GLM-5.2 vs Claude Opus 4.8 (2026): Open Weights vs a Superseded Flagship

Our Recommendation

The honest 2026 answer starts with a correction: Claude Opus 4.8 is no longer Anthropic's flagship. Claude Opus 5 shipped on 24 July 2026 at the identical list price — $5 per million input and $25 per million output tokens — and Anthropic reports it more than doubling Opus 4.8 on Frontier-Bench v0.1 at a lower cost per task. If you are choosing an Anthropic model today, you are choosing Opus 5; Opus 4.8 is the pinned legacy option. On measured coding, hosted Opus still wins the wide-margin tests: SWE-bench Pro 69.2% to 62.1%, Terminal-Bench 2.1 85.0% to 81.0%, and the ultra-long-horizon SWE-Marathon 26.0% to 13.0%. Where the task is bounded rather than multi-hour, the gap collapses — FrontierSWE 75.1% against 74.4%, MCP Atlas 77.8% against 77.0% — and at that point the price sheet decides. GLM-5.2 lists at $0.76 in and $2.38 out per million tokens: 6.6x cheaper on input, 10.5x cheaper on output, with a slightly larger 1,048,576-token window. The strongest new argument for GLM-5.2 is not price at all. In July 2026 Hugging Face published the technical timeline of the agent intrusion on its own infrastructure and wrote that Claude Opus and Fable "refused a large part" of the forensic work — guardrails tripped every time the team tried to analyse the attack logs. Hugging Face stood up NVIDIA's quantized nvidia/GLM-5.2-NVFP4 on its own hardware, rerouted the entire pipeline through it, recovered the attacker's chunk+XOR+gzip scheme and turned up roughly four times the secrets its first automated scan had found. That is a capability argument no benchmark table shows: an open-weight model you host yourself has no refusal policy you did not write, and the evidence never leaves your network. A fourth development has changed the GLM side of this comparison since mid-August 2026: Z.ai has made GLM-5.3 available through its API. Artificial Analysis pegs it at 60 on its Intelligence Index — tied with Kimi K3 for the top spot among open models, seven points ahead of GLM-5.2 — with its biggest gains in agentic work: on GDPval-AA v2 its Elo jumps from 1,524 to 1,770 (+246), second behind Claude Opus 5 (1,855). The price point stays GLM-like: $0.68 per task, 1.5x GLM-5.2 but 19% cheaper than Kimi K3. What has not shipped is the open-weight release: Z.ai is delaying the GLM-5.3 open drop by roughly two weeks because, per the company, the model is so effective at finding security vulnerabilities that it is first tightening controls and restricting full access to select security partners under a tiered access program. The practical consequence: for API-based workflows the GLM side of this comparison is now effectively 5.3; for anyone whose decision hinges on MIT weights they can host, GLM-5.2 remains the newest open checkpoint until the 5.3 weights land — and the security-gating move is a template other open-weight labs are likely to copy. Choose Claude Opus — Opus 5, not 4.8 — for repository-wide refactoring, multi-hour autonomous runs, and regulated work where a managed Western API and its compliance paperwork are the point. Choose GLM-5.2 when volume economics, MIT weights, air-gapped deployment, or adversarial and incident-response work put you on the wrong side of a hosted model's guardrails.