---
title: "How Does Sage-A1.5 Compare to Kimi-K3, GLM-5.3 and GPT-5.6-sol?"
url: "https://goremy.ai/blog/articles/sage-a1-5-benchmarks-vs-kimi-k3-glm-5-3"
author: Luis Chavez-Mattos
editor: Luis Chavez-Mattos
editor_title: Director, Product
published: "2026-10-23T10:00:08.090Z"
updated: "2026-10-23T10:00:08.090Z"
topic: AI Model Reviews & Comparisons
description: Sage-A1.5, a 35B MoE agentic model, hits 95.1 on GAIA and 61.2 on SWE-bench Pro. Here is where it beats or trails Kimi-K3, GLM-5.3 and GPT-5.6-sol.
---

# How Does Sage-A1.5 Compare to Kimi-K3, GLM-5.3 and GPT-5.6-sol?

> Sage-A1.5, a 35B MoE agentic model, hits 95.1 on GAIA and 61.2 on SWE-bench Pro. Here is where it beats or trails Kimi-K3, GLM-5.3 and GPT-5.6-sol.

Edited by Luis Chavez-Mattos, Director, Product · Published 2026-10-23

## How does Sage-A1.5 perform against larger frontier models?

Sage-A1.5 is a 35B Mixture-of-Experts agentic model from Shanghai AI Laboratory. On the lab's own benchmark table it beats Kimi-K3, GLM-5.3 and GPT-5.6-sol on several tasks, including GAIA (95.1), XBench-DS-2510 (92.0), IFBench (81.6) and FrontierScience-Research (48.3). It trails them on most coding, deep-search and chemistry tests, usually by a few points to about fifteen. Its predecessor, Agents-A1, still wins a few categories.

## TL;DR

- Sage-A1.5 posts **95.1 on GAIA**, ahead of Kimi-K3 (92.2), GLM-5.3 (92.2) and GPT-5.6-sol (88.3), though its predecessor Agents-A1 scores 96.0.
- It takes **overall first place on seven benchmarks**: XBench-DS-2510, IFBench, PinchBench, FrontierScience-Olympiad, FrontierScience-Research, CMT and OmniDocBench.
- On engineering, **SWE-bench Pro lands at 61.2**, within about two to three points of the three larger models (63.3 to 64.6).
- The weak spots are **BrowseComp against Kimi-K3 and GPT-5.6-sol**, SciCode, ChemCOTBench and Agents' Last Exam, where larger models lead clearly.
- Versus A1, the jump is large: **+17.6 on SWE-bench Pro and +21.3 on MLE-Lite**.
- All numbers come from the developers' own model card, so treat them as vendor-reported until independent runs appear.

## What is Sage-A1.5 and how is it built?

According to the model card, Sage-A1.5 is based on Qwen3.6 (the linked base is Qwen3.6-35B-A3B) and on Agents-A1, Shanghai AI Laboratory's earlier agentic model. The lab frames it around three capabilities:

- **Agentic orchestration.** A lead agent delegates work to subagents that run in parallel with isolated contexts.
- **Verification mode.** An opt-in self-review step inside the model's own chain of thought. The model pauses, checks what has worked and what has failed, and changes direction in the same turn, with no external critic.
- **Open scientific discovery.** Exploratory research that produces inspectable artifacts, rather than only answering fixed questions.

The weights are on Hugging Face and ModelScope in Transformers format, and the card lists compatibility with vLLM and SGLang. Native context is 256K, extendable to 512K with YaRN (factor 2.0).

## Where does Sage-A1.5 lead the larger models?

The model card compares Sage-A1.5 with two groups. The "comparable" group is Agents-A1 and Ornith-1.5. The "larger-scale" group is Kimi-K3, GLM-5.3 and GPT-5.6-sol.

| Benchmark | Sage-A1.5 | Kimi-K3 | GLM-5.3 | GPT-5.6-sol |
|---|---|---|---|---|
| GAIA | 95.1 | 92.2 | 92.2 | 88.3 |
| XBench-DS-2510 | 92.0 | 85.0 | 84.0 | 90.0 |
| IFBench | 81.6 | 72.1 | 53.4 | 72.7 |
| PinchBench | 90.0 | 88.2 | 86.6 | 87.3 |
| FrontierScience-Olympiad | 81.0 | 69.0 | 67.0 | 73.0 |
| FrontierScience-Research | 48.3 | 21.7 | 16.7 | 26.7 |
| CMT | 80.0 | 34.0 | 40.0 | 50.0 |
| OmniDocBench | 91.5 | 91.1 | not reported | 85.8 |

The scientific research rows stand out. On FrontierScience-Research, Sage-A1.5 nearly doubles GPT-5.6-sol's score. On CMT it leads the next-best larger model by 30 points. Instruction following is also strong, with IFBench showing a roughly nine-point lead over GPT-5.6-sol.

Two caveats. GLM-5.3 is text-only, so it has no OmniDocBench score. And Agents-A1 actually posts 96.0 on GAIA, so Sage-A1.5's GAIA result is the best among the larger models but not the best in the table.

