AI Agents Autonomously Beat Minecraft as ML Engineering and Cheaper Models Advance Together

AI agents are moving from demos to genuinely long-horizon autonomous tasks. In the latest round of technical updates, a team of Gemini 4 Argon agents beat Minecraft with no human intervention, MLExpert nearly doubled the Kaggle medal rate, and Anthropic and DeepSeek continued to push down the cost of models and tools.

Long-Horizon Agent Tasks: From Games to Software Engineering

Gemini 4 Argon Beats Minecraft

Four Gemini 4 Argon agents entered a sandboxed Minecraft survival world with the goal of defeating the Ender Dragon. Each agent controlled one player character through the open-source library Mineflayer, coordinating via in-game chat and shared memory, with no human assigning roles or writing game logic.

According to arXiv paper 2610.22319, the agent team wrote 2,783 programs during the run, taking 27 hours and 14 minutes to defeat the Ender Dragon while dying 605 times along the way. The run started from a randomly generated world on Easy difficulty. This is the first time AI agents have completed Minecraft end-to-end autonomously.

Mineflayer and the Minecraft agent benchmark

Mineflayer is an open-source Minecraft bot library written in JavaScript. It wraps the game's low-level protocol into a high-level API, letting programs move, mine, craft, and interact with blocks and entities just like a player. Researchers commonly use it to build agent environments because it requires no modification to the game client—agents act entirely through standard interfaces.

Minecraft has become a long-horizon agent benchmark because it combines an open world, sparse rewards, and long-range dependencies: to defeat the Ender Dragon, agents must first gather resources, craft tools, enter the Nether, and find a fortress—any failed step can break the plan. Most prior agent demos relied on pre-scripted behavior, hand-written game logic, or specific seeds, whereas end-to-end autonomous completion requires agents to decompose goals themselves, write programs, and recover from failure—which is exactly why this task is seen as a milestone.

Why This Matters

The agents maintained a coherent plan over more than a day and across thousands of interdependent steps, using only the standard Mineflayer library—no pre-scouted seeds, no human-written game tools. Long-horizon programming and multi-agent coordination are equally critical for real-world software engineering.

The value of this result lies not in the game itself, but in validating the full loop of goal decomposition, code writing, and trial-and-error recovery in an open environment. The 605 deaths show the process was far from elegant, but the final goal was achieved—which is closer to the norm in real engineering projects.

ML Engineering Agents: Medal Rate Nearly Doubles

A 35B MoE Agent Model

Researchers introduced MLExpert, a 35-billion-parameter mixture-of-experts (MoE) agent model aimed at autonomous machine learning engineering. The team built Kaggle-MLE Factory to generate executable tasks, used hierarchical agent supervised fine-tuning to transfer coding ability to MLE tasks, and applied windowed interactive reinforcement learning to treat multi-turn decisions as execution windows with shared rewards.

Mixture-of-Experts (MoE) architecture and MLE-Bench

Mixture of Experts (MoE) is a sparse-activation model architecture: the model contains multiple parallel "expert" subnetworks, and each token is routed to only a few of them for computation, so the total parameter count can be large while the actual compute per inference is far smaller than an equivalent dense model. MLExpert's 35 billion total parameters use exactly this approach to gain substantial capacity while keeping inference costs manageable—which is the premise for comparing it with larger frontier models.

MLE-Bench is a benchmark for evaluating machine learning engineering agents. Its tasks come from real Kaggle competitions, and agents must complete data exploration, feature engineering, model training, and submission within a limited time. MLE-Bench Lite is a streamlined subset that measures performance by average medal rate (the proportion reaching bronze medal or above). A jump from 37.9% to 72.7% means the agent can now produce competitive submissions on most tasks, not merely complete the pipeline.

Medal Rate Data

On MLE-Bench Lite, MLExpert raised the base model's average medal rate from 37.9% to 72.7%, on par with larger frontier models such as Kimi-K3 and GLM-5.3. According to arXiv paper 2610.22344, this post-training method can extract useful supervision signals from long experiment trajectories, and asynchronous collection supports policy updates before trajectories end.

Generalization Ability

MLExpert also improves on NatureBench Lite and helps optimize agent scaffolding and algorithm optimization. The learned optimization ability generalizes to other tasks, pointing to an AI-for-AI path and the possibility of recursive self-improvement.

