Every major AI lab shipped a frontier model in roughly the same eight-week window this summer. Claude Sonnet 5 landed June 30, 2026. Grok 4.5 followed on July 8. GPT-5.6 went generally available July 9. Gemini 3.1 Pro has been iterating through the same stretch. When four capable labs are all releasing within days of each other, "which model is smartest" stops being the useful question, since the honest answer changes weekly and none of the gaps are large enough to matter for most real work. The more useful question is what a given product actually lets you do with that intelligence once you have it, and that's where ChatGPT Work is worth a real look rather than a passing mention.
Powered by GPT-5.6, ChatGPT Work is built to take a goal, not just a question, and turn it into a finished spreadsheet, a working dashboard, or a completed research report, pulling context from your actual files and tools rather than waiting for you to paste everything in yourself.
This article walks through what specifically earns that reputation, where the underlying models genuinely hold up against the competition, and where the honest limits are.
Where ChatGPT's Models Actually Outshine the Competition
GPT-5.6, the model family powering ChatGPT Work, made an unusual packaging decision when it shipped: instead of one flagship model with an effort dial, OpenAI split it into three named, durable tiers:
- Sol for the hardest agentic and coding work
- Terra as a balanced option priced at roughly half of Sol's rate
- Luna for fast, cheap, high-volume tasks
That's a genuinely different strategy than shipping one model and letting cost scale with a reasoning-effort slider, and it means a team can route routine work to Luna and save Sol's cost for the problems that actually need it, without switching products.
On raw capability, the honest picture is that GPT-5.6 doesn't sweep every benchmark, and it's worth saying that plainly rather than pretending otherwise. On Terminal-Bench 2.1, an agentic coding benchmark, Sol scored 88.8% in its standard mode and 91.9% in its higher-compute Ultra mode, edging out both GPT-5.5 and Claude Mythos 5's 88.0%. But Claude Fable 5, Anthropic's premium Mythos-tier flagship, actually beats Sol on SWE-Bench Pro (80% versus 64.6%) and the Artificial Analysis Intelligence Index. What GPT-5.6 wins on instead is the trade that matters for most real teams doing real work: Fable 5 costs \$10 input and \$50 output per million tokens, double Sol's rate, while OpenAI's own benchmarks suggest Sol reaches comparable or better results on many agentic and coding tasks using meaningfully fewer tokens and less time to get there. That's not a benchmark-chart win; it's a genuine "good enough, dramatically cheaper, and much faster" bet, and for the volume of work most teams actually run through these models, that bet is a real, defensible strength rather than a consolation prize.
There's a second, less-discussed strength worth mentioning on the model side: OpenAI is one of the few frontier labs that also ships open-weight models a team can run entirely on its own infrastructure. gpt-oss-120b and gpt-oss-20b, released under the permissive Apache 2.0 license, were OpenAI's first open-weight language models since GPT-2, and they're built specifically for teams that want data residency guarantees, full customization through their own fine-tuning pipeline, or a model that runs on common inference stacks like vLLM, Ollama, or llama.cpp without touching OpenAI's API at all.
It's worth being precise about what this is and isn't: these weights aren't served through the OpenAI API and don't appear inside ChatGPT itself; they're a separate, self-hosted option. But their existence still matters for the overall picture — it means a team evaluating OpenAI isn't only choosing between "use our cloud" and "use nothing," which is a genuinely different position than most closed frontier labs offer.
What You Can Actually Do Inside the ChatGPT Interface
This is where ChatGPT Work stops being a spec sheet and starts being something people describe as changing their actual jobs. A few concrete pieces of the interface are doing most of that work.
- Plan mode is the first thing worth understanding, because it changes the whole shape of how a session unfolds. Rather than diving straight into execution, ChatGPT gathers context, asks clarifying questions, and produces a step-by-step plan you can adjust or approve before anything actually happens. That single design choice is what makes handing over a multi-hour project feel less like a gamble; you see the approach before it runs, not after.
- Sites turns a prompt into something you wouldn't expect from a chat window at all: a live, interactive webpage. Dashboards, project trackers, launch calendars, prototypes — built from a conversation and kept up to date as the underlying information changes, rather than a static export that goes stale the moment you close the tab. Paired with the desktop app's built-in browser, which supports multiple tabs and lets ChatGPT work across your files and accounts directly, this is a genuinely different interaction model than typing a question and copying an answer into another program.
The part that's hardest to fake is what actual teams report doing with it. At Zapier, a lead-triage process that used to take 35 to 45 minutes per lead across HubSpot, Gong, and email turned into an automated QA system that traces every lead's journey and surfaces drop-offs — work that now identifies seven figures in pipeline handed to sales every month, according to Zapier's own Head of Enterprise Marketing.
At NVIDIA, a Go-to-Market Manager described roughly 40% of their time going to manual number-crunching before GTC events, now automated into a workflow that runs twice a week, freeing that time for actual strategy work with the field team.
Shopify's Lead for Applied AI and Enablement described using it as a daily operating layer, pulling context from Slack into a running "second brain" and coordinating a research program across 3,500 non-R&D employees.
Those are named people, at named companies, describing specific before-and-after numbers, not anonymous marketing copy — which is worth more than any benchmark score for judging whether a tool actually changes how work gets done.
Underneath a lot of that automation sits a feature worth understanding on its own: Scheduled Tasks.
