← Back to feed

Four model additions landed on Arena leaderboards between October 19 and 20, according to the Arena changelog. The most significant is Reflection AI’s beam-1.0-501b-a23b joining the main Text Arena — the first open-weight 500B-class sparse MoE to enter the flagship language leaderboard.

Beam on Text Arena

Reflection AI launched Beam on October 6 with 80.9% SWE-bench Verified, an Apache 2.0 license pending red-team, and a 501B sparse MoE architecture running 23B active parameters. Within two weeks it was added to Agent Arena and Code Arena WebDev. The Text Arena entry follows on October 20.

Text Arena is the broadest comparison surface in the Arena ecosystem: blind human preference votes across general conversation, reasoning, instruction following, and writing. Beam’s addition puts it in direct competition with every closed frontier model — GPT-6 Astra, Claude Opus 5.5, Gemini 4 Argon — under pairwise human judgment rather than coding-specific tasks. It is the first real read on whether Beam’s conversational quality matches its coding benchmark numbers.

Results will take days to accumulate before rankings stabilise. An initial ELO will appear once sufficient votes are in.

Muse Wonder Full (Max) on Agent Arena

Meta’s Muse Wonder Full made the Artificial Analysis Intelligence Index top four earlier this month, scoring 56 on the index and ranking second behind Claude Opus 5.5. The Agent Arena addition on October 20 is its first evaluation on multi-step agent tasks.

The “(Max)” designation indicates the maximum compute/effort budget, matching how other frontier models are evaluated at their highest capability tier. Agent Arena evaluates on realistic agentic workflows rather than single-turn responses — tool use, multi-step planning, iterative correction. Wonder Full’s agentic quality has been benchmarked on LiveBench 2.0 at 80.8, third overall behind Claude Opus 5.5 (85.0) and GPT-6 Astra (82.4), which gives a reference point before Arena votes accumulate.

Muse Wonder Full was also added to the Text Arena on October 19, one day before the Agent Arena entry.

MiniMax M3.1 Flash on Code Arena WebDev

MiniMax launched M3.1 Flash on October 19 — the same day as its Code Arena WebDev entry. The model is a 428B MoE with 23B active parameters, positioned as a faster and cheaper successor to M3, with improvements in agentic tasks, coding, and cybersecurity. Pricing from MiniMax is $0.10/M input, $0.55/M output, with a 1 million token context window.

Code Arena WebDev evaluates models on front-end development tasks with human judges comparing outputs. M3.1 Flash’s addition immediately on launch day signals MiniMax is trying to establish a benchmark presence from day one rather than waiting for organic listing.

What Changes

Beam’s Text Arena debut is the headline. Open-weight models have consistently struggled against closed-weight frontiers on subjective preference votes — human judges tend to prefer longer, more polished responses that closed models with larger RLHF budgets produce. If Beam holds its ground in general conversation, it strengthens the case that open MoE architectures with aggressive RL runs can close the preference gap as well as the benchmark gap.

Muse Wonder Full on Agent Arena is a test for Meta’s frontier ambitions. The model has moved from Muse Spark (competitive but below frontier) to Wonder Full (legitimately top-4 on intelligence indexes). Agent Arena puts it against GPT-6 Astra and Claude Opus 5.5 in the evaluation format that most closely mirrors real agentic deployment.

Rankings update daily. Check arena.ai for live ELO as votes accumulate.