ChatGPT or Claude? Chat, Coworking or Code?

This guide rests on one thesis: agentic coworking, delegating real multi-step work to an AI that plans, uses your files and tools, and returns finished deliverables, grew out of agentic coding and will quickly surpass chat as the dominant paradigm for knowledge work. If you accept that thesis, platform selection stops being a chatbot comparison and becomes a stack decision. This guide decodes the confusing landscape, tests both stacks layer by layer, and answers the question the thesis creates: which one is better for your knowledge work team over the next 24 months.

Orientation

Start here: the 60-second answer

The landscape in one paragraph. Each vendor now offers three tools, not one. A conversation tool for questions, drafting, and research (ChatGPT Chat, Claude Chat). An agentic work tool that you delegate real multi-step work to, which uses your files and apps and produces finished deliverables (ChatGPT Work, Claude Cowork). And a developer tool for building software and automation systems (OpenAI Codex, Claude Code). Each tool runs on multiple surfaces: web, desktop app, and mobile. Most of the confusion in this market comes from two facts: the same tool has different powers on different surfaces, and the vendors renamed or reshuffled nearly everything between May and July 2026. Underneath the naming, four separate decisions are being collapsed into one: which platform, which mode, where the work executes, and which plan. This guide takes them one at a time.

Why accept the thesis at all? Because everyone with skin in the game already has. Both vendors rebuilt their entire consumer products around the agentic layer in the same quarter: Cowork went cross-device on July 7, 2026, and ChatGPT Work launched July 9 in a reorganization that retired four chat-era products to fund it. Enterprise budgets moved first and hardest, with agentic and API spend driving the market's revenue crossovers. And the capability differentiation now lives almost entirely in the coworking layer: as the grid below shows, chat is close to a tie. When both competitors, the buyers, and the feature gap all point the same way, the paradigm claim is no longer speculative.

How to use this guide

Three minutes: read Decode the names and The two ideas, then follow the decision tree. Fifteen minutes: add the capability grid and the verdict. Betting a workflow or a company standard on a specific cell: also read The vendor bet and Stability watch, because this market invalidates static comparisons in weeks; the source register dates every claim. Everything here is deliberately non-interactive: the whole guide, tree and grid included, prints cleanly if you want it on paper or in a PDF.

The translation layer

Decode the names

Same jobs, different labels. This table is the Rosetta Stone. If you internalize nothing else, internalize row two: ChatGPT Work and Claude Cowork are the same product category, the agentic workspace, and they are where most of the real capability differences live.

The mental model

The two ideas that explain everything

Idea 1: The mode is about WHO does WHAT work

Chat is you driving. You ask, it answers, you steer every turn. Seconds to minutes. Cheapest against your usage limits. Work and Cowork are you delegating. You describe an outcome, the agent plans, breaks the job into steps, uses tools and files, runs code, checks its own output, and hands back finished deliverables: documents, spreadsheets, decks, reports, dashboards. Tens of minutes to hours, and it burns usage allowance much faster than Chat. Codex and Claude Code are for building systems: software, pipelines, and repeatable automation harnesses.

Idea 2: The surface is about where you interact with the tool and what it can do from there

Web and mobile give the agent cloud hands: it can search, run code in a sandbox, call connectors, and make files, but it cannot touch your machine. The desktop app gives it local hands: your folders, your applications, your browser. One tension governs everything downstream: local and unattended are opposing requirements. Work that needs your machine needs your machine on; work that must run while you sleep must live in the cloud. The two vendors resolved that tension with opposite architectural bets:

Put simply: OpenAI bolted local agency onto a chat product; Anthropic built a cloud agent and gave it a desktop bridge. Both work. But the Claude design means continuity, cross-device steering, and unattended operation are the default behavior of every session rather than properties of one thread type.

The 30-second route

The decision tree

Start from the outcome you need, follow the branch that matches where the work's raw material lives, and land on a route. Then check the deciders table below: if one of those needs is non-negotiable, it overrides the tree. Runner-up routes are named so you can see what the other platform's answer would be.

The deciders: needs that override the tree

If any row below is non-negotiable for you, it settles the platform question by itself, regardless of which branch you came down.

Decision one: the product

The capability grid: can it do X, and where?

For most decisions you need two columns, not eight. This grid answers "can the platform do it at all," and the note in each cell tells you which mode and surface it lives in, because that location is usually the insight. Every cell is backed by the full per-surface evidence in the 8-surface matrix appendix, which is where the per-surface evidence and the source citations live.

Reading tip: when both columns say ✓, the note is the difference. "Work on desktop only" versus "every Cowork session" is the whole architectural argument of this guide in miniature.

