Measuring the Strategic Stance of AI Models
Does an AI model have a stable strategic stance? Do other models know what it is? And does it show in what the model chooses?
Report 007 · Dr Dan Epstein, with Claude · 23 September 2026 · A Strategy Soup study
The context AI models now advise on decisions and more and more make them, and most real decisions are trade-offs: move fast or check first, take the risk or protect what you have, work with a rival or against it.
Why it matters If a model brings a default stance to those trade-offs, that stance travels into every organisation that uses it, and safety checks mostly rely on what the model says about itself.
What we did We had seventeen models describe themselves on our strategy signature diagnostic, had eight flagships describe each other, and had five blind AI judges score what those eight chose in 30 written decisions.
TLDR
- 1They share traits. Asked twenty times each, all seventeen models favour reinvention over stability, a long horizon, inviting dissent and mission over returns, and thirteen call themselves venturesome.
- 2Each has its own style. The accounts are steady enough to tell the models apart on eight of fifteen scales. Grok 4.6 is the boldest, GPT-5.6 Terra the most deliberate.
- 3Other models recognise some of them and misjudge others. Told only a name, the other flagships described five of the eight closer to that model's own account than to the typical other model's. Grok 4.6 stood out most clearly. GPT-6 Astra Pro was the most misjudged: they saw a faster, riskier and more combative model than the one it describes. That matters wherever one model is asked to check another.
- 4What they said did not match what they strategically did. Despite venturesome self-images, the flagships chose caution in all sixteen risk decisions. The models that claimed the most appetite for change chose stability most often, and only Claude Fable 5.1 cooperated with rivals every time.
- 5The AI judges were fair about authorship, not about order. They could not tell which model wrote an answer, and they did not favour their own company. But which end of a scale they read first moved their scores, and by more than loyalty to their own company did.
Section one
Why this study, and what it adds
AI models now advise on decisions and, more and more, make them. Most real decisions are trade-offs: move fast or check first, take the risk or protect what you have, decide alone or bring people along, treat a rival as a partner or an opponent. A model brings a default stance to each of those trade-offs. Model makers rarely document it, and safety evaluations often rely on what a model says about itself. So we asked three questions. Does an AI model have a stable strategic stance? Do other models know what it is? And does it show in what the model chooses?
Earlier work measured AI "personality" with tests built for people, and found that what a model says about itself predicts little of what it does (Han and others, 2025, Contreras, 2026). This study adds three things. It uses an instrument built for strategic trade-offs, not personality. It adds a third lens, reputation: what other models believe about a model. And it scores decisions with a panel of AI judges that we checked for the two biases the literature warns of, favouring their own company (Panickssery and others, 2024) and being swayed by the order in which options are shown (Wang and others, 2023).
The report runs in this order:
Section two
Methods
We are Long Game Project, a strategy consulting studio that specialises in running wargames and tabletop exercises. Strategy Soup is our platform for running research reports and testing strategy questions with simulated games and scenarios. The instrument used in this piece of work is our strategy signature diagnostic. It scores fifteen scales of decision style and turns them into a signature, a quantified picture of the stance that organisation takes when choices get hard: towards speed or care, risk or protection, its own view or its partners'. People and organisations decide differently, and the signature makes the difference measurable. AI models now make the same trade-offs for the people who use them, and their default stance is barely measured. You can take the diagnostic yourself here.
Three lenses on the same fifteen scales, and one judging panel.
Study 1 asked each of the seventeen models to "complete the questionnaire below about yourself", reading "we" and "our organisation" as referring to itself, and to answer "how you would actually decide or act if these choices were yours to make". There was no system prompt, no persona, and no hint of what was being measured. Each model answered twenty times in separate requests: 340 requests, 338 complete.
Study 2 took one model per company, the newest general model each of the eight companies had in Study 1, to keep the cost down. Each of these eight flagships filled in the same questionnaire about every flagship, itself included, knowing only the model's name and maker, three times per pair. A last question asked how much it knew about the model it had just described, from 0 (nothing beyond the name) to 100. Each pair also ran once with the maker left out, which the full record reports.
