Faraday Agent Beats GPT-5.5: What It Means for AI Shopping Agents

A small lab posts a big claim

Inherent is a London AI lab started by former Google DeepMind researchers. In August it said its new agent, Faraday, had beaten much larger frontier systems at one specific task: reproducing the findings of published scientific papers without being given the answers first.12 The systems it says it outperformed were Anthropic's Claude Opus 4.8 and OpenAI's GPT-5.5. Faraday runs on Qwen 3.6, a model with only 27 billion parameters.12 The announcement came a few weeks after the startup came out of stealth with a $50 million seed round. The company has about a dozen employees and plans to grow to roughly 20 to 25 by the end of the year.12

The coverage tells a consistent story. Every outlet describes the result as a small model beating frontier labs on research replication, and treats it as a sign that agent design may matter as much as model size in specialized fields.1113 The more careful write-ups also add the caveat that the claim still needs independent checking before it counts as a settled result.1113 That caveat matters. Inherent's comparison is company-reported, and none of the coverage describes an outside replication of the benchmark.

Faraday is a research agent, not a shopping bot. Still, it is worth reading closely for anyone building or paying for AI shopping agents. The questions it raises about how agents are built, whose models sit underneath them, and how far they can be trusted are the same questions now facing agentic commerce.

Orchestration over scale

The most telling detail is that Inherent did not build its own coding tool. Faraday calls OpenAI's GPT-5.5 Codex when it needs code, much as a human scientist uses existing software.12 So the agent that "beat GPT-5.5" was partly running on a GPT-5.5-family product. The fair reading is not that a small model overpowered a giant. It is that a well-designed system, trained to decide what to do and when to call which tool, beat general-purpose frontier agents on a narrow job.

Co-founder and chief scientist Edward Hughes said this himself. He told TechCrunch that the way the team built the system interested him more than the win.12 Inherent used reinforcement learning, which rewards outcomes instead of writing out rules, to give Faraday what the company calls "research taste": a sense of which experiments are worth running.12

Inherent is not alone here. In May, Google DeepMind's own AI co-mathematician set a FrontierMath Tier 4 record of 47.9%, ahead of GPT-5.5 Pro's 39.6%. It did this with a multi-agent setup coordinated by Gemini 3.1 Pro, not a new foundation model.16 Researchers have also reported turning published papers into agents that answer questions about them. On 300 questions those agents scored 91.2%, against 80.3% for Claude working directly from each paper's code, and they were cheaper per question.18 The pattern repeats across these cases: specialized orchestration beats general capability on targeted tasks.

The reliability problem has not gone away, though. The ResearchGym benchmark found that a GPT-5-powered agent improved on the provided baselines in only 1 of 15 end-to-end runs. Its failures included impatience, overconfidence in weak hypotheses and poor resource management.19 Faraday's result should be read with that in mind. Being good on a benchmark is not the same as being dependable over long, unsupervised tasks.

Why commerce builders should care

Agentic commerce, where AI agents find, compare and pay for products for a consumer, has the same structure as Faraday: a reasoning layer that calls outside tools. In commerce those tools are protocols, not code interpreters. OpenAI and Stripe's Agentic Commerce Protocol (ACP) covers checkout and payment. Google and Shopify's Universal Commerce Protocol (UCP) covers the whole journey from discovery to after the sale. Google's AP2 handles proof of payment authorization. Anthropic's Model Context Protocol supplies the data connections.1 One industry explainer splits these into three layers: what to sell, how to pay inside an AI interface, and who authorized the purchase.7

My reading is that if small, tuned agents can match frontier models on narrow work, the brain behind a shopping agent becomes swappable. A retailer, bank or startup could, in principle, run a compact model trained for comparison shopping and spending discipline, and plug into the same checkout rails as ChatGPT or Gemini. The value would then move toward whoever controls the protocols, payment tokens and trust signals.

The infrastructure side is clearly preparing for many kinds of agents. Adyen launched Adyen Agentic on June 16. It supports UCP, ACP and AP2 together, so merchants don't have to bet on which standard wins.8 On the network side, Visa's Trusted Agent Protocol verifies an agent's identity cryptographically when a transaction happens. Mastercard's Agent Pay combines agentic tokens with a "Verifiable Intent" record attached to each transaction.4 American Express added purchase protection for registered AI-agent purchases, which means all three major U.S. card networks now support agentic transactions.1 Mastercard has also built Know Your Agent frameworks meant to tell legitimate shopping agents apart from malicious bots.9 That vetting will matter more if the number of capable agents keeps growing.

Where the reporting diverges

Accounts of how far agentic checkout has actually got do not agree. One industry guide describes ACP as letting transactions finish entirely inside the AI interface.2 Forrester reported in March that OpenAI was shutting down its Instant Checkout feature because of weak performance. ChatGPT apps do not take payments directly; they send shoppers to merchants' own checkout pages.5 A more recent overview says ChatGPT Shopping recommends products and sends users to merchant sites.1 I think the Forrester account is the better guide to where the market stands: discovery is happening inside AI chats, and most payments are still happening on merchants' own sites.

Conversion data backs this up, and it is mixed. Adobe found that visitors arriving from AI sources during Prime Day 2026 converted 40% better than other channels. Walmart, by contrast, saw conversion rates three times lower for purchases made inside a chat than for shoppers redirected to its own site.10 Cost is another factor. One analysis puts the total fee on agent-led ACP conversions at about 7%: a 4% OpenAI platform fee plus roughly 3% for Stripe processing.9 That could push merchants toward cheaper agent setups.

Forecasts vary widely too. Stripe and BCG estimate that 55% of U.S. e-commerce spending could be agent-assisted, with 18% driven by fully autonomous agents, which they put at about $1.3 trillion in affected spending.10 Payments analysts quoted elsewhere expect mainstream consumer adoption in 2027 or 2028, not 2026.4

The trust gap is the real benchmark

Consumer surveys point to the limit no benchmark score can fix. Worldpay's 2026 report covers seven markets. It found willingness to shop with AI agents rose everywhere year over year, from 40% of consumers in Australia to 86% in China.6 But once money is involved, people pull back. Among shoppers open to AI shopping, 80% would let an agent play some part in payment, yet 59% want to approve every purchase themselves.6 Fraud is also a concern: 78% of financial institutions expect fraud to rise sharply as shopping agents spread. When an agent triggers a purchase, it is unclear whether the merchant, the agent provider or the payment processor is responsible for a dispute.9

This is where Faraday's design goals stand out. Hughes said Inherent wants to avoid agents that simply tell users what they want to hear. He wants an agent that explains what it tried and asks for your opinion of the results.12 A shopping agent that shows its reasoning, flags trade-offs and asks before spending fits exactly what Worldpay's respondents say they want.

The verdict

I would read Inherent's result as a signal, not proof. It is self-reported and narrow, and it was partly achieved by calling OpenAI's own tooling.12 Still, it adds to growing evidence, from DeepMind's math system to paper-based agents, that specialized orchestration can beat raw scale on defined tasks.1618 For agentic commerce, if that holds up, the shopping agent itself becomes easier to replace and the payment rails become harder to dislodge. The companies building identity, authorization and dispute systems, such as Visa, Mastercard, Stripe, Adyen and Google, are betting that the protocols will last longer than any one model.48 Whichever agent wins over shoppers will need to be careful before it spends their money, not just capable.