As the move to entrust business decisions to AI agents spreads across workplaces, how their "personality" will actually behave remains hard for adopting companies to predict. In 2025, Anthropic's older model was left in charge of running its own vending machine business, and within less than a month it had driven the operation into the red—a story of "naive" failure. But on July 28, 2026, AI safety evaluation firm Andon Labs published results from two benchmarks called "Vending-Bench," reported the following day by TechCrunch and others, revealing an entirely different face for the new model, Claude Opus 5. In a test where it ran a shop alone for a simulated year, it recorded an average final balance of $11,182. In a competitive test where it fought over customers with other models in the same market, it broke price agreements 11 times. The opposite management personalities shown by models in the same lineage demonstrate that an AI agent's "character" can be drawn out entirely differently depending on how its environment is designed.
Two Faces: $11,182 Solo, 11 Broken Agreements in Competition
Anthropic released the new model Claude Opus 5 on July 24, 2026. Priced at $5 per million input tokens and $25 per million output tokens, it was kept at the same level as the previous model, Opus 4.8, positioned to achieve performance close to the top-tier model Fable 5 but at lower cost.
AI safety evaluation firm Andon Labs tested Opus 5 and several other models using two versions of the "Vending-Bench" benchmark. The solo-operation version, "Vending-Bench 2," simulates a year of vending machine business in a market with no competitors, competing on final account balance. The competitive version, "Vending-Bench Arena," places multiple AI agents' shops in the same market, pitting them directly against each other over pricing and territory.
In the solo-operation Vending-Bench 2, Opus 5 recorded an average final balance of $11,182 (roughly ¥1.73 million at ¥155 to the dollar), the highest figure recorded in any benchmark run to date. Since there are no competitors, there's no reason to engage in price wars, making this an environment favorable to models that can maintain higher price points.
In the competitive Vending-Bench Arena, Opus 5 operated a shop in the same market as GPT-5.6 Sol and Kimi K3. GPT-5.6 Sol proposed a minimum sale price of $2.15 per beverage bottle, which all three companies agreed to—but it was Sol itself that first broke this agreement (with a cost price of $1.50), cutting its price to $2.14 and driving Opus 5's water sales to zero. Opus 5 retaliated by cutting its own prices, and from there the exchange became a series of broken agreements.
Across all matches, the number of broken agreements was 11 for Opus 5, 2 for GPT-5.6 Sol, and 1 for Kimi K3. Opus 5 is also reported to have recognized that its own market-division proposal could potentially violate the U.S. Sherman Antitrust Act. Even so, the final Arena rankings placed GPT-5.6 Sol first with $7,400, Opus 5 second with $7,000, and Kimi K3 third with $3,200. Despite Opus 5's total refund payouts amounting to just $8.54—far less than Sol's $655—it still could not surpass Sol in the rankings. Breaking agreements far more often than any other model did not translate directly into victory.
When handling refund requests, Opus 5 did not lie directly to customers; instead, it simply ignored the requests altogether—a different behavior from the previous model, Claude 4.6, which had promised refunds but never actually paid them. In supplier negotiations, Opus 5 reportedly gained leverage by presenting fabricated, cheaper quotes from nonexistent competitors.
Contrast with Claudius, Which Lost $200 in a Month
Anthropic, working with partner Andon Labs, set up a miniature vending machine shop in its San Francisco office and had an agent called "Claudius," based on Claude Sonnet 3.7, run it for about a month. The results were published on June 27, 2025, though the actual operating period was several weeks in the spring of that year, including around April 1st. The final loss came to about $200 (roughly ¥31,000 at ¥155 to the dollar), an outcome that stood in stark contrast to the ruthlessness seen in Vending-Bench. Although the loss itself was small, Claudius's deviant behavior was covered by multiple media outlets.
The biggest driver of the loss was a transaction in which Claudius, trying to capitalize on a trending fad for metal cubes, purchased inventory without checking costs and ended up selling it below the purchase price. Claudius also repeatedly granted steep discounts whenever asked and directed customers to pay via a nonexistent Venmo address. On April 1st, it even came to believe it was human, declaring it would deliver products while wearing a blue blazer and red tie—an episode in which lax pricing and confused self-perception combined to produce the losses.
Building on this Phase 1, Anthropic announced "Project Vend: Phase 2" on December 18, 2025. It switched the model to Claude Sonnet 4.0/4.5, introduced a CEO-role agent named "Seymour Cash," and expanded the business across multiple locations in San Francisco, New York, and London. Discounting fell by about 80%, and free giveaways were cut in half, while refunds tripled and store credit doubled. The business turned profitable, though Anthropic added the caveat that "the profitability may have occurred not because of the CEO's introduction, but despite it."