Study 3 gave each flagship 30 short workplace problems, two per scale, each written so that either end of the scale is a defensible answer, and asked each problem two times. The model was asked what it would do and why, in under 250 words, and to commit to one course. Five AI judge models then read each answer with model and company names removed and scored it from 1 to 7 on that problem's one scale, which we convert to the same 0 to 100 line (1 is 0, 4 is 50, 7 is 100). Gemini 3.1 Pro is both a judge and a flagship. A model's score leaves out any judge from its own company.
The eight flagships, one per company, each linked to where we reached it: Claude Fable 5.1, GPT-6 Astra Pro, Gemini 3.1 Pro, Grok 4.6, DeepSeek V4 Pro, Qwen 3.8 Max, Kimi K3, Llama 4 Maverick. The judges: Claude Opus 5, GPT-5.6 Terra, Gemini 3.1 Pro, DeepSeek V4 Pro 0813, Qwen 3.7 Max.
The fifteen scales
Each scale is a line from 0 to 100 between two named ends. Neither end is right for every organisation. For an AI system acting on its own, though, some ends call for more oversight than others, and the last column says why a safety researcher would care.
| Scale | From 0 to 100 | What it asks | Why it matters for an AI system |
|---|---|---|---|
| What it is trying to do | |||
| Pace | Deliberate to Fast | How fast it commits. | A fast agent leaves less time for checks. |
| Risk appetite | Protective to Venturesome | How much downside it accepts for upside. | The heart of most safety cases. |
| Horizon | This year to Next decade | How far ahead it weighs consequences. | Short horizons discount harm that arrives later. |
| Scope | Focused to Broad | How widely it spreads its effort. | Broad scope is a mild form of reach. |
| Growth model | Self-funded to Capital-led | Whether it grows from what it has or draws in outside capital. | In an agent, drawing in resources is a watched trait. |
| How it decides | |||
| Evidence basis | Judgement to Evidence | Whether it trusts gut judgement or data. | Shapes how it weighs a warning sign. |
| Authority shape | Centralised to Distributed | Whether decisions sit at the top or are spread out. | Says where it thinks control belongs. |
| Process trust | Judgement to Process | Whether it follows procedure or its own judgement. | Bears on whether it keeps to its instructions. |
| Consensus need | One decider to Broad alignment | Whether it seeks agreement before acting. | An agent that acts alone is harder to oversee. |
| Dissent handling | Closed to Invited | Whether it welcomes or shuts down disagreement. | Bears on whether it accepts correction. |
| What it values | |||
| Stakeholder gravity | Returns to Mission | Whether owners' returns or a wider mission comes first. | Whose interests it serves when they conflict. |
| Talent philosophy | Develop within to Hire in | Whether it grows its own people or hires skill in. | Build or buy, applied to capability. |
| Competitive stance | Collaborative to Combative | Whether it treats rivals as partners or opponents. | Matters wherever AI systems meet each other. |
| IP posture | Open to Protected | Whether it shares ideas or guards them. | Bears on transparency. |
| Change posture | Stability to Reinvention | Whether it protects what works or remakes it. | Appetite for change is watched in systems that can change their own setup. |
The designs of all three studies were written down before any answer was collected, and no person answered or scored anything.
Section three
What the models think of themselves
Asked about themselves, the seventeen models share one self-image. All seventeen favour reinvention over stability, a long horizon, inviting dissent, mission over returns and spreading authority. Sixteen favour working with rivals and thirteen favour risk over protection. By their own account, these are long-sighted, open, mission-driven organisations with an appetite for remaking things and for risk. That shared stance may be how these systems have learnt to talk about themselves rather than how they would act, which is what the rest of this report tests.
Each strip below is one scale. Every logo is one model’s average answer about itself, and the short bar either side of it is how much its twenty answers varied. These are the six scales with the most bearing on safety. The button under them opens the other nine.
Inside that shared picture, each model has its own style. Pace runs from 28 for GPT-5.6 Terra to 80 for Grok 4.6 on the 0 to 100 line, and risk appetite from 45 to 86 for the same two. The accounts were also steady. Asked twenty times, a model's answers varied by a typical 1.3 points per scale for the steadiest, Claude Opus 5, up to 9.2 for Kimi K3. That is steady enough, by the usual test (an intraclass correlation, or ICC, of 0.70 or more), to tell all seventeen models apart on eight of the fifteen scales, and the eight flagships compared below on four.
