Artificial Analysis Intelligence Index (September 2026): Fable 5.1 and GPT-6 Astra tie at 53.
On 17 September Fable 5.1 scored 53.37 and GPT-6 Astra 52.81, at identical list prices. Running the index cost $7.63 against $3.26.
In September the top of the Artificial Analysis Intelligence Index was a tie. Claude Fable 5.1 scored 53.37 and GPT-6 Astra scored 52.81. Both round to 53, and the gap between the two best models anyone could buy was 0.56 points.
They also list at exactly the same price: $10 per million input tokens, $50 per million output. At which point the score has told you nothing useful and the interesting number is somewhere else entirely.
0.56
Points between first and second. Both round to 53.
2.3x
What Fable cost to run the index against Astra. $7.63 versus $3.26 per task.
45%
Of the v4.3 index weighted to evaluations with private questions or answers.
Same score, same list price, half the bill
List price per token is not what a model costs you. What it costs is tokens consumed times price, and a model that reasons longer to reach the same answer burns more of them.
On Artificial Analysis's own cost-per-task figures published alongside the v4.3 board, running the index cost $3.26 with GPT-6 Astra at max effort and $7.63 with Claude Fable 5.1 at max effort with its default fallback. Identical sticker price, 2.3 times the bill, and a difference in score too small to round apart.
What open weights cost at 45
The row worth staring at is GLM-5.3, sixth at 44.86. That is eight points behind the leader, which is a real gap and not one to wave away.
It lists at $1.40 per million input and $4.40 per million output. A seventh of the input price of the two models at the top, and a tenth of the output price. For the large share of production work that is not frontier reasoning, that ratio is the whole argument, and it is why the open-weights conversation stopped being ideological this year and became a budget line.
Two of the top eight models in September carried open weights. A year earlier that number was zero.
And what 42 costs
Keep going down the board and the ratio gets harder to ignore. GLM 5.3 Flash scored 41.91, about eleven points off the lead, at $0.15 per million input and $0.50 per million output. Running the whole index with it cost $0.25 against Fable's $7.63.
Thirty times cheaper, for a model that lands within about 80% of the top score on a composite of ten evaluations. Gemini 3.8 Flash sat beside it at 41.19, and GPT-5.6 Terra at 42.25 ran the index for $1.40.
None of these is the model you reach for on your hardest reasoning. All of them are plausible defaults for classification, extraction, summarisation and routing, which is where most production token volume actually goes. The practical question in September was never which model tops the board. It was how far down the board your workload can sit before anybody notices.
What version 4.3 changed
Version 4.3 raised the private share of the index from 40% to 45% and swapped two evaluations: τ³-Banking out, AutomationBench-AA in, the latter built with a private test set, and Terminal-Bench moved from 2.1 to 4.0.
Both changes point the same way. Public benchmarks get optimised against, so the index keeps moving weight toward questions nobody can train on. That is the right call and it has a cost, which is that scores stop being comparable across versions. Six weeks later that cost arrived in full, when Fable 5.1 fell from first to eighth without shipping a thing.
How to use a board like this
Read cost per task, not cost per token. It is the only figure that accounts for how many tokens a model burns getting to its answer, and it separated the September leaders by 2.3x when their list prices were identical.
Treat sub-point gaps as noise. Fable led Astra by 0.56 on a composite of ten evaluations. That is not a capability difference you can plan around, and it reversed by November anyway.
Price the gap before you dismiss the cheaper tier. Eight index points sounds decisive until you see it costs a seventh as much. Work out what those points are worth on your workload, which is a question only your own evals answer. The longer record of these scores is in the benchmark index.
Write the version on everything. This board is v4.3 as displayed on 17 September 2026. The published history shows v4.3.1 and v4.3.2 also landed that month, so even inside September the ruler was being adjusted.
Two models tied at the top, at the same list price, and one of them cost twice as much to run.
Price the task.


