Best AI chatbot

Why Fable 5 leads, and why that is stranger than usual this time

One number and one contradiction. The number: in the 2 September update of the Arena text leaderboard Fable 5 sits first at 1507 and takes the whole forty-five-weight column. The contradiction: Anthropic moved that same model to legacy on 1 September. The top of this table today is a model its own maker no longer recommends.

Its successor is directly underneath it. Fable 5.1 is second at 1504, three points back, at the same price and the same context window. Three points on this board is not something anyone feels in conversation. So the ranking is telling the truth and the practical advice is a different thing, and we show both: under the Fable 5 name in this very table we print which model replaces it. Being superseded is a fact, not a score, and the rubric has no way to score it and should not have one.

Both of them lose the score-per-dollar column outright. Output on either costs $50 per million tokens, the dearest rows in the table, which puts their score-per-dollar at 30.1, the floor. Opus 5 is fourteen points lower at 1493 and half the price. So the real answer splits by question: if you are building chat on an API and paying a monthly bill, Opus 5; if you are buying an app subscription the price column does not apply to you and the top of the table is what you want, with the caveat that you should take 5.1 rather than 5.

One thing against our own leader, which no column here measures: Fable 5 is the only model Anthropic own table marks as slower, and its adaptive thinking is always on, so there is no lever to turn it down. In conversation, latency is felt. Our table does not see that and you will inside ten minutes of using it.

This table ranks models, not apps

That is a deliberate choice with a cost, so let us be direct about it. All five columns here are numbers published about a model: a leaderboard score, a price per million tokens, a context window. No vendor publishes any of the three for its chat app. Rank apps instead and three of the five columns go empty for most rows, which stops being a table.

The second reason matters more: one model runs under several apps, and an app swaps the model underneath it without telling you. Opus 5 sits under the Claude app, the API and Claude Code at the same time. Make "Claude" a row and we count the same number again in the coding table. It is the same reason the video category ranks versions rather than tools.

We made the opposite call for studying and it was right there, because Google publishes the notebook own limits as figures: fifty sources, five hundred thousand words. For a chat app no vendor publishes anything comparable. The cost is that Claude, Gemini and ChatGPT as apps have no row here, and their own pages are where those questions get answered.

Why ChatGPT has no row, and what we do have from OpenAI

The short answer is that ChatGPT is a product and this table ranks models, so it was never going to be a row. The longer answer counts against us and matters more: even if we wanted to, we could not describe it. Every path on openai.com returns 403 to this server, re-tested on 12 August 2026, while that domain own robots.txt does not forbid crawling.

One thing is open, though: developers.openai.com answers in full. GPT-5.6 Sol is listed there at $4 input and $20 output, a 1.05M token context window, 128K max output and a knowledge cutoff of 16 February 2026. That price was $5 and $30 as recently as 12 August, which makes it the only model in this table that has got cheaper.

And one thing we corrected in this run, at the expense of something we had written ourselves. We had said the OpenAI supported-countries list lives only on openai.com and that the domain will not answer us. Half of that was right: the domain still returns 403, but the same list also sits on developers.openai.com, and that one is open. It names about 195 countries and territories, Iran is not among them, and the page itself says access from outside the list can get an account blocked. So the Iran column on this row is filled as of today.

Those ten weights, plus the price drop, took GPT-5.6 Sol from 54.2 to 65.2. It still moved down a place, to fourth, because Fable 5.1 entered the table above it. That is exactly what a relative score means, and it is why the arithmetic for every row is printed on this page.

The summary is one sentence: the model can be ranked, the app cannot be described. What we still cannot tell you is which model runs under ChatGPT today and what its plans cost. Both live on the domain that will not answer. The OpenAI page writes that boundary out in full.

What reading the leaderboard properly taught us

Two rows we had recorded as unverified had been on the board the whole time. GLM-5.2 Max is 37th at 1472 and DeepSeek V4 Flash is 85th at 1438. An earlier run read the top of the page; this one read the JSON the page itself ships. So both candidates that could not be ranked in this category now can be, and the coding table moved as well. That gap was ours, not the vendors.

The second concerns what a row is named. This reference already argued that an Arena row names a serving configuration rather than a model. For OpenAI it goes one step further: the WebDev row reads, literally, gpt-5.6-sol-xhigh (codex-harness). That score belongs to the model inside OpenAI own coding agent, not to the bare model. Price and context attach to the base model; the score attaches to whatever was actually tested.

The third was a trap we did not fall into. The same board carries a row called deepseek-v4-pro at 1458, sitting above Flash. That is a different model in the same family, not a setting, so we did not borrow its number. Had we done so, DeepSeek would have climbed two places and nobody would have noticed.

Where our own method exaggerates

All twelve rows fit inside 69 Elo points, from 1507 down to 1438. Min-max normalisation stretches those 69 points into a spread from 78.6 to 32.2. So the score column makes the distance look larger than it is, and the last row is not weak: it is last in a table of high-end models.

Three new candidates joined the table today and not one of them had changed anything about itself. Fable 5.1 second, Meta Muse Spark 1.2 fifth and Google Gemini 3.8 Flash sixth. Their arrival pushed Qwen3.8 Max from fourth to eighth, while the only real thing that happened to Qwen is that its figure on that same leaderboard moved from 1490 to 1480. If you noted the order down somewhere yesterday, that is the explanation for the difference.

A sharper example, because it surprised us too. GPT-5.6 Sol wins the context column with 1,050,000 tokens and Kimi K3 Max holds 1,048,576. The gap is 1,424 tokens, about a tenth of a percent, and in practice it changes nothing. Both take essentially the full column, which is correct. But add a model with a two-million window tomorrow and both lose that column together, without anything about either having changed.