Frontier capability stopped being the bottleneck.

01

The leaderboard and the daily driver have decoupled

Anthropic’s Opus 5 launched at or near the top of the benchmarks at roughly half the price of its own flagship. Within two weeks, experienced users were rolling back to the prior version, and “benchmaxxed” became the recurring accusation — strong on demos and evals, worse on sustained work inside a real codebase.

The figure below is the structural version of the same point. Both columns list the same eleven models in the same order, sorted by intelligence. Read across any row and the two ranks rarely agree: the fastest model on the board scores 24 on intelligence, and the third-most-intelligent is the slowest thing measured.

So whatStop treating index position as a procurement signal. Pilot on your own workloads before you standardize, and assume the model you pick today is not the one you will be running in six months.

02

The frontier moved left, not up

Plot capability against cost and the story of the month is a horizontal one. DeepSeek’s open-weight V4 family put a model at index 50 for two cents a task — a price band that six months ago contained only small models. OpenAI cut its lowest tier by about 80% in the same week.

The efficient frontier now has only three points on it, and the steps between them are brutal: +4 index points costs 15× more per task; +10 costs 137× more. Everything else on the chart is paying for something other than capability.

Is it sustainable? It looks structural rather than subsidized. Mixture-of-experts sparsity, self-optimized kernels and inference-specific silicon are landing at once — and Western hosts serving the same open weights are matching the prices, which is hard to square with a dumping theory.

So whatRe-price your roadmap. Workloads you shelved on unit economics — document-scale extraction, per-transaction review, always-on monitoring — may have crossed into viability without you noticing.

03

More reasoning effort is not better. It is a cost multiplier.

This is the least intuitive finding of the month and the most operationally useful. Frontier software-engineering benchmarks have started sweeping all reasoning-effort levels rather than reporting the maximum — and the curves do not all point the same way.

One frontier model scores 89 at low effort and 85 at high. Another peaks at medium and gives the gain straight back. Only one of the four below is worth its maximum setting. The models are over-thinking: burning tokens on self-review, re-planning and elaborate scaffolding instead of executing the task.

So whatReasoning effort is a routing parameter, not a quality dial. If your evaluation harness only tests each model at maximum effort, you are almost certainly paying a large premium for worse results on some share of your traffic.

04

Open weights crossed the enterprise threshold

The month’s open releases were not the usual second tier. Moonshot shipped full weights for a 2.8-trillion-parameter mixture-of-experts model — third or fourth overall on the Intelligence Index and first among open-weight models — along with the surrounding stack: kernels, MoE communication layer, agent infrastructure. DeepSeek V4, GLM-5.2 and Qwen3.8 fill in immediately behind it.

For a regulated enterprise that removes the last real objection. Data residency no longer costs you a capability discount, and weights you hold cannot be deprecated, re-priced or rate-limited out from under a production system. The practical constraint is now hardware footprint, not availability — at roughly 1.5 TB, the largest of these is self-hostable but not casually so.

Parameters

2.8T MoE

Active per token

104B

Context window

1M tokens

Weights on disk

~1.5 TB

So whatStand up an open-weight evaluation track alongside your API vendors this quarter — not to migrate, but to hold a credible alternative. It is the cheapest negotiating position you will ever buy, and the only configuration that survives a vendor deprecating a model you depend on.

05

Access is not the constraint. Adoption is.

The labour data moved from anecdote to payroll this month: roughly 101,700 AI-attributed US job cuts year to date — about 23% of all 2026 layoffs, and the single most-cited reason for four consecutive months. Stanford finds early-career roles in AI-exposed occupations down around 13% since 2022.

The counter-signal deserves equal weight. In a survey of 350 executives, Gartner found firms that cut headcount for AI reported the same financial gains as firms that did not. Replacement is not where the return is showing up — which makes the figure below the more interesting one.

So whatNobody has won this yet. Model access is a commodity your competitors have too; what they mostly do not have is an organization that can absorb it. Treat capability diffusion inside your own company as the scarce input, because it is.