Skip to content
Tiki AI
Benchmarks

Benchmarks & Leaderboards

Benchmark results from the Tiki AI team. When a weakness we fix for a client is common across the industry, the eval behind it hardens into a public benchmark.

Latest Publication

Dataset & BenchmarkOctober 2026

Image-to-Code: A Dataset & Benchmark for Full-Stack Web Reconstruction

50 projects that ask a model to rebuild a multi-page web application from its prototypes, with a working backend, data and sign-in. Scored on visual fidelity, workflows run by a GUI agent in a real browser, and whether the project starts at all.

Image-to-CodeFull-StackAgentsMultimodal
LeaderboardMean overall score
  1. 01GPT-6 Astra54.15
  2. 02GPT-6 Sol46.29
  3. 03DeepSeek V4.1 Flash42.26
  4. 04Gemini 3.8 Flash35.68
  5. 05Qwen3.8 Max35.11
View full results
BenchmarkSeptember 2026

Assistant Reliability Bench

An evaluation of persistent, tool-driven agent workflows: 28 long-horizon tasks across 10 workflow categories, scored with deterministic checks, rubric assessment, and repeated-run analysis.

AgentsTool UseLong-HorizonReliability
LeaderboardTotal / 100
  1. 01Claude Fable 5.186.45
  2. 02GPT-6 Astra85.38
  3. 03Kimi K378.23
  4. 04DeepSeek V4.1 Flash76.69
  5. 05GLM-5.375.35
View full results
Dataset & BenchmarkJune 2026

Image-to-Code: A Dataset & Benchmark for Static Web Page Replication

A dataset and benchmark for turning web page screenshots into code, scored on 8 metrics validated against human review. Fine-tuning on 1,000 samples lifts the base model 16.5 points.

Image-to-CodeMultimodalDatasetFine-tuning
LeaderboardMean score
  1. 01GPT-5.480.02
  2. 02Qwen-3.6-plus72.64
  3. 03Doubao-seed-2.072.10
  4. 04Kimi-k2.670.58
View full results
What's Next

More on the way

New benchmarks will be added here as our team publishes them, each with a leaderboard and the full method behind its scores.