The frontier is shorter than your planning cycle

On July 31, DeepSeek published V4 Flash 0731.

When we retrieved Artificial Analysis's live results on August 3, its Intelligence Index v4.1 score was 50 and the estimated cost of the full benchmark suite was $72.02.

Evidence:DeepSeek release logV4 Flash resultsGPT-5.4 resultsGPT-5.4 release

GPT-5.4 xhigh, released on March 5, scored 51 at $2,185.46.

The qualifiers matter. The index combines nine text-oriented evaluations; equal totals can hide very different strengths.

Evidence:Index methodologyDeepSeek pricingV4 Flash resultsNIST evaluation

V4 Flash is text-only. Its public endpoint was still labeled beta. It produced far more output tokens during the suite than GPT-5.4.

Benchmark-run cost is not production cost. It says nothing about support, data handling, latency tails, or geopolitical exposure.

A separate NIST evaluation of V4 Pro found that its cost advantage changed by task and that its overall capability trailed the frontier. The result is a price-performance shock — not proof that every workload should switch.

Stanford previously measured a 280-fold fall in the cost of inference at roughly GPT-3.5 capability between November 2022 and October 2024.

Evidence:AI Index cost trend

DeepSeek's result is another version of the same economic force: yesterday's scarce capability becomes today's commodity faster than most annual roadmaps can react.

Our theory is that the model itself should be treated like a rapidly depreciating dependency.

The durable assets are the representative tasks, hidden constraints, evaluation cases, tool contracts, traces, and rollback paths around it.

Owning those assets lets a team capture a price-performance jump without rebuilding its operating model around a new vendor.

The strategic question is no longer which provider will win. It is how many days your team needs to decide whether a new provider wins on your work.

Release knowledge now has a short half-life.

A senior engineer who tries to stay ahead by memorizing model names and prompt tricks is volunteering for an endless news cycle.

The higher-leverage job is to make the system replaceable: define what good looks like, expose repository-specific constraints, capture representative failures, and insist on comparable evidence before changing defaults.

This also changes architecture. Provider-specific prompts and APIs should sit behind narrow adapters.

Task contracts, expected outcomes, and failure traces should remain portable. Portability is not abstraction for its own sake — it is an option on the next price collapse.

The team that can rerun its real workload this week can exploit a breakthrough. The team with no task bank can only debate public benchmarks.

Build a monthly swap test from 20 to 50 completed tasks, weighted by risk and drawn from your own repository.

Give the incumbent and two challengers the same tools, context, time limits, and hidden acceptance checks.

Record first-pass acceptance, severe failures, latency, tokens, human steering, review minutes, and rework.

Use one decision metric: cost per verified outcome equals model and infrastructure cost, plus human verification and expected rework.

The theory predicts that the best-value model will change more often than the harness — and that a portable team will reach a trustworthy decision in days rather than weeks.

It is falsified if public benchmark rank reliably predicts your winner, token price dominates total accepted-change cost, or model switching stays expensive after the adapters and task bank exist.

Those are useful failures. Each one identifies whether your real bottleneck is integration, verification, or a genuinely unique model capability.