Claude Opus 4.6 vs. GLM-5.1: The Closed-Source King Meets Its Open-Source Challenger

On April 7, 2026, something happened that would have been unthinkable a year ago. An open-source model — with weights on Hugging Face, under an MIT license, for free — topped the SWE-Bench Pro leaderboard, beating both GPT-5.4 and Claude Opus 4.6. That model is GLM-5.1, built by Z.ai (formerly Zhipu AI), a Tsinghua University spinoff that became the first publicly traded foundation model company in the world when it listed on the Hong Kong Stock Exchange in January 2026 at a market capitalization around $52.83 billion.

Claude Opus 4.6, meanwhile, remains Anthropic's flagship — the model that powers Claude Code, leads on SWE-Bench Verified, and has become the backbone of agentic coding workflows for hundreds of thousands of developers. It's the model 70–80% of Anthropic's own technical staff use daily.

So what happens when the closed-source king meets its open-source challenger? The answer is more nuanced than any single benchmark can capture.

The models at a glance

Claude Opus 4.6 launched on February 5, 2026. It's a proprietary model with a 1 million token context window (beta), 128K max output tokens, and adaptive thinking that dynamically decides when and how much to reason. It introduced Agent Teams for multi-agent orchestration, conversation compaction for infinite-length workflows, and four effort levels that let developers trade off intelligence, speed, and cost. Its pricing sits at $5 per million input tokens and $25 per million output tokens.

GLM-5.1 is a post-training upgrade to the GLM-5 base model. It runs a 754-billion parameter Mixture-of-Experts architecture with 40 billion active parameters per token, a 200,000 token context window, and up to 128,000 output tokens per response. It was trained entirely on 100,000 Huawei Ascend 910B chips — zero NVIDIA involvement — using Z.ai's custom asynchronous RL infrastructure called "slime." The API is priced at $1.40 per million input tokens and $4.40 per million output tokens. The weights are freely available under the MIT license.

The philosophical difference matters as much as the technical specs. Anthropic built Opus 4.6 to be the most capable model inside its own product ecosystem — Claude Code, Claude.ai, the API. Z.ai built GLM-5.1 to be the most capable open model that anyone can deploy anywhere, on any infrastructure, for any purpose.

Where GLM-5.1 wins

SWE-Bench Pro. GLM-5.1 scored 58.4, topping GPT-5.4 at 57.7 and Claude Opus 4.6 at 57.3. This benchmark tests a model's ability to resolve real-world GitHub issues using a 200K token context window, making it one of the closest proxies we have for actual software engineering capability. The margin is narrow — just 1.1 points over Opus — but the direction matters. An open-source model leading this benchmark was genuinely unprecedented.

Long-horizon autonomous execution. This is GLM-5.1's headline differentiator. Z.ai claims it can work continuously and autonomously on a single task for up to 8 hours, completing a full plan-execute-test-fix-optimize loop without human intervention. In their most impressive demonstration, GLM-5.1 built a complete Linux-style desktop environment from scratch — functional file browser, terminal, text editor, system monitor, and playable games — through 655 autonomous iterations. It also optimized a vector database to run at 6.9x its original throughput during the same session. Previous models, including GLM-5, tend to plateau early — they apply familiar techniques for quick gains, then run out of ideas. GLM-5.1 was specifically trained to avoid this pattern through what Z.ai calls progressive alignment: multi-task supervised fine-tuning, followed by reasoning RL, agentic RL, general RL, and on-policy cross-stage distillation.

Price. At $1.40/$4.40 per million tokens, GLM-5.1 is roughly 82% cheaper on input and 82% cheaper on output compared to Opus 4.6's $5/$25. For a team processing 10 million output tokens per month, that's the difference between roughly $44,000 and $250,000 annually. And since the weights are MIT-licensed, you can self-host and eliminate API costs entirely if you have the GPU infrastructure.

CyberGym performance. GLM-5.1 achieved a 68.7 score based on a single-run pass over 1,507 tasks, demonstrating a nearly 20-point lead over the previous GLM-5 model and strong performance on cybersecurity-related challenges.

Open-source flexibility. GLM-5.1 can be deployed locally via vLLM, SGLang, xLLM, and KTransformers. You can fine-tune it, customize it, integrate it into air-gapped environments, and audit every weight. For organizations with data sovereignty requirements, regulated industries, or simply a preference for full control over their AI stack, this is a meaningful advantage that no proprietary model can match.

Where Claude Opus 4.6 wins

SWE-Bench Verified. While GLM-5.1 leads on SWE-Bench Pro, Opus 4.6 still holds the crown on SWE-Bench Verified at 80.8% compared to GLM-5.1's 77.8%. The two benchmarks test related but different capabilities — Verified focuses on cleaner, well-scoped issues, while Pro tests messier real-world scenarios. The 3-point gap on Verified suggests Opus 4.6 still has a meaningful edge on well-defined engineering tasks.

Broader coding composite. On the broader coding composite that includes Terminal-Bench 2.0 and NL2Repo together, Claude Opus 4.6 leads at 57.5 versus GLM-5.1's 54.9. So "GLM-5.1 beats Claude" is accurate on one benchmark and incomplete as a full picture.

Context window. Opus 4.6 offers a 1 million token context window versus GLM-5.1's 200,000. On MRCR v2 — a needle-in-a-haystack retrieval test at scale — Opus 4.6 scores 76%. That 5x context advantage matters enormously for developers working with large codebases who want the entire project in working memory.

