Best AI for coding

Why Fable 5.1 leads, and why price did not stop it this time

One number settled it. In the 2 September update of WebDev Arena the claude-fable-5.1-max row sits first at 1765, and the next row we carry, claude-opus-5-max, comes in at 1687. Seventy-eight points, in a column whose whole top end fitted inside twenty-two points a week ago.

And this model loses the value column outright. Its output costs $50 per million tokens, twice Opus 5, and it scores zero there. Every previous time, that column has been what knocked an expensive model down. This is the first time a score gap has outweighed a price, and the reason is plain enough: it takes the full forty weight of the leaderboard column and also carries the tooling, context and Iran columns.

So the short answer is not one answer. If you work against the API and get a monthly bill, Opus 5 is half the price and second in this same table. If your work is the kind where getting it right once is cheaper than doing it twice, seventy-eight points is what you are paying for.

Where this table falls short

Our primary criterion is a human-preference leaderboard, not an automated test. WebDev Arena gives an absolute score, a date and a vote count, which is what publishing requires, but human preference also responds to readability and answer shape rather than only to whether the code is right. SWE-bench Verified would be a sharper measure, but its leaderboard renders in JavaScript and we have not read the figures directly. We do not print a number we have not read, even where someone else has quoted it.

One candidate on this page carries no rank: Gemini 3.1 Pro. It has a real score on the general text leaderboard, but we have not read its WebDev Arena row and Google publishes no context window for it at all, so more than half of the criteria weight is empty. Its name appears below the table; its position does not. GLM-5.2 Max and DeepSeek V4 Flash sat there too for a while, and now that their price and context window are verified from their own documentation, they carry a rank.

One thing about reading this leaderboard that came up this time and is not written down anywhere else: Qwen has two separate rows. There is qwen3.8-max at 1669 and qwen3.8-max-0902 at 1688, which is second on the whole board. We record the first, because that is the id in Alibaba own documentation and a dated row is a snapshot of one serving configuration. If you see Qwen listed second somewhere else, that is why, and not a mistake on our side.

One more thing, so the numbers do not confuse you: the scores are relative. Every column is normalised between its own minimum and maximum, so adding a new candidate moves everybody's score without any model having changed. Adding Grok 4.5 on 12 August 2026 did exactly that: the order held and the numbers rose. If today's figure does not match a screenshot from last week, that is why, and not because we touched anything.

Two more data changes moved the numbers on the same day. First, we read the text leaderboard in full this time and found that GLM-5.2 Max and DeepSeek V4 Flash had been on it all along; an earlier run had read only the top of the page. Both filled their general-text column. Second, GPT-5.6 Sol became an eighth candidate: openai.com still does not answer our server, but its developer documentation publishes the price and the context window, and its leaderboard row exists. One caveat about that row, since this is the coding page: it is named gpt-5.6-sol-xhigh (codex-harness), so the score belongs to the model inside OpenAI own coding agent rather than to the bare model. GPT landed between Kimi and Qwen, and no existing row changed place relative to the others.

28 August: DeepSeek got dearer and three rows moved

This is the only number in this reference that has ever gone up. On 12 August, DeepSeek V4 Flash cost $0.14 in and $0.28 out. Its own pricing page now carries two columns, peak and off peak, and even the cheaper column is 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 hours fall in the middle of the Iranian working day.

DeepSeek itself did not lose a place, because it is still the cheapest row in the table and still holds the top of the value column. But its lead narrowed, and since every column is normalised 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 those three models changed anything at all. That is what a relative ranking means, and it is why the arithmetic is printed on this page.

3 September: the gap at the top went from 22 points to 78

Until last week the three highest scores in this column sat inside 22 points, at 1692, 1674 and 1670. This page concluded then that the models differ by less than the numbers suggest, and that what genuinely changes a working day is tooling maturity: a command line tool, batch execution, presence on a cloud your company already has a contract with. For ranks two to eleven that still holds, and it is why the criterion carries fifteen points. For rank one it no longer does.

Three other rows also moved in the same update without changing anything about themselves. Kimi K3 Max went from third to fifth and Qwen3.8 Max from fifth to sixth, while Kimi leaderboard figure stayed at exactly 1674. GPT-5.6 Sol held fourth but its score fell from 51.6 to 48.8. All three are the arithmetic of new candidates entering and the columns normalising again.

Three new candidates, and one model deliberately not in the table

Muse Spark 1.3 from Meta and Gemini 3.8 Flash from Google joined the table today, both with a real row on the same leaderboard. Muse Spark carries a point other lists rarely make: Meta spent years being the maker that publishes its weights, and this one is not published. The model at the top of Meta line today is a closed one.

And the model that is not in the table: GPT-6 Astra, OpenAI new flagship at $10 in and $50 out. Neither leaderboard on this page has a row under that name, so seventy of its hundred weight would be empty and putting it in the table would produce one line reading "not ranked". It is in the model catalogue, and we wrote down here why it is not in the table so nobody assumes we missed it.