Has ChatGPT Finally Caught Up with Claude? A Side‑by‑Side Look at Two Office Agents

Release Overview

On July 9 OpenAI released three items simultaneously: the GPT‑5.6 model series (Sol, Terra, Luna), a desktop application that combines Chat, Work, and Codex into a single shell, and the ChatGPT Work agent.

GPT‑5.6 Series

Sol is the flagship model, Terra is a balanced variant for everyday work, and Luna offers the lowest cost. All three are being rolled out in ChatGPT, Codex, and the OpenAI API. An “ultra” mode can coordinate multiple agents in parallel, substantially shortening complex task completion time.

Codex Integration

The standalone Codex app was merged into the new desktop client, creating a three‑in‑one structure where Chat, Work, and Codex share a single interface.

ChatGPT Work

ChatGPT Work is positioned as a general‑purpose autonomous agent. It can operate applications, process complex files, and continue working for hours after receiving a final goal. The agent performs task decomposition, step execution, and product delivery without further user intervention.

Model‑Level Benchmarks

In the Agents' Last Exam benchmark, GPT‑5.6 Sol achieved a new high with roughly one‑quarter the estimated cost of Claude Fable 5, while Terra and Luna surpassed Fable 5 at about one‑sixteenth of the cost.

In the Artificial Analysis Intelligence Index (AAII) at maximum reasoning intensity, Sol lagged Fable 5 by one point but completed tasks 61 % faster and at about half the cost.

In the Artificial Analysis Coding Agent Index, Sol outperformed Fable 5 by 2.8 points, reduced token output by more than 50 %, halved execution time, and lowered cost by roughly one‑third.

Real‑World Task Comparison

Task 1 – WAIC Conference Preparation

ChatGPT Work took 18 min 50 s and delivered an interactive card with an aggressive “skip list” and interview scripts. Claude Cowork finished in 6 min, providing a markdown document covering more venues. Both missed some information; ChatGPT’s output was more ready‑to‑use.

Task 2 – Luo Yonghao User Guide

ChatGPT Work produced a styled web page resembling a commercial feature article, complete with citations, printable layout, and a checklist. Claude delivered a media‑style card with emojis and a “danger zone” list. The ChatGPT version was more polished and functional for direct deployment.

Task 3 – Comparative PPT

ChatGPT Work generated a 16‑page PPT that highlighted conclusions and noted that Work had been available for only one day. Claude produced a 12‑page research‑style report with market data. Both disclosed limitations; overall, ChatGPT achieved higher completion quality but used more tokens and was slower.

Current Limitations

Quota consumption can quickly exhaust a subscription because background agentic usage drains tokens faster than ordinary chat.

The three‑in‑one redesign hides the legacy chat interface as “ChatGPT Classic,” which may inconvenience users who only want simple chat.

Cross‑device continuity is stronger in Claude Cowork: Work’s desktop and web sessions do not share state, and local threads remain on the originating device.

Clarifying Chat, Work, and Codex

Chat answers questions, Work delivers finished products, Codex modifies code.

Chat handles Q&A and content generation within the current conversation and any user‑provided material.

Work targets knowledge work across applications, reading email, documents, calendars, and business systems, and delivering spreadsheets, documents, PPTs, or web apps. It autonomously gathers context, integrates information, and drives an entire workflow.

Codex operates on code repositories, reading project files and Git context, and returning diffs, test results, or pull requests.

Quota and Pricing Details

Chat has separate limits for messages, images, and voice. Work and Codex share an “agentic usage” pool. All subscription plans—from Free to Enterprise—include access to Work and Codex; higher tiers provide larger token pools and more capable model variants.

Free: limited Work/Codex quota.

Go: low‑price upgrade.

Plus: ad‑free, full feature set.

Pro: higher quota for power users.

Business/Enterprise: adds SSO, compliance, and granular controls.

Quota is measured in token consumption; more powerful models and richer agentic modes consume tokens faster. Plus and Pro users can purchase extra credits once their pool is exhausted.

Choosing Between Sol, Terra, and Luna

Sol: strongest capability, suited for complex reasoning and heavy coding.

Terra: balanced performance and cost, recommended as the default.

Luna: fastest and cheapest, ideal for simple or high‑volume tasks.

The decision should match the task’s value to the appropriate cost tier rather than simply selecting the strongest model.

Implications of the Update

OpenAI is bundling model improvements, agent capabilities, a unified desktop entry point, and a plugin ecosystem into a more complete work system. GPT‑5.6 emphasizes per‑dollar performance, Work orchestrates cross‑application knowledge work, and Codex handles development tasks. This marks a shift from chat‑only assistance toward autonomous agents that can read/write files, integrate across systems, sustain long‑running tasks, and deliver finished results.

Code example

@