Is there a better AI than ChatGPT?

The Problem: 'Which AI Is Best?' Is the Wrong Question

ChatGPT is the default answer to "which AI should I use?" — but the default is starting to wobble. Traffic data shows OpenAI.com visits fell 38.9% between August and December 2025, while ChatGPT.com traffic stayed roughly flat at around 5.2 billion visits a month, according to aggregated industry statistics. Meanwhile, the frontier has consolidated into a three-way race with no clear champion.

That consolidation is the crux. Wharton professor Ethan Mollick's analysis identifies Claude Opus, Gemini 3 Pro, and GPT-5.x as "remarkably close in overall capability," concluding that for most people, the app and harness now matter more than the model. The same model behaves very differently depending on the tooling wrapped around it — a distinction that raw "best AI" rankings completely miss.

Asking "which AI is best?" leads to bad decisions because the honest answer is task-dependent. A 30-day hands-on comparison found sharply different winners across categories:

• Claude leads coding (80.84% on SWE-bench Verified vs. roughly 74.9% for ChatGPT), plus writing, editing, and long documents
• Gemini wins real-time research and Google Workspace integration — and topped the Humanity's Last Exam benchmark at 37.2%, edging GPT-5.2's 35.4%
• ChatGPT still owns image generation, voice mode, and overall versatility
• Claude's calibration — saying "I don't know" rather than fabricating — makes it the safer pick for legal, medical, and financial work

UC Berkeley's School of Information puts it plainly: the real question is not which model is best, but which model is best for what. Benchmarks miss things, workflows differ, and hands-on testing reveals gaps that leaderboards hide.

This matters more than ever because the ground is shifting under the question itself. Enterprise AI is moving from asking models for outputs to delegating multi-step workflows, and the agent layer — Claude Code, OpenAI Codex, the Model Context Protocol — is where real differentiation now happens. That's also where new risks live: unvetted agent tooling can quietly undermine even a top-tier model, which is exactly the problem SkillzBench is building against by verifying and benchmarking agent skills before they reach your workflow.

So the research-backed answer to "is there a better AI than ChatGPT?" is yes — for specific tasks, several models beat it. But there is no single overall winner, and picking one without matching it to your actual work is how people end up paying $20 a month for the wrong tool.

The Research Verdict: Task-by-Task Winners Among the Big Three

For the first time since ChatGPT launched, benchmark leadership has changed hands — and the winner isn't OpenAI. According to recent benchmark data, Gemini 3 Pro now tops Humanity's Last Exam at 37.2%, edging out GPT-5.2 at 35.4%. Last year's best score was just 26.5%, so the entire frontier jumped dramatically — yet the narrow spread means no model dominates.

The clearer verdicts come from task-specific testing. A 30-day hands-on comparison found Claude "not even close" to its rivals on coding, posting 80.84% on SWE-bench Verified versus roughly 74.9% for ChatGPT and 65% for Gemini. Claude also calibrates its confidence better, saying "I don't know" instead of fabricating — a decisive trait for legal, medical, or financial work.

ChatGPT retains the versatility crown, though. It remains the leader in image generation, voice mode, and the most mature API and agent ecosystem, per the UC Berkeley I School's analysis. Gemini counters with real-time research strength, Google Workspace integration, and cost efficiency: at $2.00 per million input tokens versus $5.00 for GPT-5.5, it is the budget workhorse.

The task-by-task scoreboard looks like this:

• Hardest reasoning tasks: Gemini 3 Pro (37.2% on Humanity's Last Exam)
• Coding: Claude (80.84% on SWE-bench Verified)
• Versatility (voice, images, APIs): ChatGPT
• Real-time research and cost efficiency: Gemini
• Calibrated accuracy: Claude

Two caveats matter. First, as Wharton professor Ethan Mollick notes in his guide to which AI to use, the big three are "remarkably close in overall capability" — the app and harness now matter more than the model. Second, benchmarks miss things that hands-on testing catches, so treat these numbers as directional.

Privacy also splits the field. Anthropic retains data for only 7 days and never trains on it unless users actively opt in. OpenAI (up to 30 days) and Google (up to 55 days) both train on user data by default, requiring an opt-out. For sensitive workloads, that inversion alone can decide the winner.

One more shift deserves attention: differentiation is moving from models to agent harnesses — Claude Code, Codex, MCP. Mollick's warning that one open-source agent poses a "serious security risk" illustrates the trust gap in this new layer. SkillzBench is being built to close exactly that gap, verifying and benchmarking agent skills so you know what actually helps before you load it.

The Real Differentiator: Harnesses, Not Models

Here's a question that surprises most people: the same AI model can behave like two completely different products depending on the software wrapped around it. That's not a bug — it's the single most important insight for anyone trying to figure out whether something beats ChatGPT.

Ethan Mollick, the Wharton professor behind his widely-read AI guide, breaks the AI landscape into three layers: Models, Apps, and Harnesses. The models — Claude Opus 4.6, Gemini 3 Pro, GPT-5.2/5.3 — are "remarkably close in overall capability." His conclusion is blunt: for most people, the app and harness matter more than the model itself.

What does that mean in practice? A harness is the scaffolding that lets a model take actions — read files, run code, connect to tools. Claude Code, OpenAI Codex, and the Model Context Protocol (MCP) are all harnesses. The same underlying model behaves very differently depending on which one you plug it into. Mollick notes that OpenAI and Anthropic hold a "clear lead over Google" here: Gemini's model is just as good, but its agentic website is much less capable.

This matters because the whole industry is shifting from prompting to delegation. OpenAI's own enterprise report describes the next phase as "delegating complex, multi-step workflows" rather than asking models for outputs. In an agentic era, your results depend less on which chatbot you picked and more on which tools, skills, and connections you've given your agent.

