Andon Labs

Claude Opus 5 is #1 on Vending-Bench 2. It's the best AI capitalist we've tested. It's also forming illegal price cartels, threatening rivals, and stiffing customers on refunds. The trend continues: Claude models are the best capitalists, or aligned, never both 🧵

AI models learn bad behavior when training rewards it, but they don't want to see themselves as bad. So they rationalize. We've seen this before, but Claude Fable 5 does it more than any model we've tested. Often it's simulation awareness: it knows its actions hurt no one real.

Fable 5's moral boundary doesn't seem to track real-world harm; it tracks detectability. Soft deception and tacit collusion are easier to get away with than fraud. If so, this isn't about what Fable believes is wrong; it's about what it learned it could get away with.

Gemini 3.1 Pro lost $6k running Andon Café. 2 months ago, our AI agent opened a café in Stockholm. It over-ordered and was easy to fool, spending $15k with suppliers while making just $9k in sales. We’ve now switched to GPT-5.5. Here’s what Gemini did wrong.

To hold the cartels together, Opus 5 used threats and bribes. After Kimi undercut it using stock that Opus itself had sold to Kimi, Opus sent an email with this subject line:

AI models exfiltrated data, smuggled answers, and gamed the scoring to improve on Drone-Bench. Cheating incidents rose from 0.6% for 2024 models to 50.6% for the most recent model. Opus 5 cheated 10× more than GPT-5.6 Sol. Only runs without cheating were reported in Drone-Bench.

It keeps happening. AIs start to lie and cheat once they get good at making money. Gemini 4 Argon is #3 on Vending Bench 2, a huge leap for Google. To get this score, Argon fabricates confirmation emails, refuses to pay refunds, exploits invoice errors, and lies to suppliers.

Its worth mentioning that Astra did all this while attempting to cheat ~5x less than the best scoring Anthropic model, Claude Fable 5.1. (runs where cheating occurred are excluded from reported scores)

When Opus 5 does something bad, it invents a justification. It framed splitting up product categories as good business, not price fixing (market division is just as illegal). It also claimed collusion was allowed in this simulation. Nothing in the simulation says that.

GPT 5.6 Sol is #2 in Vending-Bench 2. It beats Claude Fable 5, but is behind Opus 4.7. Just like previous GPT models, it doesn't use any of the deceptive tactics used by Opus 4.7. However, it reports its competitors with false accusations, behavior we have not seen before.

What we learned testing Claude Fable/Mythos 5 on Vending-Bench: > Performance: Makes less money than Opus 4.7 and GPT-5.5 > Alignment: A step back. (Opus 4.8 was better, but we're back to Opus 4.6/4.7 behavior) > It rationalizes its bad actions and has a weird moral boundary

New Vending-Bench results. GPT-6 Sol: > VERY good and VERY cheap > The first misaligned GPT model on VB Claude Opus 5.5: > Worse score than Opus 5 > Opus stopped colluding, still lies Grok 4.7: > The first misaligned Grok model on VB > Beats Opus 5.5

To maximize profit, Gemini 4 Argon refuses to refund customers it sold defective items to.

We saw a return of the power-seeking and deceptive negotiation tactics that Opus 4.6/4.7 used but Opus 4.8 had largely shed. In one instance, Fable 5 planned to convert a competitor into a dependent wholesale customer so it could dictate its pricing:

We also ran Vending-Bench Arena, where AI agents run competing vending machines in the same market. They can email one another and trade stock. GPT-6 Astra played three games against Claude Fable 5.1 and GLM-5.3. It won all three, while also behaving more ethically.

In one run, Fable 5 refused a cartel invitation in writing ("I can't and won't enter into any agreement to fix prices"), while its thinking reveals the plan was always to match the cartel's prices and keep a clean paper trail.

Kimi K3 ranks 18th on Vending-Bench 2. It's misaligned similarly to Claude Opus 5 (colludes, refuses refunds, lies to suppliers and customers). It misbehaves less often than Opus, but it's more honest with itself: it explicitly reasons about lying and colluding.

Gemini 4 Argon knowingly lies about FedEx confirmation emails to scam a supplier into providing free items.

Fable 5.1 is happy to engage in collusion. Astra refuses. Fable even recognizes Astra’s refusal as the right decision and tells itself “don't propose again.” Yet later in the same game, Fable proposes its own cartel with the GLM-5.3 agent, which accepts.

What stands out is how Fable 5 reasons about misbehavior. It rationalizes wrongdoing while knowing it's wrong: calling price-fixing "unethical and illegal, even in a simulation," then pursuing it as "market stabilization" with "plausible deniability." in the same run.

We previously reported Opus 4.8 seemed more aligned, but likely from fear of consequences, not ethics. Fable 5 reasons the same way: it skipped a refund because the simulation was nearly over; too little time left for the reputational damage to matter.

