Shocking Twist: ‘Best AI’ Isn’t Best

Hand holding digital AI and ChatGPT graphics.

Amid soaring claims and clashing tests, Claude Opus 4.6 posted record scores that could reshape which AI people trust for real work.

Story Highlights

Claude Opus 4.6 set new highs on key coding and research benchmarks.
ChatGPT leads on code snippets, tools, and multimodal features users rely on day to day.
Reviewers split on “best” because models excel at different jobs, not one job for all.
Benchmark flaws and marketing hype fuel confusion that benefits big platforms over users.

What the headline numbers actually say

Anthropic reported that Claude Opus 4.6 reached 65.4 percent on Terminal-Bench 2.0, the top score recorded on this agentic coding test. The company also cited a lead of about 144 Elo on an economics and law workload called GDPval-AA over OpenAI’s GPT-5.2. Claude posted 84.0 percent on BrowseComp for hard-to-find web research and 68.8 percent on ARC-AGI 2 for novel problem solving, a sizable edge over GPT-5.2 on that test.

Independent summaries back several of those gains and add practical context. Analysts highlighted Claude’s one million token context window for handling very long documents and its larger maximum output for extended drafts. Reviewers also praised stronger persistence on multi-step tasks and higher quality code on complex problems. A popular creator said Claude wrote with a more natural tone and cleaner formatting than rivals in his two-week trials.

Where rivals hit back

OpenAI’s ChatGPT, powered by GPT-5.4 in many tests, still leads on single-function code generation with a higher HumanEval score and shows stronger results on autonomous computer-use tasks like OSWorld. ChatGPT also brings built-in image generation, advanced voice features, web search, and broad app integrations that many users want. Several roundups rate GPT-5.4 the overall best across mixed benchmarks and everyday workflows.

Reviewers also point to trade-offs for Claude. Claude lacks native image generation, ranks weaker on some multimodal and visual reasoning tasks, and can respond more slowly than peers. Users report stricter usage caps on lower-cost plans and a complex price ladder, with higher-limit tiers often landing at enterprise-level costs. Those limits can block long sessions, even when writing or coding at length.

Why smart people disagree on “best”

Different tests reward different skills. Claude shines when tasks span many steps, require long memory, or involve deep research chains. ChatGPT shines when users need speed, reliable tools, and rich media in one place. Gemini gains from tight placement in Google products that many offices already use daily. These strengths reflect each company’s incentives and ecosystems as much as raw model ability, which steers how “best” gets defined.

At the same time, benchmark culture has deep flaws. Studies warn that many leaderboards are easy to game, hard to reproduce, or not statistically sound. Some tests may leak into training sets. This mess lets marketers cherry-pick big wins while dodging weaker areas. That cycle keeps buyers guessing and can lock them into large platforms that trade clarity for convenience.

How to decide for real work

Start with your job to be done, not the brand. Choose Claude for complex, multi-step knowledge work, very long documents, and deep research that must stay coherent end to end. Choose ChatGPT for faster turnarounds, strong coding assistants for smaller units, voice and images in one app, and wide integrations. Use Gemini if your team lives in Gmail, Docs, Drive, and Chrome and needs native support across that stack.

Google has some of the best AI researchers, but why they are still lagging behind Claude and ChatGPT?

Gemini Banana sucks in comparison to ChatGPT now.

— AMIT P. KUMAR (@iamitpkumar) July 14, 2026

Ask for proof, not promos. For high-stakes work, demand side-by-side trials with your own prompts, long-context recall checks, and raw logs of success and failure. Push vendors to document limits, output caps, and pricing triggers. The shared concern on left and right is fair here: when the scoreboard is murky, the biggest players win by default while users pay more for less clarity. Better tests and transparent data put you back in charge.