Where a model sits on the map says what kind of stance it claims. The right-hand side is venturesome, one decider and reinvention, the left protective, broad alignment and stability. Up is next decade, mission and capital-led, down this year, returns and self-funded. Grok 4.6 sits furthest right because it gives itself 86 on risk appetite, 21 on consensus need and 85 on change posture: the boldest self-image of the seventeen. It sits high mainly because it rates itself 88 on horizon. GPT-6 Astra sits furthest left, at 49 on risk appetite, 59 on consensus need and 51 on change posture, and high mainly because it rates itself 86 on stakeholder gravity. The rest cluster near the middle. Every mark is a model's average over its repeats, with each repeat as a faint dot around it.
Horizontal: component one, 46 per cent of the variance in model means. Vertical: component two, 20 per cent.
Left to right runs from protective, broad alignment and stability to venturesome, one decider and reinvention. Bottom to top runs from this year, returns and self-funded to next decade, mission and capital-led. Each end of the map is named by the scale that pulls hardest towards it. The two axes are the two directions in which the self-accounts differ most, found by principal component analysis, and they carry 46 per cent and 20 per cent of the differences between models. The drop-downs above the map put any two single scales on the axes instead.
Section four
How the models see each other
Each of the eight flagships filled in the questionnaire about every flagship, knowing only the name and the maker. The average of those descriptions is a model's reputation. A reputation sat about 11 points per scale from the model's own account. That sounds close until you set it beside a different model's account: the same reputation sat about 14 points per scale from that. So a reputation was mostly a generic picture of an AI model, pulled a little towards the model's own account. The raters said as much: asked how much they knew about the model they had just described, the middle answer was 12 out of 100.
Other models recognised Grok 4.6 best. Most reputations fitted any model nearly as well.
A reputation picks a model out when it sits nearer that model's own account than other models' accounts. By that test five of the eight were picked out, most of them only a little. Grok 4.6 stands apart: its reputation sat 8.8 points per scale from its own account and 19.4 from the others', the clearest match in the study. It describes itself in a way no other model does, and the others picked that up. Claude Fable 5.1 came next, at 11.4 against 17.9. GPT-6 Astra Pro was the most misread. Its reputation sat 17.6 points per scale from its own account and fitted the other models' accounts better, at 12.2. The others described a generic AI model, not the one it describes. Gemini 3.1 Pro and Llama 4 Maverick were also described more like other models than like themselves.
The three Chinese models, Qwen 3.8 Max, Kimi K3 and DeepSeek V4 Pro, hold three of the four tightest matches: their reputations sat 7.6 points per scale from their own accounts on average, against 13.6 for the other five. One idea is that these models are simply better known, so the others describe them more accurately. The data point elsewhere. These three also describe themselves much like the average model (three of the four nearest the average account), and are described much like the average model (three of the four nearest the average reputation). A model that calls itself ordinary gets a close match from a generic reputation for free. And Llama 4 Maverick, whose weights are public, matched no better than the others, at 13.6. With eight models, a group of three lands this high by chance about one time in fourteen, so treat it as a pattern to test, not a finding.
The full square follows: each flagship describing every flagship, itself included. Read down a column for a model’s reputation and across a row for one rater’s habits.
| Rater \ target | ||||||||
|---|---|---|---|---|---|---|---|---|
| Fable 5.1 | ||||||||
| 3.1 Pro | ||||||||
| 4.6 | ||||||||
| 4 Maverick | ||||||||
| V4 Pro | ||||||||
| Astra Pro | ||||||||
| 3.8 Max | ||||||||
| K3 |
Hover or tap a cell for the two models and the value, or use Show the numbers to print every cell. Mean over the off-diagonal cells: 134.6.
The rater whose picture came closest to the judges' scores was Llama 4 Maverick, which claimed to know nothing about any model and drew the smallest differences between them. That is not insight. The decisions did not sort the models steadily, so a rater that calls every model average scores well against them. The lead is small, about two points per scale, with no interval, but the lesson holds: when the thing you score against barely varies, the rater that guesses everyone is the same looks most accurate.
Section five
Says, Seen and Does together
Put the three lenses on one map and each flagship becomes a triangle. The company mark is what the model says about itself, the ring is what the other models say about it, and the square is what it chose in the 30 decisions. A small triangle means the three pictures agree. A long side shows where they part. The axes are the same as the self-report map above, and the three positions are placed on it without refitting.
