Contrary Research argues in a report by Michael Kim and Claire Schultz that the benchmark scores AI labs use to sell their models have become the thing those models are trained to chase. That shift, the authors say, changes what a high score can tell a buyer. The report is analysis. The authors did not run these tests themselves; they draw on audits by OpenAI, METR, FutureHouse and others.
Their central claim is narrow. For most of AI’s history, benchmarks stayed honest because models fell well short of them, and scores rose as a by-product of general progress. The authors say that has ended. Recent gains cluster in the areas labs deliberately go after, because today’s models improve mostly through reinforcement learning: a model tries a task over and over, receives a score for each try, and is nudged toward the best-scoring attempts. The report calls this loop hill climbing. A flawed benchmark then does more than mismeasure progress. It points the climb in the wrong direction.
The idea has a pedigree. In 1975, Charles Goodhart, who served the Bank of England as Chief Adviser, observed that “any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.” The everyday version is that a score people aim at stops measuring what it once did. Contrary Research argues AI has now hit that wall at industrial scale.
The report’s main case study is Arena (formerly LMArena), where users vote between two anonymous chatbot answers and the votes become an Elo rating. In April 2025, Meta announced Llama 4 Maverick at an Elo of 1417, second on the board. Users soon noticed that the version tested was an experimental chat variant tuned for conversation, a detail left to a chart footnote. When Arena retested the public release, it ranked 32nd. Mark Zuckerberg said it was “relatively easy” for Meta’s team to tune a version to land at the top. Researchers separately reported that, before launch, Meta ran 27 unreleased Llama 4 variants through private trials on Arena. Meta denied rumors that it trained on benchmark test sets.
The report adds that Arena still lets labs test unlimited variants in private and withdraw the ones they dislike. Meta is second on Arena’s text leaderboard again as of October 2026, behind only Anthropic. Contrary Research also notes that more than $225 million traded on Polymarket markets that name the “best” AI model by Arena rank, between February 2025 and October 2026.
Other failures follow a pattern. OpenAI reported in February 2026 that SWE-bench Verified no longer measured real coding ability, because models had met its open-source repositories in training and could recite the fixes. Poolside found a leftover git history had let an agent mine reference solutions; after the gap was closed in September 2026, GLM-5.2’s SWE-Bench Pro score fell from 78.8 percent to 57.3 percent. The report also says Google’s Gemini 4 Argon drew accusations of “benchmaxxing,” meaning tuning a model to look good on tests at the cost of being useful, after Google employees reportedly found it still struggled on some coding tasks. The report does not say Google confirmed this.
The money explains the pressure. The report says the largest tech companies plan to spend more than $700 billion on capital expenditure in 2026, and later puts guidance from Microsoft, Google, Amazon, and Meta at $720 billion to $745 billion combined. That is a plan, not money out the door. Enterprise buyers, the authors add, use benchmark rankings to shortlist models before testing them, in a market where businesses spent $12.5 billion on foundation model APIs in 2025.
Good tests are costly, which is the authors’ other worry. Their own back-of-the-envelope estimate puts a 100-task clinical benchmark written by physicians at about $800,000, rising to $1 million with yearly refreshes. Their remedy is expert-built, private tasks checked against live usage and expert preference data, with Surge AI’s Hemingway-bench and Cursor’s CursorBench 4.0 as early examples.
At those prices, few buyers will build a better public scoreboard. A private set of a few dozen tasks drawn from your own work is the cheaper way to check any vendor’s ranking before you sign.
Contrary Research, “AI Benchmarks: Climbing the Wrong Hills,” by Michael Kim and Claire Schultz, surfaced 19 October 2026.
