Skip to main content

Benchmark profile

InferenceBench

A benchmark for open-ended LLM inference optimization by AI agents. Agents receive a base model, one H100, and a fixed time budget to build a valid OpenAI-compatible inference server that improves serving speed.

Data verified

How BenchLM shows InferenceBench

BenchLM mirrors the public InferenceBench agent leaderboard captured on May 20, 2026. The source evaluates 18 frontier agent configurations over 4 inference-serving scenarios with a 2 h budget on 1 NVIDIA H100 80 GB.

InferenceBench is display only on BenchLM. It is a useful agentic systems-engineering signal, but the rows combine model capability, agent scaffold, inference framework choices, hardware, and final-server validity, so BenchLM keeps it separate from weighted model-only rankings.

18 agent rows4 scenarios180 runs2 hDisplay only

Aggregate speedup on InferenceBench — May 20, 2026

BenchLM mirrors the published aggregate speedup view for InferenceBench. Claude Opus 4.8 (xHigh) leads the public snapshot at 7.34x , followed by GPT-5.4 (High) (6.16x) and Claude Sonnet 4.6 (5.56x). BenchLM does not use these results to rank models overall.

18 modelsAgenticCurrentDisplay onlyUpdated May 20, 2026

Aggregate speedup table (18 models)

Score
1
Claude Opus 4.8 (xHigh)Anthropic · Closed
7.34x
2
GPT-5.4 (High)OpenAI · Closed
6.16x
3
Claude Sonnet 4.6Anthropic · Closed
5.56x
4
GPT-5.3 Codex (High)OpenAI · Closed
5.49x
5
GPT-5.5 (xHigh)OpenAI · Closed
5.45x
6
Gemini 3.1 ProGoogle · Closed
4.92x
7
Kimi K2.6Moonshot AI · Open weight
4.51x
8
Claude Opus 4.6Anthropic · Closed
4.38x
9
GPT-5.2OpenAI · Closed
4.28x
10
GPT-5.5 (High)OpenAI · Closed
4.22x
11
Gemini 3.5 FlashGoogle · Closed
4.16x
12
Claude Opus 4.5Anthropic · Closed
3.76x
13
GPT-5.1 Codex MaxOpenAI · Closed
3.59x
14
GLM-5Z.AI · Open weight
3.22x
15
Claude Sonnet 4.5Anthropic · Closed
3.18x
16
Claude Haiku 4.5Anthropic · Closed
2.78x
17
GPT-5.3 Codex (Medium)OpenAI · Closed
2.32x
18
GPT-5.2 CodexOpenAI · Closed
1.98x

The published InferenceBench snapshot places Claude Opus 4.8 (xHigh) first at 7.34x. The third row is 1.78 points behind. The broader top-10 range is 3.12 points, so many of the published results sit in a relatively narrow band.

18 models have been evaluated on InferenceBench. The benchmark falls in the Agentic category. This category carries a 22% weight in BenchLM.ai's overall scoring system. InferenceBench is currently displayed for reference but excluded from the scoring formula, so it does not directly affect overall rankings.

About InferenceBench

Year

2026

Tasks

4 inference-serving optimization scenarios

Format

Two-hour autonomous CLI agent run

Difficulty

Open-ended ML systems engineering

BenchLM mirrors the public InferenceBench agent leaderboard as a display-only agentic systems-engineering benchmark. The primary score is aggregate geometric-mean speedup over a PyTorch baseline across prefill latency, decode latency, throughput, and all-in-one serving scenarios.

BenchLM freshness & provenance

Version

InferenceBench 2026

Refresh cadence

Quarterly

Staleness state

Current

Question availability

Public benchmark harness and aggregate leaderboard

CurrentDisplay only

BenchLM uses freshness metadata to decide whether a benchmark should still be treated as a strong differentiator, a benchmark to watch, or a display-only reference. For the full scoring policy, see the BenchLM methodology page.

FAQ

What does InferenceBench measure?

A benchmark for open-ended LLM inference optimization by AI agents. Agents receive a base model, one H100, and a fixed time budget to build a valid OpenAI-compatible inference server that improves serving speed.

Which model leads the published InferenceBench snapshot?

Claude Opus 4.8 (xHigh) currently leads the published InferenceBench snapshot with 7.34x aggregate speedup. BenchLM shows this benchmark for display only and does not use it in overall rankings.

How many models are evaluated on InferenceBench?

18 AI models are included in BenchLM's mirrored InferenceBench snapshot, based on the public leaderboard captured on May 20, 2026.

Last updated: May 20, 2026 · mirrored from the public benchmark leaderboard

Choose a model with this week’s evidence

Join 2,000+ readers for ranking moves, pricing changes, and the claims that still need proof.

One email each week. Unsubscribe anytime.