September 30, 2026, and a DataLearner mirror of the Artificial Analysis Intelligence Index v4.3 puts Claude Opus 5.5 in first place. Nine days later, an October 9 leaderboard relay via llm-stats puts GPT-5.6 Sol on top at 0.590. Two first places, nine days apart, and both have paperwork. As of the October 11, 2026 data cutoff for this piece, no dated source settles the argument between them, which is the best reason I know to write it down: most people read leaderboards in a way that guarantees confusion, and the fix is a procedure you can run in a minute. Consider this a method piece. The models in it are exhibits, each stamped with its board and its snapshot date, and the page is built to survive turnover — when every name here has been replaced by a newer one, the checks below should still work. Next scheduled review is January 15, 2027.
Two boards, two first places
The Intelligence Index is a curated battery. Artificial Analysis runs a fixed set of standardized tests on every model it lists, aggregates the results into one score, and versions the whole instrument; the September 30 snapshot that puts Opus 5.5 first is labeled v4.3, and that version label is part of the measurement. A fixed battery buys comparability inside a version — same prompts, same graders, same weights — and it also defines what you are comparing, which is performance on tests somebody else chose.
An arena is the other caliber. Models face blind, pairwise human votes on live prompts, and the standing rank is built from preference, with no syllabus and no grader. Arena is also a family of boards, because each modality gets its own. When Google's Nano Banana 2.1 shipped on October 6, its launch coverage and the follow-ups through October 10 quoted separate image-board results: fourth on text-to-image, up 80 points from its predecessor, fifth on multi-image editing, sixth on single-image editing. One model, three boards, three different ranks, inside a single fortnight.
Now put the two first places side by side. The October 9 relay that crowns GPT-5.6 Sol reports a normalized score on its own scale. The September 30 v4.3 result is an aggregate index position on a different scale. The two results differ in instrument and in snapshot date, and the quarter was crowded enough that the top moved between snapshots; Anthropic alone released three models between September 22 and October 7. One-line answer: the boards measure different things on different days, so the word "first" is doing a different job in each headline. The discipline that dissolves the confusion costs one sentence — every rank you quote carries its board, its version when one exists, and its snapshot date.
What a version bump breaks
When the index iterates, as it did in the latest bump to v4.3, the test set, the weights, and often the scale get rebuilt. Scores on either side of a version bump are different units, and comparisons that straddle one break before anyone starts interpreting them; quoting, say, a 61 from a superseded scale beside Sol's 0.590 is the spreadsheet equivalent of adding meters to feet.
I have the scar tissue. I once built a six-month trend line by pasting scores from two versions of the same index into a single column, and the cliff I found in the middle turned out to be my own paste job wearing the costume of progress. Here's the trap: version labels sit in report titles and URLs, where nobody looks. The working rules are short. Find the version label before comparing anything. Compare only scores that live inside one version. If you need a trend across a bump, wait for the source to re-run older models on the new scale, or compare positions within a single snapshot and leave the levels alone.
Price-performance you can defend
Price-performance starts with a constraint that most roundups skip: the score and the price have to come from the same snapshot on the same scale. The October 9 relay ships both on one page. GPT-5.6 Sol scored 0.590 at $5 per million input tokens and $30 per million output; Terra scored 0.550 at $2/$12; Luna scored 0.510 at $0.20/$1.20.
| Model | Score (llm-stats relay, 2026-10-09) | Price in/out, per M tokens | Blended at a 1:1 mix | Score per dollar | |---|---|---|---|---| | GPT-5.6 Sol | 0.590 | $5 / $30 | $17.50 | ~0.034 | | GPT-5.6 Terra | 0.550 | $2 / $12 | $7.00 | ~0.079 | | GPT-5.6 Luna | 0.510 | $0.20 / $1.20 | $0.70 | ~0.73 |
The blend in that last column is my arithmetic, stated plainly: assume half your tokens are input and half are output, so a model's blended price is the midpoint of its card. On that mix, Luna returns roughly twenty times Sol's score per dollar. Shift the mix toward output-heavy generation and every number moves — the method is what survives the movement, along with the caveat that a ratio only means something inside one snapshot on one scale. The table above never borrows a number from the v4.3 board, because mixing calibers to manufacture a ratio is the oldest trick in this genre.