Rebuilt with a dedicated Scheduled page in mid-2026, it lets you turn a one-off request into something that runs on its own: a daily briefing, a recurring status report, a monitoring task that checks for a change and only notifies you when something's actually worth seeing. Since the same update, a scheduled task can use the same tools an interactive chat can, including live web search and connected apps like Gmail, which is what turns "remind me about X" into something closer to a standing employee than a calendar alert.
It's worth knowing the real limits rather than assuming it's unbounded: tasks can't run more than once an hour, and the number you can have active at once scales with plan — 3 for Go, 5 for Plus, up to 15 for Pro, Business, and Enterprise. That's a sensible, disclosed constraint rather than a hidden one, and it's the kind of detail worth checking before building a workflow that assumes more headroom than the plan actually provides.
MCP, Agents, and Tool Calling: How It Actually Connects to Your Work
None of the above works without a real way to connect ChatGPT to the tools a team already uses, and this is a place where OpenAI made a genuinely collaborative choice rather than building a closed alternative. OpenAI adopted the Model Context Protocol (MCP) across its products in March 2025, the same open standard Anthropic created, and in December 2025 Anthropic handed MCP's governance to a vendor-neutral foundation, meaning it's genuinely shared infrastructure now rather than one company's walled garden. The same MCP server can serve ChatGPT, Claude, and anything else that speaks the protocol, which means a team's investment in building one internal integration isn't locked to a single AI vendor.
In practice, this shows up in ChatGPT as Developer Mode for connecting remote MCP servers on paid individual plans, and workspace-published MCP apps for Business, Enterprise, and Edu accounts, with full support for write actions — not just read access — rolling out to those workspace tiers. The adoption curve since OpenAI's Apps SDK shipped has been fast by any standard: more than 35 enterprise software vendors launched ChatGPT apps or MCP integrations within 60 days of the release, including Salesforce, Box, Dropbox, Atlassian, and Adobe. Combined with the more than 1,400 plugins available directly inside ChatGPT Work for pulling context from existing workflows, the practical effect is that "connect ChatGPT to the tool my team already uses" is a solved problem for a large and fast-growing share of common business software — not a custom integration project.
How ChatGPT Gets Its Data, and What's Free vs. What Isn't
A model's training data has a cutoff. Real work doesn't wait for the next training run, which is why live web access matters as much as raw model quality. ChatGPT's browsing capability lets it search and read the current web mid-conversation, pulling in today's information rather than answering from a frozen snapshot, and Agent Mode extends that further into taking multi-step actions — browsing, running code, and calling tools in sequence — entirely within one session.
On the free tier specifically, it's worth being precise rather than repeating "it's free" as if that settles the question. As of this writing, free users get roughly 10 messages every 5 hours on the default Instant model before the conversation automatically drops to a lighter, mini fallback model until the window resets. Plus, at \$20 a month, lifts that considerably — around 160 messages every 3 hours on the standard model, plus a separate weekly allowance of up to roughly 3,000 messages on the dedicated Thinking reasoning model. Even at the top, Pro and Enterprise plans are described as effectively unlimited on standard usage, but every tier — including the most expensive ones — still operates within fair-use guardrails rather than a genuinely uncapped ceiling. That's a real, worthwhile distinction for anyone deciding whether the free tier covers their actual usage pattern or whether the heavier, ChatGPT Work-specific features are worth paying for.
Comparison Table
|
GPT-5.6 (ChatGPT Work) |
Claude Sonnet 5 |
Gemini 3.1 Pro |
Grok 4.5 |
| Release date |
July 9, 2026 |
June 30, 2026 |
Rolling 2026 updates |
July 8, 2026 |
| Pricing (per million tokens) |
Sol \$5/\$30, Terra \$2.50/\$15, Luna \$1/\$6 |
\$2/\$10 intro through Aug 31, 2026, then \$3/\$15 |
Varies by access path |
\$2/\$6 |
| Context window |
1.05M tokens |
1M tokens |
1M input tokens, 65K output |
500K tokens |
| Task Management |
|
|
|
|
| Standout strength |
Agentic execution, tiered cost/speed trade-off, MCP and plugin ecosystem breadth |
Strong in-repo coding value at intro pricing |
Long-context and multimodal workflows |
Cost-efficient coding, tight Cursor integration |
This table was sourced from three different sources: EdenAI's GPT-5.6 benchmark guide, Falconer's July 2026 model comparison, and Merge's Sonnet 5 vs. Grok 4.5 breakdown.
Wrapping Up
The honest case for ChatGPT Work isn't that GPT-5.6 wins every benchmark against Claude, Gemini, and Grok; it doesn't, and pretending otherwise wouldn't hold up past the first fact-check. The real case is narrower and better supported: a genuinely useful cost-and-speed trade-off on the model side, a Plan-mode-and-Sites interface built around finishing real work rather than just answering questions, a fast-growing, standards-based MCP integration ecosystem instead of a walled garden, and enough named, attributed customer results to suggest the "turns scattered work into finished output" pitch isn't just a slide in a launch deck. That's a specific, earned answer to what ChatGPT is actually good at, and it's worth more than a claim to being the smartest model in the room — which, this particular summer, nobody gets to hold onto for very long anyway.
Shittu Olumide is a software engineer and technical writer passionate about leveraging cutting-edge technologies to craft compelling narratives, with a keen eye for detail and a knack for simplifying complex concepts. You can also find Shittu on Twitter.
Instagram Engineering
Shopify Engineering