Acquisition
Own the name the industry measures by.
$17,888
Gemini 4 Carbon leads Artificial Analysis’s rebuilt index at 60, one point ahead of Claude Opus 5.5. The score is the least interesting number on the board — the cost of a finished task runs from $1.27 to $10.04.
The machine-learning platform team re-ran its model selection in the first week of November 2026 and found its October comparison was already history. Artificial Analysis had rebuilt its Intelligence Index as version 5.0.1, combining nine evaluations under a new methodology — and scores from the old version do not compare with scores from the new one. The spreadsheet that had justified a vendor decision a month earlier was now a document about a benchmark that no longer exists.
The new standings, on the November 6–8 snapshot, are tight at the top. Google DeepMind’s Gemini 4 Carbon leads at 60. Anthropic’s Claude Opus 5.5 follows at 59.1. OpenAI’s GPT-6 Astra sits at 56.9, GPT-6.1 Sol at 54.6, and Claude Sonnet 5.5 at 55. The index is a composite of nine evaluations Artificial Analysis runs itself — agentic office work, terminal coding, hard science, knowledge, and instruction-following among them — and version 5.0.1 rebuilt the set, which is why the firm warns that v4 scores are not comparable with v5.0. The Artificial Analysis leaderboard is the primary source; ModelCap’s tracking of the same board confirms the November 6 snapshot and the 37-model field.
The score is the least interesting number on the board. The cost of a finished task is the one that decides budgets. On the same November snapshot, the cost per index task — a weighted average across the benchmark suite at each model’s top effort setting — was $7.76 for Gemini 4 Carbon, $10.04 for Claude Opus 5.5, $6.25 for GPT-6 Astra, and $1.27 for GPT-6.1 Sol. Sol trails Opus by four index points and finishes the same work for about one-eighth the money. Requesty’s November 4 ranking analysis puts the same figures side by side and draws the obvious inference: a leaderboard that ranks by intelligence alone is ranking by the wrong variable for anyone who pays per task.
The open-weights side of the board tells its own story. Moonshot AI’s Kimi K3.1 Pro released its weights on November 4 and went straight to first place among open models at 46, ahead of DeepSeek V4.1 Flash at 33 and Z AI’s GLM 5.3 at 32. The best open model now sits 14 points behind the overall leader — close enough that a team with the infrastructure to self-host has a real decision to make, and far enough that nobody should pretend the gap has closed. Axiomi’s open-weight leaderboard guide, checked November 8, records the same ordering and the same caveat: open weight is not open source, and the two boards measure different things.
And the other scoreboard disagrees. On LMArena’s human-preference vote, dated November 6, the highest open-licence models are GLM 5.3 at 1482 and MiMo V2.6 Pro at 1479 — a different winner from a different method. On a separate composite of eleven capability benchmarks, GPT-6 Astra ranks first, leading GPQA Diamond at 96.0% and Humanity’s Last Exam at 54.80%, while Gemini 4 Carbon ranks second. AI Productivity’s November benchmark review records all three orderings without reconciling them, because they cannot be reconciled. The leaderboard you pick is a decision, not a fact.
A benchmark is a ruler. In November, the ruler changed — and every comparison made with the old one became a historical document.
Index scores measure test performance, not production value, and the November board has three specific weaknesses worth naming. First, Carbon’s rollout is restricted: Google said it would go first to a subset of cyber defenders and trusted partners through its Fairwind Program, so the index leader is not broadly available to the buyers reading the ranking. Second, effort settings move scores by points — Carbon at high effort, Opus at max, Astra at max — and a buyer comparing a max-effort score with a medium-effort price is comparing two different products. Third, a one-point gap between 60 and 59 is noise on any real workload. The teams that switch models over a single index point are optimizing for the press release, not the invoice.
Price by finished task, at stated effort settings, with stated cache assumptions, and with a date on every number. Re-run the comparison monthly, because the board updates several times a week and models are superseded mid-quarter. Keep an open-weights option in the evaluation set — Kimi at 46 is now good enough to be the fallback that disciplines the frontier price. And treat any price chart without a date as a historical document, because in this market that is literally what it is: the same check that produced the $1.27-to-$10.02 spread also found the most-shared price charts out of date within days.
The market now generates enough spend to pay for independent measurement, and the missing institution in inference economics is a referee who sells no models and publishes task prices for stated workloads. The rankings will keep churning — new models, new index versions, new effort settings. The buyers who win will be the ones who stopped asking which model is smartest and started asking which finished task costs least, on a board they can trust, measured this week.
The category-defining name for that referee — inferencebenchmarks.com — is available for $17,888.