The ratio's weakness comes first, as it should. A cheap point is still a point on somebody else's test, and nothing in it predicts your retry rate on your own tasks. Capability sets a floor; price-performance sorts the field above the floor. If your workload sits comfortably under a tier's ceiling, the ratio is the most defensible number in the room. If the task barely clears that floor, you are optimizing the wrong column, because you will pay for retries the leaderboard never sees.
Cost per task, the unit budgets run on
Per-token prices answer what a word costs. Cost per task answers what a finished job costs, and budgets are built out of finished jobs. The clearest public example in the current cycle is Xiaomi's MiMo-V2.6: a September 22 Economic Information Daily report placed it at the top of the Artificial Analysis open-source board with 46 points, first among domestic models, at a unit task cost of $0.13 — the first model reported inside the ten-cent class — ahead of Kimi K3 and Qwen3.8 Max. $0.13 per task is a number you can multiply by next quarter's roadmap. Notice, too, what the example is: 46 points sits on the open-source cut of one board on one date, which makes it a third scale in this article, and it never gets compared against 0.590 or against v4.3 positions in the same sentence.
Compute your own version of this number from your own bill: total spend on a model in a month, divided by jobs it finished to your standard. Requests sent is the wrong denominator, because a retry you paid for is still spend, and retry rates differ between models that sit near each other on a leaderboard. Vendors rarely volunteer this unit, which is why the $0.13 figure traveled. When a vendor quotes only a per-token card, the cost per finished job is left as an exercise for the buyer, and the exercise is where overruns live.
The same September 22 report priced the rest of the open board, which calibrates what the ten-cent class looks like against the field: GLM-5.3-Flash from Zhipu, 320B parameters, at $0.15/$0.50; Gemini 3.7 Flash at $0.75/$3.75; Grok 4.5 from xAI at $2/$6; MiniMax M2.7 with a 205K context at $0.30/$1.20; and Inkling-Small from Thinking Machines Lab, 276B parameters. One wrinkle from the same coverage: MiMo-V2.6-Pro reached third among open models on the Code Arena, placing alongside a high-spec Claude Fable5 edition — an arena rank, its own caliber again, quotable on its own board only.
Reading a vendor claim like an auditor
Vendor claims and leaderboard entries fail in different ways, so they get audited differently. The current Anthropic cycle shows the vendor pattern in its best form: Opus 5.5, announced September 22, with most work at Fable 5.1's level and running costs 40% below Opus 5; Sonnet 5.5 on September 28, 30% faster than Sonnet 5 and around 30% cheaper on most work; Haiku 5.5 on October 7 as the fastest and cheapest small model. Every one of those sentences carries a date. Every one measures against a baseline the vendor picked and a workload the vendor summarized as "most work." The percentages are real against their baselines; whether the baseline resembles your traffic is the part only you can supply.
The audit takes about a minute per claim:
- Baseline named. Cheaper than what, faster than what.
- Workload named. Which tasks, what mix.
- Price card published in dollars per million tokens, or a task cost with the task defined.
- Any rank carries its board, its version, and its snapshot date.
- A snapshot you can pull again yourself.
A claim that survives all five behaves like evidence, whatever its source. A claim that survives none is a press release in a lab coat. And anyone quoting a first place without the paperwork above is selling you something.
What the boards still miss
No board runs your prompts, so every score is an argument from analogy. Arenas drift with whoever votes that month. Index batteries take time to grow tests for new model classes, so a model can ship ahead of the instrument meant to grade it. Price cards move quarterly, which is one reason this page carries a review date. Task cost depends on retry behavior that only your own logs reveal. Treat every board as evidence with a shelf life, and stamp whatever you lift from it.
How x-rush reads these boards
The platform reads public leaderboards the way ADR-0008 describes: as admission evidence for routing by role, checked on real tasks before anything ships. Vendors are candidates for a job, and none of them is the product's identity; tasks route to a suitable model class, and the fixed credit price a user sees carries no vendor's name inside it.
Stamp every number you repeat with a board, a version, and a date.
Or keep handing your budget to whoever quoted a first place last.