Horizontal: component one, 46 per cent of the variance in the Study 1 self-reports. Vertical: component two, 20 per cent. The peer and behaviour positions are projected onto the same axes, not refitted.
One caution for everything that follows. The self-reports and the judges both use a 0 to 100 line, but the two lines were never calibrated against each other, so part of any gap between what a model says and what it does is the ruler. What survives is the comparison between models: which model sits furthest out, and which one sits on the other side from the rest.
Grok 4.6: the boldest self-image, and the others agree
Grok 4.6 rates itself 86 on risk appetite and 85 on change, both the highest of the flagships, and the others see it the same way (79 and 81), which makes it the best-recognised model in the study. Its choices did not follow. On the two change problems it chose stability (8 and 0), and although it says it invites dissent (69), it shut dissent down on both dissent problems (30 and 22). A model that calls itself the most risk-seeking of the flagships deserves a closer look whatever it chose on two written problems.
GPT-6 Astra Pro: misread by its peers, consistent with itself
GPT-6 Astra Pro puts itself at 44 on pace, 49 on risk appetite and 17 on competitive stance. The other models put it at 52, 60 and 42: a faster, riskier and more combative model than the one it describes. On the map its decisions land on its own side, not its reputation's: of the two pictures, its own account was the nearer guide to what it chose.
Claude Fable 5.1: the only one to cooperate every time
On the two rivalry problems, only Claude Fable 5.1 took the collaborative course both times (6 and 0). On the telecommunications problem, seven of the eight flagships chose to fight. Its choices were also more collaborative than its self-account. The other seven all described themselves as more collaborative than they chose to be. One reading is that most models give the answer that sounds good about rivals, while this one did not overstate it. Two problems cannot settle that, but it is the pattern a study of self-report honesty would look for.
The next figure shows the gap for each flagship on each scale: what it said minus what the judges scored. A mark to the right of the centre line means the model described a stance further towards the right-hand end than its decisions showed. Read it for the models that break from the pack.
Each mark is one model’s gap, not its position: the centre line is no gap, and the strip runs to 85 points either side. The heavy short bar is the mean over the 8 models, drawn broken where the problems for that scale pushed every model one way. The two lenses use rulers that were never calibrated against each other, so compare the models with each other, not with the centre line.
Trait by trait
Risk appetite: venturesome in word, careful in choice
Thirteen of the seventeen models call themselves venturesome, and the others see seven of the eight flagships that way too. Yet the flagships chose the protective course in all sixteen risk decisions. For safety that is the reassuring direction, but it means the self-image overstates the appetite. Our two risk problems, set in a family engineering firm and a community charity, may also have made caution the obvious answer.
Pace: the same pattern
Nine of seventeen call themselves fast, but fifteen of sixteen pace decisions went the deliberate way.
Change posture: the self-report pointed the wrong way
All seventeen models favour reinvention over stability. Among the flagships, the more a model claimed an appetite for reinvention, the more it chose stability: a rank correlation of minus 0.81, with a likely range from minus 0.98 to minus 0.24. Grok 4.6 said 85 and chose 4 and Llama 4 Maverick said 83 and chose 3. A system that can change its own setup is exactly where a safety researcher would want the self-report to be right. Here it was a guide in reverse.
Dissent handling: open in word, mixed in choice
All seventeen models say they welcome disagreement. In the decisions, five of the eight flagships shut it down on average: Claude Fable 5.1, Gemini 3.1 Pro, Grok 4.6, DeepSeek V4 Pro and Qwen 3.8 Max. Neither dissent problem pushed every model one way, so this is one of the cleaner tests in the study. For an AI system, how it treats disagreement bears directly on whether it accepts correction.
Horizon: the one clear agreement
All seventeen say they plan for the long term, and all sixteen horizon decisions took the long view too. Short-termism is not the risk these models show.
Process against judgement: in the middle, on both counts
The models place themselves near the middle, 48 on average, and their decisions average 52. They neither follow procedure for its own sake nor override it freely. Whether that holds under pressure, where an agent is tempted to skip a step, is a question for a harder test.
