Gemini 3.5 Flash vs Hy Image 3.5 PreviewWhich Is Better in 2026?

A 657-task benchmark of multi-step work across simulated SaaS applications in six business domains. The Artificial Analysis protocol reports a guardrail-aware score and is distinct from both the public Zapier split and the unrelated dynamic AutoBench framework.

Guardrail-aware score

A benchmark of difficult, verifiable information-seeking questions designed to measure an agent's ability to locate hard-to-find facts through web browsing.

Accuracy

A verified computer-use benchmark in which multimodal agents operate desktop applications and are graded from the resulting environment state.

Mean task reward

A long-horizon software-engineering benchmark with 113 original tasks graded by hand-written tests.

Pass@1

A human-validated subset of real GitHub issues used to measure whether a coding agent can produce repository patches that resolve the associated tests.

Resolved

Version 2.1 of the benchmark for completing realistic tasks in terminal environments. Harness, resource limits, and attempt count are part of the protocol.

Mean task success

A dynamic LLM evaluation framework in which models generate questions, answer them, and participate in reciprocal peer assessment. AutoBench is distinct from Zapier's AutomationBench.

Weighted peer-assessment score

The highest-quality subset of Graduate-Level Google-Proof Q&A, designed to test expert-level scientific reasoning in biology, physics, and chemistry.

Accuracy

A 2,500-question expert-level benchmark spanning dozens of academic fields. Tool-assisted and no-tools results are separate protocols and must not be merged.

Accuracy

A contamination-resistant benchmark refreshed on a fixed release cadence. Scores from different LiveBench releases must never be compared as the same protocol.

Mean of category averages