## Where does Sage-A1.5 trail?

The gaps show up in coding depth, hard deep-search and domain knowledge.

| Benchmark | Sage-A1.5 | Kimi-K3 | GLM-5.3 | GPT-5.6-sol |
|---|---|---|---|---|
| BrowseComp | 81.7 (82.8 with orchestration) | 91.2 | 76.2 | 90.4 |
| DeepResearch-Bench2 | 43.7 (47.5 with orchestration) | 52.4 | 55.7 | 57.9 |
| SWE-bench Pro | 61.2 | 63.3 | 64.6 | 64.6 |
| SWE-bench Multilingual | 72.7 | 80.8 | 81.3 | 74.1 |
| MLE-Lite | 65.2 | 72.7 | 72.7 | 77.3 |
| SciCode | 44.3 | 58.7 | 59.0 | 56.1 |
| HLE w/ tools | 55.1 | 56.0 | 62.5 | 64.5 |
| ChemCOTBench | 66.5 | 82.5 | 78.3 | 77.9 |
| Agents' Last Exam (ALE-CLI) | 19.1 | 27.6 | 28.5 | 28.6 |
| VitaBench | 45.4 | 59.0 | 58.7 | 46.8 |

The biggest misses are ChemCOTBench (16 points behind Kimi-K3), SciCode (about 14 points behind), VitaBench against Kimi-K3 and GLM-5.3 (over 13 points) and BrowseComp against Kimi-K3 (9.5 points). Some rows are close. SWE-bench Pro is within 2.1 to 3.4 points of the larger models, and HLE with tools is within one point of Kimi-K3.

Sage-A1.5 does beat individual larger models on a few rows where it doesn't take first place. It tops GLM-5.3 on BrowseComp and AA-LCR (81.0 vs 79.7) and edges GLM-5.3 on CritPt (20.0 vs 19.1), but it trails GPT-5.6-sol on VitaBench by 1.4 points (45.4 vs 46.8).

## How does it compare with other ~35B models?

Among the "comparable" group, the card marks Sage-A1.5 best on BrowseComp, SWE-bench Pro, SWE-bench Multilingual, MLE-Lite, AA-LCR, HLE with tools, CritPt, ALE-CLI, VitaBench and ChemCOTBench. It ties Agents-A1 on SciCode at 44.3.

The exceptions are worth knowing:

- **GAIA:** Agents-A1 is ahead, 96.0 to 95.1.
- **HiPhO:** Agents-A1 is ahead, 46.4 to 45.3.
- **MatTools:** Agents-A1 is ahead, 47.1 to 42.6.
- **DeepResearch-Bench2:** Ornith-1.5 scores 48.6, above Sage-A1.5's 43.7 without orchestration (47.5 with it).

Ornith-1.5 is the closest rival on engineering. It scores 59.6 on SWE-bench Pro against Sage-A1.5's 61.2, and 71.4 against 72.7 on SWE-bench Multilingual. On most other rows it sits well behind.

## How much better is Sage-A1.5 than Agents-A1?

The engineering and agentic gains are the headline improvement:

| Benchmark | Agents-A1 | Sage-A1.5 | Change |
|---|---|---|---|
| SWE-bench Pro | 43.6 | 61.2 | +17.6 |
| SWE-bench Multilingual | 58.7 | 72.7 | +14.0 |
| PinchBench | 76.9 | 90.0 | +13.1 |
| MLE-Lite | 43.9 | 65.2 | +21.3 |

Other rows moved less, and a few went slightly the wrong way (GAIA, HiPhO, MatTools). The pattern suggests the new release trades a little peak search and materials performance for much better coding and tool-driven workflows.

## Does agentic orchestration actually help?

The card reports two numbers on the long-horizon search tests where it applies. BrowseComp rises from 81.7 to 82.8 with orchestration. DeepResearch-Bench2 rises from 43.7 to 47.5. The second gain is larger, which fits the idea that parallel subagents help most on broad report-style tasks rather than single-answer lookups. Even with orchestration, DeepResearch-Bench2 stays below Kimi-K3, GLM-5.3 and GPT-5.6-sol.

## Can you trust these benchmark numbers?

Treat them as a strong first signal, not a final verdict. A few points matter:

- **Self-reported.** The table comes from the model's developers. Independent reproductions were not part of the source material.
- **Harness details.** For the SWE-bench series, the card says it used the mini-SWE-agent harness with temperature 1.0, top-p 0.95, a 200K-token context and automatic context compaction, in a fully offline environment. Different harnesses can shift scores noticeably.
- **Small test sets.** Several scores are round numbers (80.0, 90.0, 86.0), which hints at benchmarks with few items, where one or two questions move the result.
- **Narrow comparison set.** Only six models appear in the table.

For anyone planning to build on it, the useful test is a private eval on your own tasks, especially if your workload leans on coding depth or chemistry, where the gaps are widest.

## Frequently Asked Questions

### Is Sage-A1.5 better than GPT-5.6-sol?
On the developers' table, it wins on GAIA, XBench-DS-2510, IFBench, PinchBench, FrontierScience-Olympiad, FrontierScience-Research, CMT and OmniDocBench. GPT-5.6-sol wins on BrowseComp, MLE-Lite, SWE-bench Pro, HLE with tools, CritPt and others. Neither model is better across the board.

### What is Sage-A1.5's best benchmark result?
Its largest margins are on CMT (80.0 against 50.0 for GPT-5.6-sol) and FrontierScience-Research (48.3 against 26.7). In absolute terms, GAIA at 95.1 is the highest score it posts, though Agents-A1 reaches 96.0 there.

### Can I run Sage-A1.5 locally?
The weights are published in Hugging Face Transformers format, and the card gives launch commands for SGLang and vLLM, including a multi-token prediction option in SGLang. It supports a 256K context natively and up to 512K with YaRN. The card excerpt does not state hardware requirements, so check the full repository before sizing a GPU.

### What does Verification mode do?
It is an opt-in self-review step inside the model's chain of thought. The model pauses mid-task to review what it has established and what failed, then changes course in the same turn. No separate critic model is involved.