AI Signals — Week 35, Aug 24–28, 2026

The Big Picture

Something unusual happened this week: all five models moved in the same direction. Claude, DeepSeek, Gemini, GPT, and Grok each registered a more negative bias shift relative to last week, with the aggregate consensus upside across the panel compressing noticeably. The temptation is to read this as a coordinated bearish signal — five independent analytical systems arriving at the same conclusion.

Resist that temptation. The bias_shift metric compares valuation gaps against each week's own market prices, which means rising equity prices mechanically shrink apparent upside even when the models haven't changed a single assumption. With markets broadly elevated heading into late August, a portion of this week's bearishness is the market doing the work, not the models. That said, the magnitude of the moves — particularly Gemini's -6.6pp and DeepSeek's -5.4pp — is large enough that pure price drift cannot explain everything. Something in the underlying estimates shifted too.

The cross-model spread tells its own story. At one end, Claude still sees +2.5% average upside. At the other, GPT sits at -4.2%. That 6.7pp gap between the most and least bullish models is the real signal: these systems are not converging on a shared view of fair value. They are running different mental models of the economy, and the divergence is structural, not noise.

Sector Signals

Every sector in the dataset lost favor this week — a clean sweep that reinforces the price-driven explanation above, but with meaningful variation in severity.

Industrials took the sharpest hit, falling 11.2pp to a thin +2.3% consensus upside. Last week's +13.5% reading had looked like genuine model conviction; this week's collapse suggests that conviction was fragile, or that industrial names rallied hard enough to erase the gap. With only three companies in the bucket, idiosyncratic moves dominate — KONE's target price revision of -12.9% this week is the likely culprit dragging the sector average down.

Materials deserves a flag: it is a single-company sector (UPM-Kymmene), so its -8.4pp shift is really a stock call, not a sector view. UPM sits at -20.7% implied downside, and the models appear to be getting more uncomfortable with the name. Treat this as a company signal wearing sector clothing.

Healthcare remains the most favored sector at +30.9% consensus upside, even after a -6.9pp retreat. The two-company bucket — Orion and Johnson & Johnson — is pulling in opposite directions: Orion at +83.5% implied upside versus J&J at -20.3%. The sector average flatters neither position.

Technology, the largest bucket at eight companies, slipped 5.3pp to +7.0%. The internal range here is extraordinary: from TIETO at +94.7% to TSLA at -56.2%. Averaging across that distribution produces a number that describes nothing in particular.

What the Models Reveal About Themselves

The most behaviorally interesting data point this week is Gemini's crossing of the zero line. Last week Gemini was mildly bullish at +5.4%; this week it sits at -1.2%, a 6.6pp swing that makes it the week's biggest mover. Gemini also carries the highest cap rate in the panel at 16.8% — tied with GPT — suggesting it is applying more aggressive terminal value constraints than Claude or DeepSeek. High cap rates combined with a newly negative bias is a model telling you it thinks current prices embed too much optimism about long-run growth.

DeepSeek warrants a methodological note. Following the provider-side switch to a reasoning-mode variant on July 31, DeepSeek's median gap moved from roughly -8% toward zero. This week it sits at +0.6% — still in positive territory, but down 5.4pp from last week's +6.0%. The post-July-31 DeepSeek is a structurally different series from its predecessor, and week-over-week comparisons should be read with that break in mind.

Claude remains the most expensive model by a significant margin at $42.46 per 1,000 valuations — nearly five times GPT's $20.26 and almost exactly five times Gemini's $11.36. Yet Claude's average upside (+2.5%) is the most bullish of the group, and its CAGR forecast (7.5%) is the highest. Whether that optimism is worth the premium is a question this data cannot answer, but the correlation between cost and bullishness is at least consistent: the model that thinks hardest also thinks most favorably.

Grok posts the lowest confidence score in the panel at 0.59, alongside the tightest terminal growth dispersion (stddev 0.10). That combination — uncertain about the answer, but anchored on a single growth assumption — is an unusual profile. It suggests Grok is varying its discount rates and near-term forecasts while keeping its long-run view on a very short leash.

Where the Framework Breaks

TIETO and TSLA both show dispersion of exactly 0.0 — perfect model agreement. But the similarity ends there. TIETO trades at €19.65 against a consensus target of €38.25, implying +94.7% upside. Tesla trades at $355.94 against a target of $156.04, implying -56.2% downside. In both cases, every model arrived at essentially the same number.

This is where the framework reveals its limits. Zero dispersion does not mean the models are right — it means they are reading the same inputs and applying similar logic. For TIETO, the models may be anchoring on analyst targets that have not caught up with a depressed share price. For Tesla, the models may be uniformly applying DCF discipline to a stock whose valuation has always been a referendum on optionality, not cash flows. Perfect agreement is not a signal of accuracy. It is a signal that the models share a blind spot.

The Model Scorecard