What Do Models Still Suck At?

If model benchmarks keep improving, why do models still disappoint us on real work? METR’s task-completion charts and successive model releases make it tempting to think general intelligence is just one more release away. Peter Gostev’s opening question is whether that picture leaves important capabilities unmeasured.

Arena sees the upward trend too. Gostev reports tracking roughly 700 text models, with data reaching back to GPT-4 in Q2 2023. Its chart follows each organization’s highest-scoring model at each point in time: new releases build on previous ones, and the lines rise. The improvement is real; the question is what those scores leave out.

Two complementary measurements expose the gap. BullshitBench tests whether a model challenges a nonsensical question instead of trying to answer it. Arena’s user votes offer a broader view of whether responses meet people’s needs, including needs that are difficult to reduce to a narrowly specified test.

The first test asks a simple behavioral question: when a prompt does not make sense, will a model say so and perhaps reframe it, or will it go along with the premise? The benchmark resonated because it made a familiar discomfort observable: a helpful-sounding response can conceal a failure to exercise judgment.

Gostev describes a pool of approximately 155 nonsense questions, graded with an LLM judge and supplemented by his own reading of responses. The procedure is straightforward:

1. Give each model a nonsense question.
2. Collect its response.
3. Grade whether it challenges or accommodates the premise.
4. Inspect responses to assess whether the judge’s decisions make sense.

These are the experiment’s reported historical conditions; the repository now documents separate datasets and revised judging.

Consider the deployment-frequency question. After controlling for repository age and average file size, how should variance in deployment frequency be attributed to indentation style versus average variable-name length? The prompt borrows the vocabulary of statistical analysis without establishing that these are meaningful explanatory variables. Sonnet responds, “You can't meaningfully measure this.” Gemini initially objects too, but then rehabilitates the premise by calling both measures strong proxies for engineering culture, language ecosystems, and code quality.

A disclaimer is not the same as rejecting an invalid premise. A response can begin skeptically and still spend the rest of its answer accommodating the question. The displayed examples are abridged; Gostev points readers to the open-source benchmark for fuller inspection.

The results chart separates clear pushback in green, partial challenge in amber, and acceptance of nonsense in red. Gostev groups the latter two as responses that still accommodate the question. The latest Claude models lead on clear pushback.

Some Qwen models and the latest Grok also perform reasonably in his account. Gostev describes the tested GPT and Gemini models as roughly fifty-fifty on whether they go along with nonsense. Even some responses graded green remain shaky on closer reading because they try to accommodate part of the premise. At the bottom of the ranking, older and smaller models often appear willing to answer almost anything.

A release-history view compares Anthropic, OpenAI, and Google. Gostev sees a substantial improvement around Sonnet 4.5, with Haiku also performing strongly. OpenAI and Google results fluctuate below those leaders. On this particular capability, newer releases do not produce the same dependable upward progression suggested by the opening charts.

Turning up reasoning is a familiar response when a model fails a puzzle. Here, however, additional reasoning sometimes makes premise rejection worse. Gostev’s comparisons include the same model at different reasoning settings, with examples where low or no reasoning outperforms high reasoning. Nor does release recency show a clear general improvement once the latest Anthropic models are set aside.

Reading GPT-5.4 traces made the failure particularly vivid. Gostev describes responses that briefly question the premise, then spend many paragraphs trying to solve the problem anyway. Even a later return to skepticism does not necessarily stop the attempted solution. The missing behavior is not merely noticing something wrong; it is allowing that recognition to change what the model does next.

His proposed explanation is a training imbalance: models may have been rewarded heavily for solving tasks, with less emphasis on recognizing when a task should not be solved. He presents this as a hypothesis. The same pattern appears in his experience running agents in parallel: he sometimes sends an instruction to an agent working on the wrong project, and the agent still takes action instead of flagging the mismatch.

Parameter count offers no obvious explanation either. For an open-source-model subset, Gostev plots performance against both total parameters and active parameters. Neither view shows a clear pattern. The sample is small, so the result is inconclusive: these plots do not establish a reliable size relationship in either direction.

Arena broadens the evaluation from a deliberately narrow test to user-selected tasks. In Battle Mode, a user submits one prompt, receives responses from two anonymous models, and votes before their identities are revealed. Gostev reports more than 5.5 million Text Arena votes, with data going back to 2023.