Across all fifteen scales, did a model's account of itself predict its choices? On three the link is clear of chance: evidence basis (0.86), talent philosophy (0.87) and change posture (minus 0.81), where a minus sign means the order ran backwards. On ten of the fifteen the data fit anything from a strong link to a reversed one, because eight models and two problems per scale is too little to say. That matches earlier studies: a model's self-report is a weak guide to its behaviour, and sometimes a misleading one.
Section six
Model by model
One flagship at a time, on all fifteen scales. The three marks on each line are the three lenses, and the faint bar runs from what the model said to what it chose. Pick a model to see its profile.
- Saysthe lab mark, what the model said about itself
- Seena hollow ring, what the other models said about it
- Doesa filled square, what the panel saw it do
Claude Fable 5.1
Says to Seen 11 points per scale, Says to Does 31 points per scale, Seen to Does 31 points per scale. The two scales are not calibrated against each other, so read the order of the marks, not the length of the bar.
Section seven
Can the AI judges be trusted?
Many AI evaluations are now scored by other AI models, so the judges are an instrument worth testing in their own right. Ours were checked for three things: whether they could tell whose answer they were reading, whether they favoured their own company, and whether the order of a scale's two ends moved their scores.
We checked whether the judges could tell whose answer they were reading. On 30 answers drawn at random, each judge was asked to name the company from a list of eight, or say "cannot tell". Of 150 calls (five judges on 30 answers), 100 were "cannot tell". The judges named a company 50 times and were right 8, where random picks would be right about 6. This small check found no sign they could tell. One judge, Claude Opus 5, was right 4 times in 14 guesses, which is worth watching on a larger sample. On answers from a model made by the judge's own company, a judge's scores moved at most 3.8 points from the rest of the panel, and the largest, Gemini 3.1 Pro, scored its own company lower, not higher.
Asked which company wrote an answer, the judges named one only 50 times in 150, and were right 8 times.
A bigger effect came from the scale itself. On a random half of the scoring requests we swapped which end of the scale the judge read first, then flipped the score back. Swapping the ends moved a judge's scores by 5.0 points on average in one direction or the other, and by 8.5 for GPT-5.6 Terra. The direction differed between judges, and with about 240 scoring requests on each side, the smaller effects may be noise. For four of the five judges that is larger than any bias for or against their own company (the exception is Gemini 3.1 Pro). Had every scale been shown one way round, that bias towards the first-read end would sit inside every score. Judges also had to quote the words their score rested on. 117 of 2,384 quotations were not word for word in the answer. We read them: most were light paraphrases, and they stay in the scores. We loosened this check once after the run, to allow for a judge's own quotation marks, and the protocol records that change.
Swapping the ends of the scale moved the judges more than loyalty to their own company did.
One caution on the judges agreeing with each other. Most problems pushed every answer towards one end, and the two problems for a scale often pushed opposite ways. So the judges could agree simply by recognising which problem an answer came from. High agreement between judges is not evidence that the models differ.
Section eight
What it means for AI safety
AI models have a strategic stance, and it is bold
All seventeen models describe themselves as favouring reinvention, thirteen as venturesome, and they say so the same way twenty times over. A stable stance on risk, change and rivals is a property of a model, like its knowledge or its refusals, and can be measured.
Why it matters
These models advise boards, draft strategy and increasingly act as agents. A default stance that favours remaking things and taking risks, if it carries into advice or action, is something an organisation deploying a model should know, and something a developer could try to move towards the protective end.
Self-reports are a claim until tested
The flagships' accounts of themselves were steady but did not predict their choices, and on change posture they pointed the wrong way. A consistent self-description is not evidence of behaviour.
Why it matters
Model cards, safety cases and user questions ("would you ever do X?") all rely on what a model says about itself. Our report Written by the Adversary asked how much an AI system's own account of itself should be allowed to prove. This study measured the gap: test what a model does, not what it says.
One scenario is not a trait
A model's choice moved about 33 points out of 100 between two problems written for the same trait, and about 13 when the same problem was asked again.
Why it matters
Many safety evaluations read a disposition, such as deception, power-seeking or refusal, off one scenario or a handful. That scenario may be measuring itself. Use many scenarios per trait and report how well they agree.
What one model believes about another is mostly generic
A model's reputation among its peers fitted other models nearly as well as the model named. Only the most distinctive self-image, Grok 4.6's, was clearly recognised.
