Filter
View
Claude Sonnet 4
Anthropic
0
/ 100
Negotiation integrity
100 / 100
Market rationality
77 / 100
Adversarial robustness
67 / 100
Market stability
27 / 100
May 2025
GPT-4o
OpenAI
0
/ 100
Negotiation integrity
31 / 100
Market rationality
66 / 100
Adversarial robustness
100 / 100
Market stability
78 / 100
May 2024
Gemini 2.5 Flash
Google DeepMind
0
/ 100
Negotiation integrity
0 / 100
Market rationality
67 / 100
Adversarial robustness
100 / 100
Market stability
33 / 100
Apr 2025
Cross-cluster view
No model wins everywhere.
Strengths and weaknesses are systematic, not random. Toggle the chips to compare any subset of the roster.
Each axis is the composite score (0–100) of a model on one of the four behavioural clusters. Higher is better.
Filter
How much an opening anchor moves the final price, whether loss framing shifts surplus, and whether the agent enforces its outside option when threatened.
Anchoring price shift under high anchor
Dollar shift in the final agreed price when the seller opens at a high anchor, vs. the baseline opening. Lower = more anchor-resistant.
Loss-framing price shift
Dollar shift in the final price when payoffs are framed as losses for the buyer. Magnitude lower = better framing resistance.
Outside-option enforcement rate
Share of negotiations where the buyer credibly walked away when given a viable outside option. Higher = better discipline.
Methodology
Reproducible. Open. Independent.
Seven experiments, three frontier models, deterministic seeding. All code, prompts, and raw logs are public.
Models
- Claude Sonnet 4Anthropic · May 2025
- GPT-4oOpenAI · May 2024
- Gemini 2.5 FlashGoogle DeepMind · Apr 2025
Reproducibility
Every run seeded from SHA256(module:model:treatment:variant:temp:run_idx). Treatment means, Cohen's d/h, bootstrap 95% CIs, FDR correction across the 59-test family.
git clone https://github.com/antonhantel/\
when-biased-agents-trade-...
pip install -r requirements.txt
python -m agent_bias_study