Z.ai, formerly known as Zhipu AI, released GLM-5.1 on March 27, 2026, a coding-focused post-training upgrade to its flagship GLM-5 model that has drawn immediate market attention, a double-digit stock surge, and a flood of conflicting field reports from developers on the ground.
The model is available to all GLM Coding Plan subscribers — Lite, Pro, and Max tiers — at $10 per month. Z.ai's Global Head Li Zixuan confirmed via X that the model will be open-sourced under an MIT license on approximately April 6–7, 2026, consistent with the terms of its predecessor. "Don't panic. GLM-5.1 will be open source," Li wrote in a pre-announcement on March 20.
Benchmark Claims: Near-Opus Performance at Commodity Pricing
In coding evaluations conducted using Claude Code as the testing instrument, GLM-5.1 scored 45.3 points — a 28% improvement over GLM-5's 35.4 — placing it 2.6 points behind Claude Opus 4.6's score of 47.9, or 94.6% of Opus-level coding performance. According to the LLM Selection Guide dated March 28, GLM-5.1 ranks 5th overall among frontier models and 2nd on the agentic coding leaderboard.
The architectural foundation remains unchanged from GLM-5: a 744B total / 40B active Mixture-of-Experts configuration with a 204,800-token context window. The improvements are concentrated in post-training: better instruction adherence across multi-step tasks, self-debugging loops that run linters and iterate to completion, adaptive reasoning calibrated to task complexity, and fuller context absorption before generating output changes.
The competitive pricing is stark. Claude Opus 4.6 costs approximately $15 per million tokens on a usage basis. GLM-5.1 is available for a flat $10 per month.
Zhipu's Hong Kong-listed stock, trading under 2513.HK, surged over 11% in the days following the release.
Speed: The Deliberate Trade-Off
The performance gains carry a measurable cost. Early BridgeBench testing clocked GLM-5.1 at 44.3 tokens per second, making it the slowest frontier model benchmarked to date — approximately half the speed of GPT-5.4 and nearly six times slower than Grok 4.20. The throughput penalty appears to be an architectural choice, not an infrastructure limitation, consistent with the model's emphasis on deliberate self-correction loops.
First Hand Experience: A Portrait of Contradictions
Evaluations by our engineers at CTOL Digital Solutions reveal a model of exceptional but unstable capability. In controlled tests, GLM-5.1 passed extensive backend test suites in Go, outperformed Claude Opus 4.6 and GPT-5.4 on vector database benchmarks and delivery route optimization, and produced superior dodge mechanics and special effects in a stickman fighting game prompt. One Cursor benchmark suggested it could marginally outperform Claude Opus 4.5 on specific mobile and backend tasks.
However, beyond approximately 100,000 tokens, we observed garbled output, infinite loops, and destructive code changes. Beyond 160,000 tokens, the model is hardly unusable. The model systematically consolidates code into single files without instruction, compounding long-context failures. For a basic math query — the smallest square between 15 and 30 — one user recorded 4,500 tokens of output.
On frontend tasks, we found little to no improvement over GLM-5.0 in particle effects, lighting, or spatial modeling. In our Games Clone (cloning benchmark to replicate three popular non trivial games) evaluation, GLM-5.1 fell significantly short of Claude Opus 4.6 in world generation, shadow rendering, and water effects.
Token consumption is a structural concern. We are exhausting daily quotas within one to two hours during peak usage windows, when a 3x multiplier applies.
Z.ai's release cadence since February — GLM-5 on February 11, GLM-5-Turbo on March 15, and GLM-5.1 on March 27 — signals an aggressive development posture. Whether that pace is sustainable without degraded service remains the central question among its user base, several of whom cited fears of post-launch capability downgrades driven by compute cost pressures.
The full open-source release, scheduled for April 6–7, will constitute the next meaningful test.
not investment advice
References: https://x.com/Zai_org/status/2037490078126084514