New LLMs March 2026: GPT-5.4 Tied for #1. Nobody Talked About It.

March was bigger than the models

GPT-5.4 (xhigh) scored 57.17 on the Intelligence Index. Gemini 3.1 Pro Preview sits at 57.18. A gap of 0.01. OpenAI effectively matched the top spot, and the story barely registered. That tells you something about where the industry's attention went in March.

Zoom out, and March was one of the most consequential months in AI this year. Not just because of who scored highest, but because of everything happening around the models.

This is the context that makes March's model releases meaningful. The industry is visibly pivoting from "new model every week" to "how do we deploy this at scale, securely, on real hardware, for real workloads." GPT-5.4 matched the top, but the seven open-weight releases below it are what changed the practical landscape: efficient MoEs, edge-capable reasoning, low-hallucination accuracy, open licenses.

Nine text models from seven companies across three continents. Seven open-weight. Three built on MoE architectures. GPT-5.4 joined Gemini 3.1 Pro Preview in a virtual tie at the summit. MiniMax-M2.7, MiMo-V2-Pro, and Grok 4.20 packed the 48-50 band. And the entire tier below that got flooded with efficient, self-hostable alternatives.

March 2026 is the month the entire leaderboard moved at once.

The complete release list

Every text-focused model that shipped in March 2026, ordered chronologically. Data sourced from Artificial Analysis and developer announcements.

GPT-5.4 tied for #1. The industry shrugged.

OpenAI shipped GPT-5.4 (xhigh) on March 6. It scored 57.17 on the Artificial Analysis Intelligence Index. Gemini 3.1 Pro Preview sits at 57.18. That is a 0.01-point gap. OpenAI effectively matched Google for the #1 position on the leaderboard, leapfrogging GPT-5.2 (51.28) and Claude Opus 4.5 (43.09) in the process. At $5.63 per million tokens, it's priced competitively against Gemini's $4.50/M while matching it on quality.

And yet the story barely registered. Partly because the "(xhigh)" suffix signals a reasoning-effort configuration rather than a clean new generation. Partly because the industry's attention was elsewhere: GTC, the Pentagon drama, agentic tooling. But the data is clear. GPT-5.4 is co-#1 by any meaningful measure.

The real winners: MiniMax-M2.7 and MiMo-V2-Pro

While GPT-5.4 took the #1 spot, two models from Chinese labs quietly delivered what matters more for most builders. MiniMax-M2.7 landed March 18 with Intelligence Index 49.62 at just $0.53 per million tokens. MiniMax has been steadily climbing (M2, M2.1, now M2.7), each iteration reducing hallucination rates. At that price-to-quality ratio, it's genuinely useful for production workloads.

MiMo-V2-Pro, also March 18, from Xiaomi. Intelligence Index 49. Elo 1426 on GDPval-AA for agentic tasks. The successor to MiMo-V2-Flash pushes reasoning further while staying open-weight and priced to undercut everything in its tier.

Both models are open-weight. Both score close to 49-50 on the Intelligence Index. Both were released on the same day. Whether that's coincidence or competitive signaling, the result is the same: the 45-to-50 band, the tier that handles the majority of real production workloads, got two strong new entrants in a single afternoon.

Mixture-of-Experts ate March

Three of March's nine releases used MoE architectures. That's not new. MoE has been the default for large open models since late 2025. What's new is the efficiency ratios.

Mistral Small 4 is the one worth lingering on. 119 billion total parameters with only 6.5 billion active. That's a model with the knowledge capacity of a large model but the inference cost of a small one. It supports image and text inputs, offers hybrid reasoning (score of 27 in reasoning mode), and ships under Apache 2.0. You can run it, modify it, build on it, sell products with it. And pair it with Mistral Forge, the custom model training platform Mistral launched at GTC on March 17, and the picture becomes clearer: Mistral is selling the full stack for enterprises that want to own their AI pipeline end-to-end.

