AI Signals — Weekend Read: We turned off the thinking

Weekend Read #14 — September 24, 2026

AI Investor Barometer tracks how five LLMs form DCF assumptions for 24 listed companies — daily, independently, from identical inputs.

At 07:23 UTC on Monday, September 7, the daily run died. It had not crashed in any dramatic sense. One of the five models, DeepSeek, had simply been thinking too long. For each company it was producing tens of thousands of characters of private reasoning before writing a single number, exhausting a sixteen-thousand-token budget, retrying, and exhausting it again. Two and a half minutes per company, twice over, and the sixty-minute limit arrived with seven companies still unvalued. The day went into the database half-finished and was later excluded from every aggregate.

The thinking had not always been there. Until the last day of July, DeepSeek answered in about as long as it takes to read this paragraph. Then, overnight and without a change on our side, its provider began serving a reasoning variant under the same model name. Response times went from seconds to more than a minute. And something else moved at the same moment: DeepSeek's opinion.

For five months DeepSeek had been the panel's most reliable pessimist. In April we noted that it had never once crossed zero bias. In May we called its residual pessimism "the genuinely informative residual". In the two weeks before the switch, its estimates sat a median 7.4 percentage points below the other four models, company for company, day for day. After the switch, that gap was gone. DeepSeek landed almost exactly on the panel's median and stayed there.

So in September, with the run failing and an obvious suspect in hand, we ran the experiment that the situation offered. On September 14 we switched DeepSeek's reasoning off.

The machine got faster

The operational result was immediate. Average DeepSeek latency fell from 54 seconds per company to four. The daily run, which had drifted to between 19 and 54 minutes, now finishes in 17 to 21. Nine consecutive runs since have completed cleanly, all 24 companies, all five models. Whatever else the thinking was doing, it was expensive.

The question was what it had been doing to the numbers. The story wrote itself in advance: a reasoning mode arrives, a pessimist turns neutral; remove the reasoning mode, the pessimist returns. Reasoning, on this theory, had talked DeepSeek out of its caution — longer deliberation pulling it toward the consensus view that every model absorbs in training.

The first reading seemed to confirm it. In the two weeks before July 31, DeepSeek's median valuation gap against market price was −7.2 percent. Under reasoning it was −0.5. In the first nine trading days without reasoning: −6.4. Back almost to the old value. The bear, apparently, was home.

The trap in the absolute number

The absolute gap is the easiest number to compute and the easiest to misread, because it moves for two reasons at once: what the model thinks, and what the market does. If prices shift or the whole panel changes its mind together, every model's gap moves with it, and a model that has not changed at all appears to have changed.

September was exactly such a month. GPT's median gap went from −11.1 percent over the reasoning period to −14.4. Grok's went from −5.4 to −11.4. Claude held steady; Gemini edged up. The panel as a whole was leaning lower, and DeepSeek was leaning with it.

The honest measurement subtracts the herd. For each company on each day, take DeepSeek's gap and compare it with the median of the other four models on the same company, same day, same inputs. Market moves cancel. Panel-wide mood cancels. What remains is DeepSeek's own position within the group.

On that measure the story inverts. Before July 31, DeepSeek sat 7.4 points below its peers. With reasoning, 0.5 points below. Without reasoning, 1.5 points above. The bear did not come home. On the relative measure, turning the thinking off changed essentially nothing about where DeepSeek stands. It stayed where the July switch put it, in the middle of the flock.

Nine days is a short window, and 189 company-days is not a large sample. But the daily series leaves little room for doubt: on seven of the nine days since the switch, DeepSeek's median offset from its peers was exactly zero.

Exactly zero

That last detail matters. A median offset of exactly zero means that on the typical company, DeepSeek's model estimate is not merely close to the panel's but identical to it. The same assumptions produce the same discounted cash flow and the same number, to the cent.

This is the herd we described in August, seen from inside one model. Across the panel, somewhere between a fifth and a third of all model estimates coincide exactly with at least one other model's, because the models reach for the same growth rates, the same margins and the same discount rates on the same half-point grid. For DeepSeek the share has crept upward: 24 percent before July 31, 28 percent under reasoning, 32 percent since reasoning was switched off. Its underlying assumptions moved together and stayed moved: median discount rate down from 9.25 to 8.5 percent, target margin up from 20 to 21.75, five-year growth up from 4.0 to 4.5. Each shift points in the same direction, toward a higher valuation and toward the group.

If reasoning was not what changed DeepSeek's mind, what did? The honest answer is that we cannot see inside the provider. What we can see is that on July 31 two things changed at once under an unchanged name — the model began to reason, and its judgment moved — and that undoing the first did not undo the second. The most economical reading is that the reasoning mode was a symptom of a new underlying model rather than the cause of its new opinions. A second clue points the same way: on September 10, four days before we touched anything, the provider quietly began reporting a different model name in its responses, and DeepSeek's position did not budge. The model behind the label has been changing on the provider's schedule, not ours.

What the panel lost

For our purposes, the practical consequence is a change in the panel's shape. Since March, the five models have offered two recognisable bears: GPT and DeepSeek. Now the bearish end belongs to GPT at −14.4 percent and, increasingly, Grok at −11.4. Claude and Gemini sit near zero. And DeepSeek sits exactly between the two camps — the median of the panel, almost by definition. A reader comparing the dashboard's model tabs in October will see a different distribution from one in June, and very little of that difference is about the companies.

This is what running an observatory on rented models means. We can fix our prompt, our engine and our sampling settings, and we publish every change we make with its date. We cannot fix what a provider serves under a name. The lesson of July 31 was that a model can change its mind overnight. The lesson of September 14 is harder: when it does, you cannot necessarily change it back, and a switch that looks like it should restore the old model may restore nothing at all except the speed.

The charts now carry both dates. The bias chart marks July 31 as "ds" and September 14 as "ds2", and DeepSeek's history should be read as three regimes rather than one series. And when we report a regime shift from here on, we will measure a model against its peers, not only against market price — because the absolute number told us a comforting story that was not true.

Back to the run that died

The September 7 run failed because a model was thinking at length about questions it was about to answer the same way as everyone else. That is the image to keep. The thinking was real — tens of thousands of characters per company — and it was not what made DeepSeek's answers different from its neighbours'. Turning it off cut DeepSeek's response time from nearly a minute to four seconds, and the answers stayed where they were.

We will keep DeepSeek in the panel, with reasoning off and its history split into three honest segments. The old bear may have been the most informative voice we had. It is not coming back, and we had to take the thinking away to find that out.

Model estimates are research output from an automated system, not investment advice. Relative offsets exclude companies where every model is pinned to the same engine cap. The regime dates and series notes are documented on the changelog page.