Artificial Analysis refreshed its Intelligence Index on 6 August, and the new table says something uncomfortable for any team that defaults to the most expensive setting on offer. Claude Opus 5 tops the ranking on 63 points. It also tops the ranking on 63 points when you turn its reasoning effort down one notch, and that version costs 23% less to run.
The detail matters because most businesses pick a model and then leave the effort setting wherever the vendor put it. On the current numbers, that single choice moves the monthly bill further than switching provider does.
What changed in version 4.1.1
The update moved the index’s banking evaluation to version 1.0.1 and upgraded the grader used for three of its hardest tests to GPT-5.6 Luna running at medium effort. Grading is the part of a benchmark nobody looks at, and it quietly decides which answers count as correct.
There is a mild oddity worth naming. One laboratory’s model now marks the exam that every other laboratory is ranked on, in the same week OpenAI made Luna the default for free ChatGPT accounts. Artificial Analysis is transparent about the change, but it is the sort of methodological detail that belongs in the footnotes of any procurement decision built on these scores.
The same score for 23% less
Claude Opus 5 at max effort scores 63 and costs $2.34 per task. Claude Opus 5 at xhigh scores 63 and costs $1.80. Identical intelligence, 23% cheaper, and it returns its first chunk of output in 34.87 seconds rather than 56.78.
The gradient continues below that. Opus 5 at high scores 61 for $1.23, and at medium it scores 59 for $0.72. Read from top to bottom, that is 69% off the price for four points of measured intelligence, which for a great many production workloads is a trade most finance directors would take without a meeting.
Where the rivals actually land
Claude Fable 5 sits second on 62 points and is the most expensive entry on the board at $3.14 per task. It is quicker once it starts producing text, at 67 tokens per second, but it takes 97.20 seconds to begin, a pattern that also shows up when Opus 5 and Fable 5 are run side by side.
The comparison that deserves the most attention is Opus 5 at high against GPT-5.6 Sol at max. Both score 61. Both cost $1.23 per task. Opus 5 returns its first output in 20.64 seconds against 145.63 for Sol, which is roughly seven times faster to respond for the same money and the same score, a gap we break down further in our Opus 5 and GPT-5.6 comparison. Kimi K3 lands on 60 for $0.84 and starts answering in 3.09 seconds.
The value tier is closer than it looks
Two entries tie on 59 points from different laboratories. GPT-5.6 Sol at xhigh costs $0.81 per task and Claude Opus 5 at medium costs $0.72. Nine cents separate them, and both sit four points below the leader.
Kimi K3 at max is the one worth a second look. It scores 60 for $0.84 and returns its first token in 3.09 seconds, the quickest start of anything in the leading group. For interactive work that number often matters more than a point of index score.
It is worth separating two figures that get conflated. Output speed measures tokens per second once the model is talking, and Fable 5 leads the front runners there at 67. Latency measures how long you wait before it says anything at all, and on that measure Fable 5 is the slowest of the group at 97.20 seconds. A model can be fast and still feel slow.
Why Australian businesses should pay attention
Cost per task is the number that scales. A team running a few hundred document extractions or reconciliation checks a day will notice the difference between $0.72 and $2.34 within a single billing cycle, and none of it shows up in a demo.
Latency matters in a different place. Anything a customer waits on, a chat reply or a quote, lives or dies on the time to first response, which is where the gap between 20 seconds and 145 seconds stops being a statistic. Working out which processes need the top setting and which are perfectly served by medium is ordinary AI consulting work, and it is usually worth more than the model choice itself.
There is a procurement point in this too. Most vendor quotes and internal business cases are still written around a model name rather than a configuration, which makes them close to unfalsifiable once the invoices start arriving. Naming the effort setting in the proposal turns a vague estimate into something a finance team can actually check.

The setting is the decision now
The top of this leaderboard has compressed to the point where the four leading entries sit within two points of each other. When the models converge, the interesting variable stops being which one and becomes how hard you ask it to think.
That is a healthier place for buyers than the last two years of launch-day leapfrogging. It also means benchmark headlines are becoming less useful on their own, because a single number now hides a threefold spread in price depending on a dropdown.
For Australian teams budgeting AI spend for the next quarter, the practical move is to test the cheaper effort settings against real work before assuming the expensive one is required. On the current table the measured difference between the top setting and the middle one is four points, while the price difference is more than threefold. That is a gap worth an afternoon of testing.