The Price-Cut Wave: A Race Toward Small Models and Low Cost

Claude Haiku 5.5 Targets High-Concurrency Scenarios

Anthropic launched Claude Haiku 5.5 on October 7, 2026, calling it its fastest, cheapest, and most capable small model to date, aimed at high-concurrency, cost-sensitive workloads. Developers can deploy large-scale inference without significantly increasing costs.

Small models and per-task cost

In agent workflows, a single task often requires dozens to hundreds of model calls, and cost is driven mainly by "price per call × number of calls" rather than the absolute capability of a single inference. Small models (such as the Haiku series) are cheap per call with low latency, making them suitable for high-frequency substeps like tool calls, format conversion, and state summarization; flagship models intervene only at key decision points. This tiered invocation is the mainstream approach to reducing agent operating costs today.

That is why Anthropic is cutting unit costs across all three product lines—Haiku, Sonnet, and Opus. The significance is not merely lower prices, but making longer autonomous workflows economically viable. When a single call is cheap enough, agents can afford more trial-and-error and verification steps—a direct link between the reliability of long-horizon tasks and cost.

Sonnet and Opus Updates

Claude Sonnet 5.5 was released on September 28, 2026, running 30% faster than Sonnet 5 and cutting the cost of most work by up to 30%. Claude Opus 5.5 launched on September 22, 2026, reaching the level of Claude Fable 5.1 on most work while running 40% cheaper than Opus 5. Anthropic is simultaneously lowering unit costs across its small, mid-tier, and flagship lines.

DeepSeek Adds Web Access

DeepSeek^Ad updated V4.1 Flash to version V4.1-Flash-1017, with significantly improved project development capability, lower task completion costs, and faster speed. DeepSeek also added API support for web search and scraping, billed at $0.75 per 1,000 successful requests. Developers can now have models work on real-time web data, building research agents that can browse and cite sources.

Three Threads Point in the Same Direction

Agent capability, engineering automation, and inference cost are advancing in parallel. Gemini 4 Argon proves long-horizon autonomous tasks are feasible, MLExpert shows post-training can push small models close to frontier level, and Anthropic and DeepSeek are lowering the price per call. For developers, this means more complex autonomous workflows are becoming economically viable; for the industry, the competitive focus is shifting from model size to per-task reliability and cost.

FAQ

What did the Gemini 4 Argon agents achieve?

They autonomously defeated the Minecraft Ender Dragon, with the team writing 2,783 programs and taking 27 hours and 14 minutes while dying 605 times.

How much did MLExpert improve the medal rate?

MLExpert raised the average medal rate on MLE-Bench Lite from 37.9% to 72.7%, on par with larger models such as Kimi-K3 and GLM-5.3.

What is Claude Haiku 5.5?

It is Anthropic's fastest, cheapest, and most capable small model, released on October 7, 2026, targeting high-concurrency, cost-sensitive work.

How is DeepSeek web search billed?

DeepSeek charges $0.75 per 1,000 successful search and scraping requests, an API launched alongside the V4.1-Flash-1017 update.

When was Claude Sonnet 5.5 released?

Sonnet 5.5 was released on September 28, 2026, running 30% faster than Sonnet 5 at up to 30% lower cost.

One-line Summary

Long-horizon agent tasks, engineering automation, and falling inference costs are all advancing in parallel.

Paragraph Summary

This article reviews three recent developments in AI agents and models. First, long-horizon autonomous tasks: four Gemini 4 Argon agents, with no human intervention, controlled characters via Mineflayer and coordinated through chat and shared memory, taking 27 hours and 14 minutes, writing 2,783 programs, and dying 605 times before defeating the Minecraft Ender Dragon—the first end-to-end autonomous completion. Second, machine learning engineering automation: the 35B MoE model MLExpert, using Kaggle-MLE Factory, hierarchical supervised fine-tuning, and windowed reinforcement learning, raised the average medal rate on MLE-Bench Lite from 37.9% to 72.7%, approaching larger frontier models and showing generalization to other tasks. Third, cost reduction: Anthropic launched Haiku 5.5 and updated Sonnet 5.5 and Opus 5.5, while DeepSeek updated V4.1-Flash-1017 and added pay-per-request web search and scraping APIs. Together, the three signal that more complex autonomous workflows are becoming economically viable.