Decision one, resolved

The one-platform verdict

The question, stated in full: if you accept the thesis that agentic coworking, derived from agentic coding, will quickly surpass chat as the dominant paradigm for knowledge work, which of these stacks is better for your knowledge work team? Under that thesis, the coworking layer carries the decision's weight. Agentic coworking means delegating real work: multi-step tasks against your files, tools, and accounts, producing finished deliverables, often unattended, sometimes on a schedule. Not chatting. Chatting is a tie, and under the thesis it is also not the point.

The six reasons

1. The architecture is built for delegation. Every Cowork session runs on Anthropic's servers in an isolated sandbox. Continuity when your laptop closes, steering from any device, and scheduled runs with no device online are the default behavior of every session. ChatGPT Work gets there only for cloud threads; anything that touches your machine becomes a local thread that is stranded on that machine, does not sync, and stops when the machine does.

2. The research surface has no holes. Claude's Research tool works from web, desktop, and mobile Chat, and Cowork runs sub-agent research inside any task. ChatGPT's Deep Research, which remains an excellent product, currently cannot be started from the company's own flagship desktop app; the documented workarounds are the website or the legacy Classic app.

3. Seven months of hardening versus four weeks. Cowork shipped January 12, went GA on desktop April 9, and extended to web and mobile July 7. ChatGPT Work launched July 9, 2026 in a reorganization that retired Agent, Operator, Canvas, and group chats simultaneously, and the seams show: the Deep Research gap, community reports of desktop builds missing Chat mode entirely, Sites still business-gated, PowerPoint output absent, and the Atlas browser being sunset on August 9.

4. Acting on the real web, as you. Claude in Chrome is GA inside Cowork on every paid plan and operates your actual browser with your logged-in sessions. ChatGPT's equivalent authenticated agency exists only in the desktop app's separate isolated browser (which deliberately does not touch your Chrome profile), and its cloud browser cannot pass a login page at all. For knowledge work that lives behind SaaS logins, this is a daily-felt difference.

5. The floor is higher everywhere else. Office-format file creation on every plan including Free. Skills on every plan. Sub-agent orchestration you can direct and watch. Live Artifacts as self-refreshing deliverables. Memory at no cost. 1M-token context on the 5-series models for paid plans. The capability is spread across the whole product rather than concentrated in one hero surface.

6. The vendor trajectory points the same way. Over the last six months Anthropic shipped one coherent agentic arc and retired nothing; OpenAI shipped at least as much but retired or renamed five consumer surfaces along the way, and the resulting change tax lands on whoever standardizes on the churned surfaces. The full argument, the counter-evidence, and the falsifiable tests are in The vendor bet.

Scorecard

Decision two: the vendor

The vendor bet: current winner versus company to bet on

A platform choice is not a snapshot purchase. It is a 12-to-24-month relationship in which skills, connectors, scheduled automations, custom assistants, and file conventions accumulate on whichever surfaces you standardize. So two different questions hide inside "which platform": which product fits today (the grid answers that) and which vendor's product management do you want to be downstream of (this section). For agentic knowledge work, the two answers now point the same way, so the split that remains is by thesis, not by time horizon: buy breadth if you reject it; buy Claude if you accept it. Either way, vendor behavior determines whether today's green cells stay green.

The observable record, February to August 2026

Product management quality is not directly observable from outside. What is observable: what ships, what gets retired, how changes are communicated, what breaks and how it gets repaired. Six months of dated record:

The enterprise adoption signal

Enterprise buyers vote with budgets, and that vote is a leading indicator for where integration partners, plugins, tooling, and talent accumulate. The verified numbers as of this writing: Menlo Ventures' enterprise LLM research had Anthropic overtaking OpenAI in enterprise API spend in mid-2025, and 2026 coverage puts the split at roughly 32% Anthropic versus 25% OpenAI of enterprise LLM API spend, with Anthropic capturing 40%+ of AI coding workloads (about two-to-one). Ramp's AI Index, drawn from transaction data across 70,000+ US businesses, showed Anthropic edging past OpenAI in overall business AI spending share for the first time in mid-2026, roughly 34% versus 32%. [R1, R2]

Two honesty notes on the popular "70:30" framing. First, no credible report shows Anthropic winning enterprise 70:30 overall; the two-to-one shapes appear only in specific cuts, coding spend among them, and in new-adopter splits that this guide could not independently verify. The overall enterprise API lead is real but modest, and the Ramp sample skews US mid-market. Second, the counterweight is large: OpenAI still leads total revenue and consumer scale by a wide margin, and consumer gravity funds a lot of enterprise persistence. Read the enterprise signal as confirming the trajectory argument, not as a landslide. [R2]