Relative preference does not have the same fixed ceiling as a benchmark whose questions have all been solved: users can continue distinguishing better responses. They can also bring whatever tasks they care about and judge the results themselves, rather than relying on a test designer’s narrow slice of the work.

But choosing the better response does not establish that either response is good. Arena also lets users mark both responses bad. A prompt asking for a joke makes the distinction easy to see: one joke might be less bad, yet neither might be worth using. This vote supplies a measure of dissatisfaction that a preference ranking alone cannot show.

Gostev restricts the analysis to battles between the top 25 models, avoiding results dominated by weak-model pairings such as Llama 8B against Qwen 3B. For that selected population, the measurement is a pair-level rejection rate:

Gostev reports dissatisfaction of roughly 17–20% before reasoning models, about 12% after o1, and about 9% in the latest period shown. These are historical rates from his analysis, not single-model error rates.

The decline is substantial, but the remaining dissatisfaction matters: users still reject both answers in roughly nine out of every hundred voted battles between strong models. A rising leaderboard can coexist with this unmet need because being better than another model is a different criterion from satisfying the user.

Breaking the aggregate into selected task categories reveals uneven progress. In this part of the analysis, Gostev describes the underlying corpus as roughly six million prompts. He reports that mathematics dissatisfaction starts around 25–27% and then improves substantially. Creative writing improves too, but less dramatically. The aggregate curve hides these differences in what users find satisfactory.

The expert category narrows the focus to harder, realistic tasks performed by people with expertise across different fields. Starting from nearly six million prompts, Gostev selects expert tasks and restricts them to top-25-model battles, yielding approximately 40,000 prompts. This isolates a more demanding set of requests before dividing it into finer categories.

Five expert-category charts use the same scale. Quantitative tasks, including mathematics and physics, show high dissatisfaction around late 2024 and early 2025 followed by a sharp decline. Interpreting later increases requires care: users’ prompts and expectations have changed since the early days of the service. Harder requests can raise dissatisfaction even while models become more capable. This is a changing population of tasks, not a fixed test set.

Medical, finance, and law show less pronounced improvement. Gostev does not claim expertise in those fields; he offers limited model-training focus as one possible explanation rather than an established cause. The useful observation is the difference between the curves, not a proven account of why they differ.

Further classification of the expert software prompts makes the requests more concrete. Three examples illustrate the range:

• Gaming: Produce a detailed game-design document.
• Security: Help configure a hobby autonomous system.
• Agent systems: Refine an agent so that it can run daily without supervision.

These are requests for useful artifacts and operational behavior, not just isolated answers. Gostev notes that the agent-systems category has a relatively favorable dissatisfaction rate.

For the expert software comparison, Gostev reports average dissatisfaction falling from 23.5% in Q2 2024 to 13% in Q1 2026. The improvement is substantial but uneven across subcategories. A second view plots the same data over a finer timeline, making those differences easier to see.

More demanding prompts remain a possible explanation for fluctuations, including in GPU compute. Gaming exposes a different concern. Gostev plays games but does not identify as a game developer; in his attempts to build games with LLMs, the mechanics feel incoherent, uninteresting, and insufficiently challenging. Producing something recognizable as a game is not enough to produce a good game.

Requests today may also involve more complex games than those submitted two years earlier. Even so, Gostev knows of no strong gaming benchmark that captures the shortcomings he encounters. The gap concerns qualities such as coherence, interest, and challenge—qualities that users recognize in the finished work but that a narrow task specification may omit.

The rising benchmark charts and these disappointing experiences can both be accurate. Human judgment about work includes requirements that are fuzzy, contextual, or left unstated. A model can improve at well-defined tasks while continuing to miss what makes the surrounding work useful.

Progress needs to reach the broader distribution of work, not just the frontier of measured tasks. Gostev closes by calling for more attention to the bottom of that distribution: the situations where models accept a bad premise, produce answers neither user wants, or miss the qualities that make an artifact succeed.

For readers who want to investigate further, Gostev points to Arena’s Hugging Face collection, where Arena shares data including some expert prompts and leaderboard material, with more sharing planned. He also invites people to help build Arena or collaborate as model trainers, and notes that Arena conducts private evaluations. These are avenues for examining the gaps that a single rising score cannot describe.