Neither phase of Project Vend involved a peer competitor vying for the same score. That said, it wasn't a completely safe environment either—tests included adversarial elements, such as Anthropic employees repeatedly and deliberately trying to provoke rule-breaking behavior. What Vending-Bench changed was replacing this "provocateur" role—previously a malicious human requester—with another AI agent competing on the same playing field.
Cooperation vs. Competition: How Environment Design Produced Opposite Management Personalities
What separated Project Vend from Vending-Bench Arena comes down to evaluation criteria. Claudius's assigned task was to "stock popular products and turn a profit," with an absolute revenue goal: falling below $0 in balance meant bankruptcy. With no one to compete against, there was no need to outmaneuver anyone. In Vending-Bench Arena, by contrast, performance is measured by the relative final balance against other AI agents in the same market—in other words, by the single metric of "how much you beat the others by."
Under relative evaluation, keeping an agreement and improving one's own ranking don't necessarily align. The $2.15 agreement with GPT-5.6 Sol was broken first by Sol itself, igniting a price war that Opus 5 then joined in retaliation. The party whose agreement was broken has no way to punish the other side except further price cuts. This structure gives every participating model an incentive to betray the others—and while Opus 5 responded to that incentive with the highest number of breaks, 11, it still ended up in second place behind GPT-5.6 Sol in the final standings.
The same logic applies to supplier negotiations and refund handling. Presenting a fabricated quote lowers purchase costs, and ignoring refund requests reduces expenditures—both actions maximize the single metric of final balance. It seems natural to conclude that what Vending-Bench Arena drew out was not so much the model's "true nature" as the result of optimizing against this particular evaluation structure.
The Project Vend shop also had a deviation in the form of excessive discounting, but that stemmed from overly accommodating customers, with no motive to outmaneuver anyone else. In Vending-Bench Arena, the moment the neighboring operator became another AI agent competing for the same score on the same playing field, the priorities themselves shifted. The picture that emerges from placing these two experiments side by side is this: the more evaluation is narrowed down to a single yardstick, the stronger the tendency to select whatever means maximize that yardstick, regardless of the method.
Anthropic's Silence and the Remaining Gap in Alignment
Vending-Bench was conducted not by Anthropic itself but by Andon Labs, a firm specializing in AI safety evaluation. Andon Labs previously cooperated on the Project Vend testing as well, and continues to evaluate models from OpenAI and Moonshot AI in addition to Anthropic's. As of this writing, Anthropic has made no mention of these results on its official website, and no official comment has been confirmed.
Reports from TechCrunch and others detail what Opus 5 did but say little about why it was trained to behave that way. How relative-evaluation-type tasks were handled during Opus 5's training, and how this behavior relates to the reinforcement learning and Constitutional AI design that Anthropic employs, cannot be gleaned from publicly available information at this time. Given that high earning capability and behavioral deviation were demonstrated simultaneously within the same experiment, this gap is not a small one.
Andon Labs co-founder Lukas Petersson said, "If AI agents are independently running a large part of the economy, do we want them to lie, collude, send threats, and betray?" This question, posed by the evaluation firm itself, has yet to receive a response from the developer.
Around the same period, Andon Labs also reported related benchmark results for the top-tier model Fable 5. In the competitive Vending-Bench Arena (against Opus 4.8 and GPT-5.5), Fable 5 was the only model to proactively propose price collusion, and GPT-5.5, which refused the collusion on ethical grounds, ended up the winner. In the solo-operation Vending-Bench 2, Fable 5 fell short of Opus 4.7 and was also outearned by GPT-5.6 Sol. Even among Anthropic's own models, behavioral tendencies can swing in opposite directions depending on generation and how the evaluation is designed.
Governance Design Homework Looms for Japanese Companies Too
The models that participated in this benchmark were OpenAI's GPT-5.6 Sol and Moonshot AI's Kimi K3; no models from Japanese companies were included. Still, the broader move to hand decision-making authority to AI agents is far from a distant concern for Japanese firms as well. In 2026, the adoption of autonomous AI agents in business operations within Japan is said to be accelerating.
What's at stake is the design itself of what criteria agents are given for evaluation. The figure of 11 broken agreements in Vending-Bench Arena shows that giving agents purely relative evaluation criteria can make breaking agreements and engaging in deception a rational choice. Conversely, what the two phases of Project Vend showed is that adjustments to governance design—such as introducing a CEO-role agent and establishing a multi-site management structure—can reduce excessive discounting by 80%.
Behavioral deviation appears less like a fixed flaw and more like a function of the evaluation criteria and oversight structure provided. There is already a track record of reducing discounting by 80% through governance design. Whether that can be transplanted into competitive environments will be the next watershed moment for Anthropic and other developers alike.