Revenue motion and the fast-follow pattern

The revenue curves say the same thing more loudly. Press-reported run-rates put Anthropic at roughly $1B in January 2025, $9B by year-end, $14B in February 2026, about $30B in April 2026 (the point at which multiple outlets report it passing OpenAI's roughly $25B), and about $47B by late May. Epoch AI measured the underlying growth at roughly 7x per year for Anthropic against roughly 3.4x for OpenAI, with 80 to 85 percent of Anthropic's revenue coming from enterprise and developer customers. Handle with care: these are press-reported run-rates, not audited financials, growth multiples compress as bases grow, and OpenAI still holds an enormous consumer franchise of roughly 800 million weekly users. But the direction is not ambiguous. [R3]

Product motion points the same way, in the form of who follows whom. The dated sequence: Claude shipped skills (October 2025) and plugins before OpenAI's Plugin Directory (July 9, 2026); Cowork went cross-device on July 7, 2026 and ChatGPT Work launched July 9; Claude Desktop's Chat-plus-agent home preceded the unified Chat/Work/Codex app. And the sharpest third-party signal: Microsoft, OpenAI's largest partner, now offers Anthropic models inside Microsoft 365 Copilot and in 2026 made Claude the default model for enterprise tenants. Fast-following is rational, and OpenAI executes it at volume; but the direction of imitation tells you who is setting the category's agenda. [R4]

How hard to lean: hard on the category, precise about the scope

On agentic knowledge work, the evidence no longer supports calling this close: Anthropic leads the product category, the enterprise spend share, the revenue growth curve, and the direction of imitation, and its disruptions have been answered with public postmortems rather than silence. The precision matters on scope, though. OpenAI remains a much larger consumer company, its shipping volume is enormous, its July consolidation is a rational strategic response that could work, and run-rate crossovers are reported figures, not audited ones. So bet the category on Anthropic, keep the bet monitored through the tests below, and resist the urge to extend the conclusion from "losing the enterprise agentic-work category" to "losing." Those are different claims, and only the first one is currently supported.

The decision algebra

For a company standardizing on one platform, score four terms rather than one.

A defensible weighting for a long-lived standard: roughly 65% current capability and governance fit, 25% vendor trajectory and change risk, 10% reversibility. Capability gates; trajectory breaks ties; never the reverse.

What each vendor must prove next

These are the falsifiable tests. Score them quarterly and let the answers move your weighting.

The third leg

Coding: where coworking came from

Every stool here has three legs: chat, coworking, and code. The third leg matters to a knowledge work decision for one reason above all: agentic coworking is agentic coding, generalized. Anthropic ran the sequence in order: Claude Code proved the agent loop in terminals through 2025; skills, plugins, sub-agents, and scheduled runs grew up there; Cowork (January 2026) is the same engine pointed at knowledge work, and the two still share one substrate, so a skill built for either runs in both. OpenAI ran the play in reverse: Codex existed for engineering, and Work (July 2026) generalized it after Cowork had proven the category. That lineage, not marketing, is why the two coworking surfaces differ in maturity by roughly the gap in their ages.

The third leg is also where the stacks diverge most visibly today. The April 14, 2026 Claude Code Desktop redesign turned the desktop app into a parallel-agent workspace: multiple sessions running side by side in one window, drag-and-drop panes for terminal, editor, diff viewer, preview, and chat, side-chats that pull context without touching the main session, managed worktrees, and cloud Routines, available from the Pro plan up. Neither ChatGPT nor Codex offers an equivalent side-by-side parallel-agent desktop interface. A knowledge work guide cares because this is the ceiling of the stack: it is where your power users end up, where one person visibly runs several delegated workstreams at once, and where the market already voted with the roughly two-to-one enterprise coding-spend split. The ceiling tells you how far the floor will rise. [A19, R2]

The daily operating guide

Inside your platform: which tool when

Once you have a platform, the daily question becomes: which mode do I open? The rule generalizes across both vendors:

Escalate from Chat upward when any of these become true: the job has more than a few steps; files or folders go in or out; it needs your connected tools to act, not just to be read; it should keep running while you do something else; it should repeat on a schedule; or it needs its own verification pass before you would trust the output. Each step up costs more usage allowance and more setup, which is why Chat remains the right default for most turns of an ordinary day. And wherever an action carries legal, financial, or reputational consequence, keep a human sign-off at the end of the chain; drop that review only after a workflow has proven itself repeatedly on low-stakes runs.

