On February 17, 2026, xAI pushed Grok 4.20 into public beta. The headline feature is not a benchmark score but an architecture: the model ships with four named agents — Grok (the captain), Harper (research), Benjamin (logic), and Lucas (creativity) — that debate with each other in real time during inference before a single answer comes out. eWeek’s coverage called the result “a debate team.”
For developers, this is the first frontier model to move multi-agent collaboration into the model itself rather than an external framework. If the early numbers hold, inference-time debate could graduate from an experimental trick to a product default.
Not One Model, Four Agents
Per NextBigFuture’s teardown, each agent has a defined remit. Grok decomposes tasks, routes subtasks, resolves conflicts, and writes the final synthesis. Harper handles research and verification, grounding answers with real-time search plus the X firehose — roughly 68 million English posts per day. Benjamin stress-tests math, code, and logic. Lucas supplies divergent thinking, flags blind spots, and polishes the output.
The critical detail is that the mechanism is native. It lives inside inference, not assembled outside the model with a framework like AutoGen or Swarm. OpenRouter’s model page describes the beta as xAI’s newest flagship with industry-leading speed and agentic tool-calling capabilities, and the lowest hallucination rate in the family.
How the Debate Loop Runs
The loop has four steps. First, Grok breaks the task down and assigns subtasks. Second, all four agents analyze in parallel rather than in sequence. Third, they enter a multi-round internal debate: Harper checks facts, Benjamin checks logic, Lucas flags bias, iterating toward consensus. Fourth, Grok synthesizes one output — and the agent traces can optionally be surfaced to the user.
Simple queries do not convene the full council. The debate triggers only for complex tasks, which is how average cost stays under control.
The Numbers: Parameters, Cost, Compute
NextBigFuture’s figures: a roughly 3-trillion-parameter mixture-of-experts base, with the four agents as specialized replicas of the same model; claimed effective-intelligence gains of 2–4x; marginal compute cost of 1.5–2.5x a single pass (versus a naive 4x); and reinforcement learning credited with 6x efficiency gains in orchestration. Training and inference run on the Colossus cluster’s 200,000-plus GPUs, with shared weights and KV cache across the replicas.
Access: consumers get it through SuperGrok at about $30 a month or X Premium+. The API is not fully open during the beta; pricing is expected to sit above Grok 4.1 Fast’s $0.20 and $0.50 per million input and output tokens, with Batch API and cached-token discounts softening the bill.
The 65% Hallucination Cut and Live Trading Tests
eWeek reports two early results: hallucinations down about 65% in internal testing, and wins over rival models in live trading tests. Both come from xAI’s own numbers and await independent verification, but the direction is consistent with the mechanism. A verification-focused agent holding a real-time data feed, plus a logic agent whose job description includes saying no, makes the debate itself a hallucination filter.
From Grok 4 Heavy to Native Multi-Agent
This line did not appear from nowhere. Grok 4 Heavy, shipped in July 2025, already experimented with multiple model instances collaborating, but without named specialization or an explicit debate loop. Grok 4.20 productizes it: every agent has an identity, a remit, and observable interaction traces.
The contrast is OpenAI and Google, whose production stacks remain single-model inference scaled with internal search; their multi-agent work stays in frameworks and research. xAI has effectively pushed an architectural bet into production — that spending a little more compute at inference beats scaling one model alone. It fits the agentic-engineering turn that has defined the start of 2026 (see our 2026 opening outlook).
What Developers Should Note
Three things. First, agent-trace visibility is about to become a standard debugging surface: when an answer is negotiated by four agents, “why did it answer this way” needs new observability tooling. Second, the cost model changes: a 1.5–2.5x marginal cost is a cheap trade for correctness inside agent workflows, where one bad step can cascade. Third, single-model benchmarks no longer capture post-debate reliability — evaluation has to measure the system, not the model.
Sources
- xAI’s Grok 4.20 Turns AI Into a Debate Team — eWeek
- How the xAI Grok 4.20 Agents Work — NextBigFuture
- Grok 4.20 Multi-Agent — OpenRouter
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.