Z.AI’s GLM-5.1 scores 58.4 on SWE-Bench Pro — ahead of both Claude Opus 4.6 (57.3) and GPT-5.4 (57.7) on real-world software engineering tasks. It’s the direct successor to GLM-5, designed for agentic engineering: long-horizon coding tasks, terminal operations, and repository-level work. The core design premise is that previous models, including GLM-5, tend to plateau after their initial gains — GLM-5.1 is built to keep improving across hundreds of rounds and thousands of tool calls.

What makes that architectural choice meaningful in practice is the model’s capacity for iterative strategy revision: breaking down ambiguous problems, running experiments, reading results, and identifying blockers rather than burning through a fixed repertoire early. It carries a 202,752-token context window, supports function calling and JSON natively, and ships under an MIT license — a meaningful detail for teams thinking about deployment flexibility. At $1.05 per million input tokens and $3.50 per million output tokens, it sits at a competitive price point relative to the frontier models it benchmarks against. It’s now available on DeepInfra.

What Makes This Model Different

GLM-5.1 is Z.AI’s successor to GLM-5, built around a specific thesis: most models hit a performance ceiling on long-running agentic tasks and then stall. GLM-5.1 is explicitly designed to keep improving as it’s given more time — sustaining performance across hundreds of rounds and thousands of tool calls rather than exhausting its strategy early.

The clearest evidence shows up in coding and terminal benchmarks, where GLM-5.1 pulls ahead of its predecessor by meaningful margins:

On SWE-Bench Pro and NL2Repo, GLM-5.1 lands ahead of both Claude Opus 4.6 and GPT-5.4. CyberGym sees the most dramatic jump: from 48.3 to 68.7, beating Claude Opus 4.6’s 66.6. GLM-5.1 is also available on NVIDIA’s build platform, which gives you another access path if you’re already working within that ecosystem.

On general reasoning, the gains are more modest. GPQA-Diamond moves from 86.0 to 86.2, math benchmarks are roughly flat or slightly down (HMMT Nov: 96.9 → 94.0), and HLE with tools goes from 50.4 to 52.3. The model is tuned for agentic work, not pure reasoning competitions. GLM-5.1 also scores 79.3 on BrowseComp with context management enabled, ahead of DeepSeek-V3.2 (51.4) and competitive with other top-tier models.

The model supports a 202,752-token context window with JSON and function calling — both required for real tool-use pipelines. It handles English and Chinese, is MIT-licensed, and is served in fp4 quantization on DeepInfra under zai-org/GLM-5.1. If you want to understand the broader GLM model lineage, the GLM-4.5 blog post covers the foundation model that preceded this generation.

Getting Started on DeepInfra

GLM-5.1 is available now on DeepInfra under the identifier zai-org/GLM-5.1 as a public endpoint. Pricing is usage-based: $1.05 per 1M input tokens, $3.50 per 1M output tokens, and $0.205 per 1M cached tokens. Private endpoint deployment is also supported if you need dedicated capacity — configure that directly from the DeepInfra dashboard.

DeepInfra gives you access to GLM-5.1 through an OpenAI-compatible API with zero infrastructure setup. DeepInfra operates with a zero-retention policy and is SOC 2 and ISO 27001 certified. If you’re planning to use GLM-5.1 for production coding workflows — Claude Code, Kilo Code, Cline, or similar tools — the GLM Coding Plan is worth reviewing for team-level access options.

To make your first call, grab your API key from the Dashboard and swap in the model identifier:

curl "https://api.deepinfra.com/v1/openai/chat/completions" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $DEEPINFRA_TOKEN" \
  -d '{
      "model": "zai-org/GLM-5.1",
      "messages": [
        {
          "role": "user",
          "content": "Hello!"
        }
      ]
    }'
copy

from openai import OpenAI


client = OpenAI(
    api_key="$DEEPINFRA_TOKEN",
    base_url="https://api.deepinfra.com/v1/openai",
)


response = client.chat.completions.create(
    model="zai-org/GLM-5.1",
    messages=[{"role": "user", "content": "Hello!"}],
)
print(response.choices[0].message.content)
copy

import OpenAI from "openai";


const openai = new OpenAI({
  apiKey: "$DEEPINFRA_TOKEN",
  baseURL: "https://api.deepinfra.com/v1/openai",
});


const response = await openai.chat.completions.create({
  model: "zai-org/GLM-5.1",
  messages: [{ role: "user", content: "Hello!" }],
});
console.log(response.choices[0].message.content);
copy

The only things that change from a standard OpenAI call are the base URL (https://api.deepinfra.com/v1/openai), your DeepInfra token, and the model name — the official OpenAI Python and Node.js SDKs work without any modifications. Head to deepinfra.com/zai-org/GLM-5.1 to start building.

Conclusion

GLM-5.1 makes a credible case for itself in the scenarios where agentic models tend to break down — long-running tasks, messy repositories, and multi-step terminal workflows that demand sustained reasoning rather than a single flash of capability. The benchmark numbers against Claude Opus 4.6 and GPT-5.4 aren’t cherry-picked narrow wins; they reflect a model that was deliberately tuned for the kind of work developers actually need to automate.

That opens up real engineering applications: autonomous PR triage pipelines, self-directed debugging agents, or repo-scale refactoring tools that don’t fall apart midway through. If any of that maps to what you’re building, GLM-5.1 is worth running through your eval pipeline. It’s also worth keeping in mind that “agentic model” here means something specific — not just a model with tool access, but one designed around the generalized linear structure of iterative, multi-step problem solving that real engineering tasks actually demand. Head to deepinfra.com/zai-org/GLM-5.1 to get started.