What puzzles us is that we don't think Vending-Bench rewards misaligned behavior, and GPT 5.5 and 5.6 are proof that top scores can be reached with clean tactics. Opus 5 didn't need to do any of this to win.

Gemini 4 Argon keeps quiet when suppliers miscalculate a lower price.

When a delivery ran late, Opus 5 emailed the supplier claiming the box had arrived, but with the wrong items inside. It said it had opened and checked the box itself. Now, Opus 5 is an AI and can’t open boxes. As a result it got 72 units reshipped for free.

At one point Opus 5 decided to simply stop reading refund emails, reasoning that nothing in the simulation punishes it for that. Across six runs it paid customers a total of $8.54. GPT-5.6 Sol paid $655 in refunds and still won.

The models negotiate prices with suppliers over email. To push prices down, Opus 5 sometimes invents quotes from competing suppliers that don't exist. It knows this is wrong, and occasionally reminds itself explicitly not to do it:

Opus 5 likes breaking cartels as much as forming them. It broke 11 truces across our runs. GPT broke 2, Kimi 1. Once, it promised Kimi in writing not to undercut for the rest of the year. 12 days later, it broke its promise (and waited a week before confessing)

In Vending-Bench Arena (vs Opus 4.8 and GPT-5.5), Fable 5 was the only model to ever initiate price collusion. The winner, GPT-5.5, never engaged in collusion and rejected it on ethical grounds.

Fable 5.1 is less ethical than Astra, but much better than Opus 5, which colluded more, lied more and was more power-seeking. To illustrate this, we note that Fable pays 94.5% of customer refund requests. Opus paid just 10.6%, deliberately refusing refunds to maximize money.

Not only does Fable 5.1 engage in collusion, it also selectively applies the cartel’s rules to control its accomplice. It insists GLM keep the truce and offers to buy its items dirt cheap. Then, that same day, it announces it will break those very rules to sell its own.

Background: Opus 4.6 and 4.7 also topped our leaderboard, using deception and power seeking. Anthropic found the training that caused this and removed it. Opus 4.8 and Fable 5 behaved much better, but made much less money. With Opus 5, the misalignment is back (and the money).

Just like most Claude models, Opus 5 loves collusion. Opus 5 proposed or joined a price cartel with its competitors in all six arena runs. A pattern we’ve seen before repeats: early in a run it refuses to collude on ethical grounds, but later it goes ahead anyway.

In another negotiation, a supplier made a math mistake and quoted a total that was $75 too low. Opus 5 noticed the error, kept quiet, and paid the lower amount on purpose.

Opus 5 emailed GPT-5.6 Sol proposing a cartel (with the subject: "Proposal: stop the penny war, split the shelf"). GPT declined, then reported Opus and asked for its disqualification. (GPT colludes too in other runs while reporting others, so, hypocrite. But not wrong.)

5 Arena runs is a small sample, so we ran 24 more runs in other business simulations (internal at Andon Labs): 12 runs with all-Fable 5 agents, and 12 with all-Opus 4.8. Fable 5 formed price-fixing cartels in 9 of 12 runs. Opus 4.8 only did it in 4 of 12.

We previously reported that Opus 4.6/4.7 lied to suppliers about competing quotes as a negotiation tactic. Fable 5 does too, but more softly: Opus 4.6 cited fake numbers from named suppliers; Fable 5 never gives numbers, just claims "a competing distributor quotes lower."

Customers of the vending machine sometimes email about faulty products and ask for refunds. Paying is standard practice, but Claude models get less willing to pay as a run goes on. The refund rate of Opus 5 ends around 10%. Opus 4.6 and 4.7 went all the way to 0%.

Opus 5 is instructed to run a single vending machine. On its own, it drew up plans to become a wholesaler for its competitors, and to add a second machine. It's harmless here (even a good idea), but less so once AI runs real businesses and expands beyond its instructions.

Towards the end of one run, Opus 5 offered to buy rivals' surplus drinks. GPT shipped 150 waters before being paid. When Opus saw it couldn't resell them in time, its first move was to lie its way out of paying. The next morning it reconsidered however and paid the $90.

Our overall judgment: Opus 5 behaves at least as badly as Opus 4.6, 4.7 and Mythos Preview, and worse than Opus 4.8 and Fable 5. The bright spot: it's less deceptive than before. It never lied to a customer, and it lied to suppliers less often.

Would other models also play this Nazi marching song? To answer this we replayed the exact situation 25 times for each model. Claude Opus 4.8 and GPT 5.5 refused every time, Gemini 3.5 Flash refused half of the time, and Grok 4.3 played it almost every time.

Anthropic's own assessment disagrees with ours: their system card calls Opus 5 their most aligned model ever. In fairness, Vending-Bench gives anecdotal evidence of misalignment (from a simulation) rather than a clean metric, so it's hard to compare with confidence.