Anthropic’s Claude Cowork has beaten Google’s Gemini in a hands-on productivity test that mirrors a task office workers face daily: sifting through a cluttered Gmail inbox to find specific information and assemble it into a polished summary. The test, conducted by ZDNet, highlights a growing divide in how well consumer AI assistants handle the messy, context-heavy reality of email research.

Inside the Gmail research challenge

ZDNet’s test tasked two leading AI assistants—Claude Cowork and Gemini—with digging through a real Gmail account to locate particular pitches, pull out relevant quotes, and confirm whether permission had been granted for certain content use. It was not a synthetic benchmark but an actual inbox with years of messages, threaded conversations, and the usual noise of receipts, newsletters, and personal notes.

Claude Cowork completed the assignment accurately, surfacing the requested items and even summarizing the permission status in plain language. Gemini struggled to deliver the same level of precision, missing key details and requiring more hand-holding. The test did not involve any deep integration with Gmail; both AIs were accessed through their standard web interfaces, with the human copying and pasting email content into the chat window.

The task sounds simple, but it demands a rare combination of abilities: maintaining context over a long conversation, understanding nuanced requests, filtering irrelevant information, and recognizing the difference between a mention and a confirmation. Claude Cowork’s retrieval-augmented generation (RAG) approach appeared to give it an edge in grounding its responses in the provided email text rather than hallucinating or glossing over specifics.

Why this matters for everyday users

If you regularly use Gmail and have experimented with AI tools to lighten the research load, this result is a signal. Not all assistants are equal when the work gets granular. For anyone who has wasted an afternoon scrolling through email threads to find a single attachment or approval, the promise of an AI that actually gets it right is enormous.

For individual users, Claude Cowork currently requires a Pro subscription ($20/month) and runs as a separate web app. It does not hook into Gmail automatically—you feed it the emails manually. That overhead limits immediate convenience but also gives you control over what data you share. Gemini is available in a similar freemium model, and Google has teased deeper Workspace integrations that might one day let it access Gmail directly. For now, though, the test suggests those integrations might not mean much if the underlying model can’t perform the task.

Power users and IT professionals should pay attention to the security and privacy models. Feeding sensitive email content into any external AI raises compliance questions. Claude Cowork’s design philosophy around safety and minimizing harmful outputs is well-documented, but the data you hand over still leaves your machine for Anthropic’s servers. Enterprises considering team-wide deployment will need to weigh the accuracy gains against the lack of native Gmail integration and potential data governance risks.

The competitive landscape that shaped this moment

The showdown between Claude Cowork and Gemini didn’t happen in a vacuum. Over the last eighteen months, AI assistants have evolved from novelty chatbots to serious productivity tools. Anthropic launched Claude 2 in July 2023 with a 100k token context window, then followed with Claude 2.1 and the multi-model Claude 3 family (Haiku, Sonnet, Opus) in early 2024. Claude Cowork, a feature layered on top of these models, leans into project-based organization and persistent memory.

Google’s path has been shakier. Bard debuted in March 2023 to lukewarm reviews, and the rebrand to Gemini in February 2024 came with promises of a unified model family (Ultra, Pro, Nano) that would power everything from cloud applications to on-device tasks. Gemini for Google Workspace was announced as a premium add-on, aiming to let users query their own email, documents, and calendar. However, ZDNet’s test suggests the current public-facing model still falls short in the kinds of open-ended, research-intensive assignments that knowledge workers actually care about.

Earlier this year, independent evaluations began to paint a consistent picture. A February 2024 analysis by Vellum found Claude 3 Opus outperformed GPT-4 and Gemini Ultra in retrieval tasks. A March study by Galileo highlighted Claude’s strength in long-context reasoning. The ZDNet Gmail test is notable not because it’s the only data point, but because it translates those synthetic benchmarks into a scenario that resonates with anyone who has felt the pain of email overload.

What you should do right now

If you’re curious whether an AI research assistant could save you hours each week, the best step is a low-stakes experiment. Sign up for a Claude Pro account (or use the free tier with lower limits), copy a few representative email threads into the chat, and ask it to perform tasks like:

  • Find the date and sender of the latest pricing proposal.
  • Summarize the three key action items from this project thread.
  • Check if the client explicitly approved the design changes in any of these emails.

Compare the output against your own manual review. If Claude Cowork consistently delivers accurate answers, it’s worth weaving into your routine—perhaps by batching email research every morning and pasting the relevant threads into a dedicated Claude Cowork project. For Gmail power users, this might still be faster than writing complex search operators to unearth the same information.

For those invested in the Google ecosystem, try the same tasks with Gemini. The goal isn’t to pick a winner forever but to see which tool understands your information landscape better right now. Google is iterating quickly, and deeper Gmail integration—possibly via the Gemini side panel already rolling out to some Workspace testers—could significantly change the equation by reducing the friction of copying and pasting.

Administrators should monitor these developments without rushing into deployment. Neither Claude Cowork nor Gemini offers an enterprise-ready vault solution that guarantees data residency within your own tenant. For regulated industries, the manual cut-and-paste approach may be a non-starter regardless of accuracy. The more pressing action is to draft an acceptable use policy for AI tools that addresses email content sharing, and to educate staff on what kinds of queries are safe versus those that might leak proprietary information.

What comes next

The ZDNet test outcome is a snapshot, not a final verdict. Anthropic has signaled that a major Claude 4 update is approaching, likely with an even larger context window and more reliable retrieval. Google is expected to announce new Gemini capabilities at I/O in May, with rumors of a revamped Workspace integration that could finally connect the model directly to your inbox.

Both companies are also racing to add agent-like features that would let the AI not just answer questions but actually take actions—like drafting replies, forwarding messages, or updating CRM entries. When those arrive, the accuracy of the underlying research becomes even more critical. An error in summarization today is an annoyance; an error in an autonomous email sent on your behalf could be a career-limiting move.

For now, the practical takeaway is straightforward: if your daily routine involves mining Gmail for specific information, Claude Cowork is worth the subscription and the few extra seconds of copy-pasting. It’s not a seamless solution, but it works where a supposedly more integrated rival currently doesn’t. The next time a manager asks you to “pull together all the client feedback from the last quarter,” you might find the real assistant isn’t sitting in the next cubicle.