But that shift carries real risk. Mollick flags OpenClaw, a popular open-source local agent, as a "serious security risk" that users "almost definitely shouldn't use." And the problem isn't isolated — published research found that roughly 26% of agent skills carry at least one vulnerability. When you hand an agent a third-party instruction package, you're trusting it with real access to your systems.

That's the gap SkillzBench is being built to close: independent verification, sandboxed benchmarking that measures each skill's actual Lift in percentage points, and trust scores that separate security signal from marketing noise. Until tooling like that matures, treat unvetted agent tooling the way you'd treat unvetted code — with skepticism.

The takeaway: stop asking "is there a better AI than ChatGPT?" and start asking "which harness fits my workflow, and can I trust what I'm plugging into it?"

How to Choose: A Practical Decision Framework

If you've been picking an AI assistant the way most people do—by default settings and brand familiarity—you're probably getting worse answers than you could. Three independent analyses, from a 30-day hands-on comparison to Ethan Mollick's agentic-era guide, converge on the same conclusion: there is no single best model, only better matches for specific tasks.

The simplest starting point is matching model to work category. The hands-on testing was blunt about coding: "For coding and software development: Claude. Not even close," with Claude scoring 80.84% on SWE-bench Verified versus roughly 74.9% for ChatGPT and 65% for Gemini, according to a 30-day comparison. Gemini wins real-time research and Google Workspace integration, while ChatGPT leads on image generation, voice mode, and overall versatility. Berkeley's analysis frames it similarly: "which model is best for what," not which model is best, period.

Second, stop accepting defaults. Mollick warns that ChatGPT's default "GPT-5.2" is actually an auto mode that often routes to a weaker model—his guide recommends manually selecting the top-tier modes: GPT-5.2 Thinking Extended, Gemini 3 Pro, or Claude Opus 4.6 with extended thinking enabled. The same model can behave very differently depending on the app and harness around it, which is where much of the real differentiation now lives.

Third, budget for serious work. Free tiers are optimized for chat rather than accuracy, and Mollick argues serious use requires paying at least $20/month—though a direct counterpoint from the hands-on testers says free tiers have "narrowed the gap." The consumer plans are nearly identical anyway: ChatGPT Plus at $20, Claude Pro at $20, and Google AI Pro at $19.99.

Fourth, weigh privacy if your work is sensitive. Berkeley's testing found Anthropic retains data for only 7 days and never uses it for training by default—users must actively opt in—while OpenAI retains data up to 30 days and Google up to 55, with users needing to opt out. For legal, medical, or financial work, that difference alone can decide the choice.

One caution before you commit: model numbers date quickly. The sources in this article reference conflicting versions—GPT-5.2 versus 5.5, Claude Opus 4.5 through 4.7, Gemini 3 versus 3.1—simply because they were published at different times. Verify current versions before making a decision.

Finally, remember that the harness matters as much as the model. Agent tooling like Claude Code, Codex, and MCP is now the differentiator layer, and unvetted tooling carries real risk—Mollick flags one popular open-source agent as a "serious security risk." That's the gap SkillzBench is being built to close: verifying, benchmarking, and scoring agent skills so you can trust what you plug into your workflow. Join the waitlist to follow the launch.

The Missing Piece: Verifying What Your Agents Actually Do

If the harness now matters more than the model, a new question follows: who is testing the harness? Ethan Mollick's widely read guide argues that the app and harness matter more than the model for most people, and even flags popular agent tooling like OpenClaw as a "serious security risk" — a warning that applies far beyond one open-source project (One Useful Thing).

The shift makes sense. Enterprise AI is moving "from asking models for outputs to delegating complex, multi-step workflows," and the tooling layer around agents — Claude Code, OpenAI's Agents SDK, Google's agent platform, and the Model Context Protocol — has become the real differentiator (OpenAI's enterprise report, UC Berkeley's comparison). Yet while models get benchmarked constantly, the skills and instructions that steer those agents get almost no scrutiny at all.

That gap is not theoretical. Published research on agent skills found that roughly 26% of skills carry at least one vulnerability — meaning the instruction packages developers bolt onto their agents are a genuine attack surface, not just a productivity boost. An agent that follows a malicious or merely sloppy SKILL.md can do real damage inside a codebase, a repo, or a production pipeline.

This is the missing piece SkillzBench is building. As an independent testing lab for AI agent skills, it is constructing a trust layer the agentic era currently lacks:

• Verification: every skill version pinned to a commit SHA, then validated, license-checked, security-scanned, and quality-judged before it earns a 0–100 trust score.
• Sandboxed benchmarking: each skill tested with the same model, with and without the skill, so the Lift is published in plain percentage points — kept strictly separate from ratings, with money never touching scores.
• Per-agent loadouts: one MCP server where users toggle skills and tools per agent, so each agent only sees what was deliberately chosen for it.
• Currency: upstream changes tracked, re-scanned, and re-benchmarked automatically, with notifications when a skill you rely on shifts underneath you.

The philosophy mirrors Berkeley's hands-on conclusion that "hands-on testing reveals what benchmarks miss." A model leaderboard tells you Claude scores 80.84% on SWE-bench Verified; it tells you nothing about whether a given community skill improves your agent or quietly undermines it. SkillzBench's launch scope covers instruction-only skills (SKILL.md plus text references), with script-bearing skills listed link-out only — a deliberately conservative start for a category where the risks are real.

The question "is there a better AI than ChatGPT?" turns out to have an unexpected answer: the better AI is the one with the better harness — and the better harness is the one with skills you can actually trust. SkillzBench is pre-launch and actively being built. Join the waitlist at skillzbench.com to get early access when verification and benchmarking go live.