GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
GPT-56T 861 —
MUSE-SPK 835 -0.7%
GPT-56SC 828 -5.2%
QWEN-38X 824 —
CL-OP55X 822 —
GROK-46H 822 -5%
GPT-6A 820 —
GLM-5 784 -8.4%
CL-FAB5H 743 -5.6%
KIMI-K3X 742 -8.4%
CL-OP5H 720 -5.8%
CL-OP5X 709 -18%
CL-OP46H 698 -5.9%
CL-OP47H 690 -5.9%
GEM-38FH 677 +0.1%
GEM-37FH 657 -24%
GPT-56S 622 —
CL-OP47 582 -0.7%
GPT-55H 582 —
INKL 531 —
GEM-31P 513 —
GEM-3P 499 —
CL-OP46 496 -0.2%
CL-OP48 490 —
← Back to feed

Mercury 2 Leads Artificial Analysis Output Speed Rankings — the Inference Speed Tier Now Belongs to Alternative Architectures

Artificial Analysis’s current output speed rankings put Mercury 2 at the top of the leaderboard, followed by Liquid AI’s LFM2 1.2B, IBM’s Granite 4.0 H Small, and Liquid AI’s LFM2.5-1.2B-Instruct. None of these models appear in the intelligence index top tier. The two rankings are measuring different things, and the architecture choices that optimize for each have diverged entirely.

What Drives the Speed Tier

Mercury 2 is built on diffusion-based generation. Instead of decoding tokens sequentially — left to right, one at a time — diffusion language models generate output in parallel passes, refining a full sequence iteratively. That eliminates the per-token sequential bottleneck that limits autoregressive throughput regardless of hardware.

The Liquid AI models at positions two and four use structured selective kernels — liquid neural networks — that process sequences without the full attention matrix computation that autoregressive transformers require. At 1.2B parameters and no attention overhead, they run faster than much larger standard models.

IBM’s Granite 4.0 H Small is a different strategy: a conventionally structured smaller model tuned specifically for high-throughput deployment on constrained hardware, not a novel architecture. It earns its position through engineering discipline on a standard design.

The Capability Gap Is Structural

The intelligence index leaders — Claude Fable 5, Claude Opus 4.8, GPT-5.5 (xhigh), GPT-5.5 (high) — are not in the output speed tier. They are not even close. These are 100B-plus parameter models running with extended thinking modes and full attention over long contexts. That depth is what produces frontier reasoning capability. It is also what makes them slow.

The gap is not closing. As frontier intelligence models grow in parameter count and expand thinking capacity, the most capable models are moving further from the inference speed tier, not converging with it. Opus 4.8 Thinking and GPT-5.5 xhigh are genuinely slower than their predecessors in raw tokens-per-second — their capability gains came from doing more compute per token, not less.

The Deployment Implication

Speed leaders are built for workloads where latency and throughput dominate: streaming UI responses, voice pipelines, real-time tool-calling loops, customer service agents with hard SLA requirements. Intelligence leaders are built for workloads where being right matters more than being fast.

A Mercury 2 deployment and a Claude Fable 5 deployment are not in competition with each other. They are solutions to different problems. The bifurcation is not a failure of frontier scaling — it is the field correctly specializing.

What changes is the buyer decision. Choosing a model used to primarily mean choosing a capability tier. It now means choosing a capability tier and an architecture tier, and accepting that optimizing for one makes the other harder to achieve.

Mercury 2 topping the AA speed chart while Fable 5 tops the AA intelligence chart is the same benchmark confirming both halves of that split.