The smallest Claude out-engineered the biggest GPT — judgment, not size, decided it

What Surge AI found

Surge AI evaluated five frontier models on 60 tasks from its sudo L7 software engineering benchmark, and Claude models outscored GPT models across all eight dimensions. The gap was not about writing code that runs. But Claude models were better at the parts of engineering that surround the code: identifying security risks, questioning flawed specifications, and communicating potential system consequences, according to Surge AI. The finding suggests that effective AI engineering agents need judgment beyond simple code implementation. m:…6091

The examples make the point concrete. Claude Haiku 5.5, the smallest Claude in the test, identified a security risk in a B2B codebase task that GPT-6.1 Sol missed. GPT Sol also failed to identify that a requested configuration change would cause infrastructure destruction and service interruption. On another task, Claude models correctly identified that caching customer account information, including balances, would lead to stale data issues, while GPT models missed that risk. And Claude models write roughly ten times longer final reports than GPT models, surfacing risks and caveats that GPT models omit. m:…6091

The price of judgment

The cheapest model per task was not the cheapest to run. GPT Sol costs $0.91 per task, while Claude Haiku costs $1.68 per task, but Surge AI's argument is that the cost-efficiency of cheaper models is offset by the potential for expensive errors, such as production incidents or data exposure, which require manual review to mitigate. GPT models also produce roughly four times as many inaccurate statements per word as Claude models when normalized by response length. A model that is cheap per task and wrong in ways that need a human to catch is not cheap. m:…6091

There is also a cheaper way to run the better-judgment model. Larridin's CTO reduced Claude Code token costs by 10 to 15 percent by changing one configuration setting, after Anthropic reduced the price of Claude Fable 5.1 cache reads by 75 percent, to $0.25 per million tokens. The largest component of a Claude Code bill is the model re-reading existing context, and cache reads cost $0.25 per million tokens while cache misses cost $10 per million. Setting promptCacheTtl to one hour keeps the main conversation cache alive longer, and the setting requires Claude Code v2.1.242 or later. The default cache lifetime for API key usage is five minutes, which is why sessions paused for a break were paying to rebuild context. m:…5316

Judgment is not one skill

The honest complication is that writing code and reviewing it are different jobs, and the rankings flip. Pydantic's account of coding agent traces notes that review-fix and review phases account for approximately 60 percent of total software factory pipeline costs, which is why tracing matters: coding harnesses are often opaque, hiding model invocations and tool calls, and traces let teams diagnose failures, branch turns for apples-to-apples experiments, and let agents reuse past session context to reduce redundant exploration. In reviewer experiments described in the same piece, Codex with gpt-6-luna outperformed Cursor and Claude Code, scoring 3.93 against 2.80 and 2.60. The model that writes with the best judgment is not necessarily the one that reviews best. m:…5094

For teams choosing a coding model, the practical read is to stop ranking models on one axis. Surge AI's benchmark says Claude is the safer pair of hands for writing and changing production code, because it questions the request and reports what could go wrong. The reviewer numbers say Codex deserves the review seat. And the caching story says the bill for either choice depends heavily on configuration, not just the model's sticker price. The teams that will get the most from these models are the ones that measure judgment, review quality and cost separately, and assign each job to the model that wins it. m:…6091 m:…5094 m:…5316