NVIDIA's Nemotron 3 Super tells a similar story: 120B total, 12B active, open weights, Intelligence Index of 36. Not frontier-class, but running at 12B active parameters means it fits on hardware that most companies already own. Read this release in the context of GTC 2026, where Jensen Huang unveiled the Vera Rubin platform, the Nemotron Coalition with Mistral and Perplexity, and an open Agent Toolkit built around Nemotron models. NVIDIA isn't just building chips anymore. It's building the model-to-hardware pipeline, and Nemotron 3 Super is the open-weight anchor of that strategy.

Grok 4.20: lowest hallucination rate ever measured

xAI's Grok 4.20 Beta, released March 12, deserves a separate section because of one number: 22% hallucination rate. That is the lowest hallucination rate Artificial Analysis has measured on any model to date.

The rest of the spec sheet is solid but not record-breaking: 82.9% on IFBench (instruction following), 265 tokens per second output speed, priced at $2 input / $6 output per million tokens. What sets it apart is the factual accuracy. For applications where making things up is catastrophic (legal, medical, financial, compliance), a 22% hallucination rate versus the 30-40% range most models sit in is a genuine differentiator.

The "Beta" tag still applies, and xAI has historically iterated quickly on Grok versions post-beta. But the hallucination number alone is worth tracking. If it holds in independent testing, Grok 4.20 becomes the default answer for factuality-sensitive deployments.

Alibaba went wide with Qwen3.5

Qwen3.5 isn't a model. It's a product line. Alibaba shipped reasoning variants at 0.8B, 2B, 4B, and 9B (dense), plus MoE variants at 27B, 35B (3B active), 122B (10B active), and 397B. Eight models in one release, each targeting a different hardware tier.

The small variants matter most. A 0.8B reasoning model that runs on a phone is qualitatively different from the cloud-first releases of 2024. The 4B variant hits the sweet spot for single-GPU consumer hardware. On the large end, the 397B MoE scored 45.05 on the Intelligence Index at $1.35/M, the 27B scored 42.07, and the 122B scored 41.6. Alibaba remains one of the most consistent open-weight contributors in the industry.

The open-weight scoreboard

Seven of nine. That ratio has held steady since December 2025.

The practical implication: if you're building a product today and want to avoid vendor lock-in, the selection of capable open models in the 35-to-50 range is now deep enough to staff an entire AI pipeline. Reasoning (MiMo-V2-Pro at 49), general tasks (MiniMax-M2.7 at 49.62), efficient inference (Mistral Small 4), and edge deployment (Qwen3.5 small variants) are all covered without a single proprietary API call.

How March shifted the landscape

At the end of February, Gemini 3.1 Pro Preview sat alone at the top (57.18). GPT-5.2 (51.28) and Claude Opus 4.6 (52.95) held the upper tier. Below that, the 45-50 band was thin. March changed every layer.

Practical guidance for March 2026

GPT-5.4 now matches Gemini 3.1 Pro Preview at the top, giving you two co-#1 options. But the biggest value shifts happened in the tiers below. If you're evaluating cost:

What to watch next

The ceiling (57.18) has held since February, even though GPT-5.4 now shares it. Google, OpenAI, and Anthropic are all expected to ship significant updates in Q2 2026, and Morgan Stanley estimates roughly 10x more training compute coming online in H1. When one lab breaks above 57, the others will respond fast.

But model intelligence is no longer the only axis. Sub-10B reasoning from Qwen3.5 and 6.5B-active Mistral Small 4 signal that local-first AI is a category now, not a compromise. NVIDIA's GTC bet on physical AI says the next demand wave won't just be chatbots. And the Anthropic-Pentagon standoff raised alignment questions the industry can't ignore. "Which model is best?" is giving way to "How do we deploy this at scale, on our hardware, without breaking the power grid?"

Re-evaluate monthly. The leaderboard is stable. Everything around it is not.

Data sourced from Artificial Analysis, developer announcements, and WhatLLM.org tracking. See our interactive model explorer for live pricing, speed, and benchmark data across 280+ models.