OpenAI vs Anthropic This Week: 3 Edges to 2, With 1 Tie

Across six head-to-heads we reported between 3 and 6 November, OpenAI came out ahead in three, Anthropic in two, and one was a tie, by our count. OpenAI led on a deception game, on cost at a tied banking-task score and on staff moves we verified; Anthropic led on a shopping-agent test and on traders’ odds of a listing this year. Each row measures something different, so the tally is a scoreboard of this week’s reporting, not a ranking of the two companies.

Business Pill · Verification Cost

An evergreen Business Engineer explainer on the cost of checking what AI produces. It teaches the general idea only and does not discuss these results.

The Six Head-to-Heads

• Deception game: OpenAI. Clocktower Bench rates GPT-6 Astra at 1859 Elo against 1595 for Claude Fable 5.1, Anthropic’s best, across 386 rated games, a 264-point gap (our piece).
• Banking tasks: tie. On DeepFinance, Claude Opus 5.5 met 54.8% of the rubric criteria and GPT-6 Astra 54.2%; Julius Research says the top four, two Claude and two GPT models, are “not statistically distinguishable” (our piece).
• Cost at a tied banking score: OpenAI. In that statistically tied group, GPT-6.1 Sol scored 52.9% at $0.81 per task against $8.63 for Claude Opus 5.5; by our arithmetic Opus cost about 10.7 times as much for 1.9 more points (our piece).
• Shopping agents: Anthropic. In Visa’s results data, Claude Fable 5.1 passed 75.6% of runs against 64.4% for OpenAI’s best, GPT-6 Astra; neither lab led the test, which Meta’s Muse Wonder and Google’s Gemini 4 Argon topped at 82.2% (our piece).
• Odds of an IPO by the end of 2026: Anthropic. At 18:10 UTC on 5 November, Polymarket prices implied 79.15% for Anthropic and 2.75% for OpenAI. These are traders’ prices, not forecasts (our piece).
• Staff moves we verified: OpenAI. Our tracker counts two people who posted that they left Anthropic and joined OpenAI, Cat Wu and Austin Lau, and no verified move the other way (our tracker).

What We Did Not Score

Some of this week’s biggest numbers have no matching figure for the other lab, so they are not on the scoreboard. Anthropic’s run rate topped $80 billion at the end of October, according to a person familiar with the matter cited by The Wall Street Journal (our piece), while OpenAI president Greg Brockman spoke of “70% quarterly growth in our annualized revenue” in the third quarter without giving a dollar figure (our piece). METR estimates Anthropic is improving its models at about twice its earlier rate (our piece); we have no comparable assessment of OpenAI.

Why the Tally Is Close

The three OpenAI edges are a game rating, a price and a head count of two people; the two Anthropic edges are a pass rate and a market price. Four of the six rows come from three benchmarks built by outside groups, each testing one kind of work, and a different week of coverage could produce a different count.

The Structural Read

As we read the scoreboard, OpenAI’s edges this week came from a game rating, a price and two staff moves, and Anthropic’s from a shopping test and the IPO market.

As we read the banking rows, the two labs were level on quality in that test, and the gap was in what the result cost.

Three Implications

FOR BUYERS OF AI AGENTS Results differ by task: Claude beat OpenAI’s best on Visa’s shopping test, and GPT-6 Astra led the deception game.

FOR FINANCE TEAMS At a tied DeepFinance score, the API cost per task differed about 10.7 times, by our arithmetic.

FOR READERS OF LAB-RACE HEADLINES This scoreboard is one week of reporting; each row is a different measure.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Agentic Harness War.

The framework says: “Expertise shifts toward reviewing outputs, coordinating multiple agents, and making strategic decisions.”

As we read it, a scoreboard like this one is part of that review work: which model to use depends on the task being checked. The framework does not discuss these results.

What Is Not Established

The edges are our reading of each figure, not a combined score, and the rows should not be added up as one contest. The benchmarks are each one organisation’s test; Polymarket prices move continuously, and the 18:10 UTC snapshot predates later moves in Anthropic’s odds. The staff count covers only moves confirmed by the people’s own posts between 24 October and 5 November.

The Bottom Line

On this week’s six like-for-like comparisons, OpenAI took three, Anthropic two, and one was a tie.

95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Every figure comes from FourWeekMBA pieces published 3-6 November 2026, each linked, which cite the Clocktower Bench site, Julius Research, Visa’s results data, Polymarket prices and the people’s own posts. The edge in each row, the count of edges and the differences are our reading and arithmetic.