatGPT Work compare to Claude Cowork for enterprise teams?
Both products run on the same underlying agentic pattern: multi-step planning, tool-call orchestration, and iterative verification. Anthropic's Claude Cowork, launched January 12, 2026, emphasizes safety validation at each reasoning step and targets enterprise customers as its primary revenue base — approximately 85% of Anthropic's revenue comes from business contracts, per a Forbes analyst report. ChatGPT Work leverages GPT-5.6's broad output range and the distribution advantage of over 900 million weekly active users as of June 2026. Enterprises running existing Microsoft integrations may find ChatGPT Work easier to onboard; those prioritizing documented safety governance may find Claude Cowork's architecture better aligned with enterprise procurement requirements. The Andreessen Horowitz enterprise AI report notes that 78% of enterprise CIOs are already running OpenAI models in production, which gives ChatGPT Work a significant incumbent advantage in expansion sales.
What are the security risks of using ChatGPT Work AI agents in enterprise environments?
The primary risk is the interaction between agentic AI and an over-permissioned SaaS ecosystem. If your organization has broad OAuth grants to tools like Google Drive, Slack, Notion, or Jira, enabling agentic mode means the AI can write, move, or delete content — not just read and summarize it. Noma Security's CISO guide characterizes this as a dramatically expanded attack surface. The mitigation is to scope down integration permissions to the minimum required for the agent's intended task before enabling agentic features, and to treat agent-initiated write actions with the same audit logging applied to human user activity. Prompt injection — where malicious content in a retrieved document redirects the agent's behavior — is a secondary risk specific to agents that retrieve and process external data as part of their workflow.
How much does ChatGPT Work cost per user for business and enterprise pricing?
As of July 10, 2026, ChatGPT Business — which provides access to ChatGPT Work features as they roll out to each tier — is priced at $20 per user per month on annual billing, or $25 per user per month on a monthly plan. Enterprise pricing starts at approximately $60 per user per month and includes additional security, compliance, and administrative controls. Neither pricing tier explicitly accounts for the token cost of multi-step agentic chains at scale. Organizations planning organization-wide deployment of agentic workflows should model expected per-chain token costs separately using OpenAI's published token pricing, in addition to the per-seat subscription fee, to avoid budget overruns when autonomous task volume scales.
Disclaimer: This article is editorial commentary for informational purposes only and does not constitute financial, investment, or legal advice. The author has not independently tested the products discussed. Research based on publicly available sources current as of July 10, 2026.