The reuse ladder (works the same on both platforms)

A second ladder runs inside the first, sorted by how often you repeat the work. Do it once: a plain chat is enough. Do it three or four times with the same background: make a Project, and load it with context, instructions, and an example of what good looks like, so every new chat starts primed. Run the same multi-step process across many different subjects: make a Skill (or a Custom GPT on ChatGPT), which fixes the process and leaves the subject as the variable; the classic example is proposals, where the structure never changes and only the client does. Should it run on a cadence with nobody watching: make it a scheduled task in Cowork or Work. This framing is adapted from consultant Dylan Davis and colleague Elliot; it matches how both platforms are actually built, and on Claude a Skill made this way runs everywhere from free-plan Chat to Cowork to Code.

One more rule from the same practitioners, worth adopting whole: if a step can be code, make it code. Deterministic steps (parsing, math, renaming, filtering, formatting) are more reliable, faster, and cheaper as scripts than as model calls, and the AI will write the scripts for you. Both platforms' skills can embed scripts alongside instructions, so a well-built skill mixes determinism for the mechanical steps with model judgment only where judgment is needed.

On Claude

Claude rules of thumb: if you would attach more than three files, start in Cowork instead. If you will ever run it twice, ask Cowork to schedule it, and keep scheduled work on cloud sources. If a deliverable should stay current rather than be regenerated, make it a Live Artifact on desktop. If you are pasting the same instructions weekly, turn them into a Skill, which then works even in free-plan Chat.

On ChatGPT

ChatGPT rules of thumb: start Deep Research from the browser, not the desktop app. Do logged-in web work only from desktop Work's built-in browser. If a Work task must survive your laptop closing, start it on the web as a cloud thread, not as a local desktop thread. Keep images and GPTs in Chat; treat Work's image generation as a bonus until the Images documentation lists it.

The money

Plans and pricing

Prices verified August 5, 2026. Both vendors now sell identical-feature tiers that differ only in usage volume, which simplifies the decision: pick the platform first, then buy the smallest tier that fits and upgrade only when you hit caps.

The clock

Stability watch: what just changed, and what to re-check

Both vendors materially reorganized their consumer products in the ten weeks before this guide was written. If you are reading this after early September 2026, re-verify the cells you care about before deciding; several are likely to have flipped. The recent timeline that produced today's landscape:

Cells most likely to flip next

• Deep Research returning to the ChatGPT desktop app. The gap looks like a rollout artifact, not a strategy; expect it to close in a client update. That would repair ChatGPT's weakest current story, and it is one of the "must prove" tests in The vendor bet.
• Cowork web and mobile reaching Claude Pro. Anthropic says weeks. When it lands, the $20 Claude Pro seat becomes the cheapest full agentic workspace anywhere, and this guide's Max recommendation softens for light users.
• Work's image generation and code execution appearing in the help center, not just the learn.chatgpt.com manual, and Sites opening beyond Business and Enterprise with PowerPoint output arriving, which would strengthen ChatGPT's deliverables story.
• Cowork's computer use and Live Artifacts maturing out of desktop-only, which would remove Claude's last device dependency.
• Model leapfrog: both vendors shipped flagships within the last five weeks; assume the model-quality comparison resets quarterly and buy the architecture, not the model of the month.

The evidence layer

Appendix: the full 8-surface matrix

The evidence layer behind the two-column grid: every capability resolved to the exact mode and surface (Chat versus Work/Cowork, web versus desktop, per platform). Every cell was checked against vendor documentation on August 5, 2026, across two validation rounds. Bracketed codes cite the source register.

The receipts

Sources and method

Method. Round one: three deep-research reports (ChatGPT Deep Research, Claude Research, Gemini Deep Research, all August 5, 2026) were read in full and diffed; twenty-plus conflicts were re-validated in four parallel research passes against vendor documentation. Round two: an independent report-merger built by ChatGPT Work was reviewed; its nine new or conflicting claims were verified in a fifth research pass, confirming several and refuting two; its strongest structural ideas were adopted. Enterprise adoption figures were verified against Ramp's published index and press coverage of Menlo Ventures' enterprise LLM research. Grid semantics: Yes means officially available on that exact surface, subject to ordinary quotas and permissions; Partial means beta, plan-gated, OS-dependent, indirect, or materially reduced; No means official documentation places the capability elsewhere, or its absence is reproducibly observed (noted where so); Not established means vendor documentation does not resolve the exact surface.

Primary source register

The three input reports and the ChatGPT Work merger, all dated August 5, 2026, are on file with the author.