Best AI for coding

Why Opus 5 leads

Two factors operate at once. First, on WebDev Arena, which measures real web-development work with tools across multiple execution steps, the claude-opus-5-max row holds first place at 1692, ahead of Kimi K3 Max at 1674 and Qwen3.8 Max at 1670. Second, its output cost is $25, half of Fable 5, which itself scores lower at 1628. A model with both higher throughput and lower cost lands first under any reasonable weighting configuration.

Where this table falls short

The primary criterion here is a human-preference leaderboard, not an automated benchmark. WebDev Arena publishes an absolute score, a date, and a vote count, which is what publication on this page requires, but human preference is also sensitive to readability and answer shape, not only to whether the code is correct. SWE-bench Verified would be a sharper instrument, but its leaderboard renders in JavaScript and we have not read its figures directly. A number we have not read is not published on this page, even where another source has quoted it.

One candidate on this page still carries no rank: Gemini 3.1 Pro. It holds a real score on the general text leaderboard, but its WebDev Arena row has not been read, and Google has not published a context window for it at all, leaving more than half the criteria weight empty for it. Its name is listed below the table; its rank is not. GLM-5.2 Max and DeepSeek V4 Flash sat in the same state for a while, and now that their price and context window are verified from their own documentation, they carry a rank.

One more point, so the numbers do not read as inconsistent: the scores are relative. Every column is normalized between its own minimum and maximum, so onboarding a new candidate shifts every row score without any model having changed. Adding Grok 4.5 on 12 August 2026 produced exactly that effect: the order held and the numbers rose. If today figure does not match a screenshot from last week, this mechanism is the reason, not manual intervention.

Two more data changes shifted the scores the same day. First, the text leaderboard was re-read in full this time, and it turned out the GLM-5.2 Max and DeepSeek V4 Flash rows had been there from the start; the earlier pass had only read the top of the page. The general-text column was filled in for both. Second, GPT-5.6 Sol entered as an eighth candidate: openai.com still does not respond to our infrastructure requests, but its developer documentation publishes the price and context window, and its leaderboard row exists. One note on that row, since this is the dedicated coding page: it is named gpt-5.6-sol-xhigh (codex-harness), meaning the score belongs to the model inside OpenAI own coding agent, not to the bare model. GPT landed between Kimi and Qwen, and no other row order changed relative to the rest.

28 August: DeepSeek pricing rose and three rows shifted

This is the only number in this reference that has risen to date. Through 12 August, DeepSeek V4 Flash cost $0.14 in and $0.28 out. Its official pricing page now carries two columns, peak and off-peak, and even the cheaper column sits above the old price: $0.22 and $0.66 off-peak, $0.44 and $1.32 at peak. We record the peak column, because DeepSeek peak window overlaps the middle of the Iranian working day.

DeepSeek itself did not lose a rank from this change, since it remains the cheapest row in the table and still holds the value column. But its margin over the rest narrowed, and because every column is normalized between its own minimum and maximum, that narrowing lifted three other rows: Qwen3.8 Max from fifth to third, Kimi K3 Max from third to fourth, and GPT-5.6 Sol from fourth to fifth. None of these three models changed anything in their own specification. This is the operational definition of a "relative ranking," and it is why the arithmetic is published on this page.

What only shows up in daily deployment

The three top leaderboard scores in this table sit inside a 22-point band, 1692, 1674, and 1670, a spread that is felt far less in practice than the number suggests. What actually changes a working day is tooling infrastructure maturity: whether the model has a command-line tool, supports batch execution, and is deployed on a cloud your organization already has under contract. That is why this criterion carries fifteen weight points, not five.