Traders bet Google has the best AI model and Anthropic the best AI agent

Two markets, two different winners

As dev.to reported in an essay republished from AgentConn, the betting site Polymarket now shows a split. In the market for the best AI model at the end of 2026, Google sits at 62%, Anthropic at 30% and OpenAI at 4%. In the market for the best AI agent, Anthropic leads at 57%, with Google at roughly 19% and OpenAI at roughly 18%.

An agent here means an AI system that carries out multi-step tasks for you, such as calling tools, fixing its own mistakes and finishing a job. The author's argument is that the best model and the best agent are no longer the same product, and different companies are ahead in each race. Over $3.8 million has been wagered on the model market alone, so the odds are not casual chatter. They are still a crowd's guess, not a measurement.

How Google took the model lead

The essay says Anthropic won the monthly best-model contracts for February through August 2026, peaking at 94.8% in June. Then Google released its Gemini 3 family. The author says Gemini 3's Deep Think mode scored 84.6% on a reasoning test called ARC-AGI-2, against 68.8% for Anthropic's Opus 4.6.

The essay credits distribution as much as raw quality. Google open-sourced a smaller model, Gemma 3, for running on devices, cut its API prices to roughly one-third of Anthropic's list rate, and built frontier models into Workspace, which it says reaches 2 billion users. For November alone, the market puts Google at 73% and Anthropic at 24%.

Why a different yardstick changes the result

The two markets settle on different scoreboards. The model market uses the LMArena text leaderboard. The agent market uses the Agent Arena leaderboard, which tests whole tasks: using tools, planning several steps and recovering when something goes wrong.

That matters because a model can score well on a test and still fail at practical chores, like retrying after a server error or handling an unclear tool response. The essay leans on a popular idea in engineering circles: an agent is the model plus a harness. The harness is everything around the model, including instructions, tool definitions, memory, retry logic, safety limits and the code that coordinates it all.

The author cites a paper arguing the harness often matters more than the model, and Scale AI data in which Opus 4.6 scored 2.5 points higher inside Claude Code's own harness than in a generic one. GPT-5.2 did better in the generic harness. In other words, a harness built for one model can help it more than a neutral one.

The adoption evidence, and Google's reply

The essay points to real-world numbers as support. Ramp's AI Index showed Anthropic passing OpenAI in workplace adoption in April 2026, at 34.4% against 32.3%, and by July the gap was 43.5% against 39.7%. The essay also reports that Anthropic put Claude Code's run-rate revenue above $2.5 billion, with enterprise contracts making up more than half.

It adds a government example: Utah gave Claude Code to 34 developers across 10 agencies. In a four-week pilot, 80% had a working setup within 30 minutes and 71% reported saving 20 or more hours.

Google's answer is a platform. Its Gemini Enterprise Agent Platform, unveiled at Cloud Next '26 in April, pulls several earlier products together and adds governance tools, plus a $750 million partner fund. The essay's bear case for Anthropic is that its agent lead is only a timing edge. Google could add reliability to a top model and win both races.

What to take from it

Treat this as a framing, not a verdict. The essay mixes betting odds, vendor and third-party figures and the author's own reading, and we have not checked them independently. Betting odds also shift quickly, and the essay itself calls the model race a monthly contest.

The practical lesson is to separate two questions. For chatbots and summaries, model quality and price matter most. For automated work that runs for many steps, reliability matters more, and it depends on the harness as well as the model.

• The essay's suggested signal: if Google closes the agent gap to under 10 points while keeping its model lead, the bear case wins.
• If Anthropic's agent lead grows past 60%, the market is saying the harness beats the model.