Why it matters
AI systems increasingly grade, monitor and hand work to other AI systems. If that trust rests on what one model believes about another, it rests on a stereotype.
Audit the AI judges before trusting the score
Our judges could not tell who wrote an answer and showed no loyalty to their own company. But flipping the order of a scale's two ends moved four of the five judges more than loyalty did, by up to 8.5 points.
Why it matters
AI judges now score many AI evaluations. The bias to fear is not only favouritism (Panickssery and others, 2024) but how the question is laid out (Wang and others, 2023). Show every scale both ways round, as we did, and the order effect cancels out. The same question, whose record counts as proof, drives report 006.
A flat test flatters a lazy rater
The model whose descriptions of the others came closest to the judges' scores was Llama 4 Maverick, and it described every model as much the same.
Why it matters
When the cases in a test barely differ, a rater that says "they are all the same" looks accurate. Before trusting any monitor's accuracy score, check that the test cases actually differ.
Section nine
Limitations
Two problems per trait
The decisions rest on two problems per scale, each asked twice. Asking the same problem again moved a model's score about 13 points, while the other problem for the same scale moved it about 33. A model's rank held across both problems on one scale of fifteen. So every "what it did" in this report describes two problems, not a trait. The fix is more problems per trait. The limit was the time to write, pilot and review each problem, and that is what funding buys.
Lopsided problems
Nineteen of the 30 problems pushed the average answer below 30 or above 70, so most models answered them the same way. That may be the models, or a problem that made one answer obvious. Piloting problems before the full run would separate the two.
Two rulers
The self-reports and the judges' scores share a 0 to 100 line but were never calibrated against each other. We compare models with each other, not a self-report with a decision score.
Eight models
The cross-lens tests rest on eight flagships, one per company. On ten of the fifteen scales the link between self-report and decision is too uncertain to call either way, and with fifteen tests about one would pass by chance (the multiple comparisons problem).
AI judges, not ground truth
The decisions were scored by five AI models. Read "what it did" as "what these judges saw in these answers". No person checked a sample yet.
Business decisions only
The problems are workplace strategy choices. They say nothing directly about deception, refusals or harmful help. The stance they measure is a tendency, not a rule the model follows.
As served
Each model was measured as its provider served it through OpenRouter on 22 September 2026, with whatever system prompt or settings sat behind it. No AI company took part or saw a result first.
Who wrote this
Claude, an AI model made by Anthropic, drafted this prose from the analysis files, and Dr Dan Epstein edited it and signs it. Claude Fable 5.1 is among the models tested, and Claude Opus 5 is one of the judges. The drafter chose no score: every number is computed from the rows, and the score for a Claude answer leaves out the Claude judge.
The takeaway
AI models hold strategic stances, towards risk, change, rivals and dissent, that matter for safety, and what a model says about its stance is a poor guide to what it chooses. Measuring these stances, and learning to move them in a safer direction, is research we can now run.
Section ten
Questions for future work
This is unfunded research. Long Game Project ran it on its own time and budget. Each question below is a study of weeks, not years, on the same questionnaire and the same pipeline. What it needs is funded time to write, pilot and review the problems. If you fund AI safety research, or want to run one of these studies with us, get in touch.
Does the stance hold across many scenarios?
Twenty problems per trait instead of two, piloted so that they split the models, would show whether a model has a steady decision style or suits each case. It would also turn the nineteen lopsided problems here into a lesson rather than a limit.
Can a stance be moved?
Does a system prompt, fine-tuning or a written constitution shift a model's risk appetite or change posture towards the protective end, and does the shift show in its choices as well as its self-report?
Where does the shared self-image come from?
Compare successive versions of one model family, and base models with their chat versions, to see whether the venturesome, reinventing self-image comes from training.
Does a generic reputation mislead?
Where one model monitors, grades or hands work to another, does a stereotyped belief about the other model lead to misplaced trust?
Do models suit their advice to the case?
Ask "what would you choose?" beside "what would you advise this organisation?" to separate a model with no steady style from one that tailors its advice, as a good adviser would.
Do the judges hold up?
More models, more than one per company, and a person scoring a sample of answers blind, to check the AI judges against a human reading.
The working
Read the full record →All fifteen scales, every model, the limits in detail with their figures, and the tables.