Reasoning breadth. Opus 4.6 leads across a wider range of reasoning benchmarks: 91.3% on GPQA Diamond (graduate-level science), 53.0% on Humanity's Last Exam with tools, 68.8% on ARC-AGI-2 (novel problem-solving), 90.2% on BigLaw Bench (legal reasoning), and 84.0% on BrowseComp (agentic search). GLM-5.1 scored 31.0 on Humanity's Last Exam base and 52.3 with tools — competitive but behind Opus on the hardest reasoning tasks.

Subjective output quality. This is the gap benchmarks don't capture well. Human evaluators prefer Claude's outputs by a 316 Elo point margin for subjective quality — nuance, clarity, and polish. When developers compare output quality side-by-side rather than measuring functional correctness, Claude consistently wins. GLM-5.1 is optimized for getting to a correct solution through iteration; Opus 4.6 tends to produce cleaner, more readable code on the first pass.

Ecosystem and tooling. Opus 4.6 powers Claude Code, which has the most mature terminal-based agentic coding environment available — Agent Teams, hooks, Skills, MCP integration, conversation compaction, and scheduled tasks. GLM-5.1 can be used through third-party tools like Claude Code (as a BYOK model in tools like Cline or OpenCode), but it doesn't have a comparable first-party agent platform built around it. Z.ai offers a GLM Coding Plan subscription and API access, but the ecosystem is younger and smaller.

Speed. GLM-5.1 runs at approximately 44.3 tokens per second — the slowest in its tier. This is fine for batch jobs and overnight autonomous work, but painful for real-time IDE autocomplete or interactive coding sessions where you're watching tokens stream in. Opus 4.6, while not the fastest model available, offers meaningfully better latency for interactive use cases, especially with the fast mode option.

Safety and alignment. Anthropic ran their most comprehensive safety evaluation ever for the Opus 4.6 release, including new evaluations for user wellbeing, more complex refusal tests, and interpretability-based detection methods. Opus 4.6 independently discovered over 500 previously unknown zero-day vulnerabilities in open-source code during pre-release testing. Anthropic's safety track record and the depth of their alignment work remain a real differentiator for enterprise deployments where safety assurances matter.

The honest nuance

The "GLM-5.1 beats Claude" headline is both true and misleading. It's true on SWE-Bench Pro — the single most demanding software engineering benchmark available. It's misleading if you take it to mean GLM-5.1 is the better model overall.

On Z.ai's own internal coding evaluation, GLM-5.1 scored 45.3 versus Claude Opus 4.6's 47.9. Z.ai themselves acknowledge that GLM-5.1 reaches 94.6% of Claude's coding performance on their composite metric — impressive for an open-source model, but not parity. The SWE-Bench Pro lead is narrow and specific; the broader picture is more complex.

The long-horizon autonomous execution claim is the more interesting differentiator. If GLM-5.1 can genuinely sustain coherent, productive work across 8 hours and 655+ iterations without degrading, that's a capability that matters independently of benchmark scores. Claude Opus 4.6 with Agent Teams can also orchestrate long-running parallel work, but the single-agent sustained execution story is different. The question is whether GLM-5.1's 8-hour demos translate reliably to production use cases or whether they represent best-case scenarios under controlled conditions. Z.ai acknowledges that reliable self-evaluation for tasks without numeric metrics remains unsolved, and the model can hit local optima when incremental tuning stops paying off.

What this means for developers

If you're choosing between these models, the decision depends heavily on your use case.

Choose Claude Opus 4.6 if: you need the broadest reasoning capability across diverse tasks, you want the most mature agentic coding platform (Claude Code), you work with very large codebases that benefit from 1M token context, you value subjective output quality and readability, you need enterprise-grade safety and alignment assurances, or you're doing interactive real-time coding where latency matters.

Choose GLM-5.1 if: you need to minimize API costs at scale, you require open-source weights for self-hosting or compliance reasons, you're running long-horizon batch jobs where the model works autonomously for hours, you want to avoid vendor lock-in and maintain full control over your AI infrastructure, or you're building in regulated environments that require model auditability.

Use both if: you're following the composability trend that's defining 2026. Many teams are routing tasks to whichever model handles them best — using Opus for interactive coding and complex reasoning, GLM-5.1 for high-volume batch processing and cost-sensitive pipelines, and other models for their respective strengths.

The bigger picture

GLM-5.1's release represents something larger than a benchmark rivalry. It's proof that the open-source gap to frontier closed-source models has shrunk to a single benchmark point. A year ago, the gap was measured in months of capability. Now it's measured in decimal points on specific evaluations.

This has real implications. It means the moat for proprietary model companies is shifting from raw intelligence — which open-source is rapidly matching — to ecosystem, tooling, safety, and user experience. Anthropic's advantage with Claude Code, Agent Teams, and their safety infrastructure matters more now than ever, precisely because the model capability gap is narrowing.

It also has geopolitical weight. GLM-5.1 was trained entirely on Huawei chips with zero NVIDIA involvement — a fully Chinese tech stack. As chip export restrictions tighten, Z.ai has demonstrated that frontier AI training is achievable on domestic hardware. That's a strategic milestone beyond the technical achievement.

For developers, the practical takeaway is simple: you have more high-quality options at lower costs than at any previous point. The "which model is best" question is giving way to "which model is best for this specific task at this price point." That's a healthier, more competitive market — and developers are the ones who benefit.