1. The Top of the Leaderboard Is Now a Cluster, Not a Single Model
The gap between the best and the tenth-best model on the Intelligence Index is small enough that token price, latency and
task-specific behaviour decide more procurement conversations than raw benchmark rank. Claude Opus 5.5 leads at 58,
but several models within a few points cost an order of magnitude less per task. That is the defining property of the October 2026 market:
capability has clustered, and cost has not.
Reasoning effort is a price dial
Artificial Analysis measures the same model at multiple reasoning efforts. Claude Opus 5.5 moves from 58
at max effort to lower scores at medium and low effort, and its cost per task falls from $5.98 toward
$5.98 as thinking tokens shrink. Picking an effort level is now as important as picking a model.
Latency is the hidden tax
Time to first answer token ranges from sub-second for small open models to several minutes for frontier reasoning models
on hard tasks. A model that is three points smarter but ten times slower changes how a team works, not just what it pays.
Open weights closed most of the gap
MiMo V2.6 Pro scores 46 on the same index while publishing weights, and several
open models are within single digits of the frontier at a fraction of the per-token price. Self-hosting is no longer a purely
ideological choice.
2. Cost per Task Is the Metric That Matters
Price per million input tokens is easy to quote and easy to mislead with. The Intelligence Index cost per task accounts for
input, cached input, cache writes, reasoning tokens and answer tokens, weighted by task type. The cheapest models on that metric
are not always the models with the lowest sticker price.
Cost per task is measured by Artificial Analysis on a standardised suite; it is not a prediction of your repository's bill.
Use it to rank models, then measure your own workload.
3. What Changed in 2026, in Practical Terms
- Context windows normalised at 1M tokens. Most current frontier and mid-tier models now advertise roughly one million tokens, and long-context pricing is flat for several of them. The differentiator is retrieval fidelity, not the headline number.
- Reasoning became a configurable SKU. The same model name can be served at low, medium, high and max effort, with different latency, token use and cost. Agent harnesses now expose those settings directly.
- Cache pricing became a first-class cost lever. Anthropic reads cached tokens at 5% of base input on Opus 5.5 and Sonnet 5.5, and 2.5% on Fable 5.1. DeepSeek off-peak cache hits on V4.1 Flash cost $0.003 per million tokens. Long-lived repository contexts are dramatically cheaper when cached.
- Peak pricing returned. DeepSeek runs two weekday peak windows (01:00–04:00 and 06:00–10:00 UTC) and charges double during them; everything else, including weekends, is off-peak. Scheduling batch agent runs is now a real cost decision.
- Subscription caps became dollar-denominated. GitHub Copilot, Cursor, Command Code and others now meter in credits or usage pools rather than request counts, which makes effective cost per task more transparent — and more sensitive to model choice.
Key takeaway: the useful unit of comparison is no longer "best model" or "price per million tokens". It is
cost per completed task at the quality bar your team actually needs, with latency and context fidelity as constraints.
The tables above give you the measured inputs; your own task mix supplies the weights.