Z.ai launches glm-5.3 on openrouter with post-training boost in benchmark scores

The new model went live on openrouter on 18 august. It shares the same training base as glm-5.2; all gains come from the post-training stage. According to the platform, terminal-bench 3.0 rose from 4.6 to 28.3 and deepSWE v1.1 climbed from 46.2 to 66.9. Both tests measure complex software-engineering reasoning and long-horizon agent tasks, so the jump signals a substantive behavior shift without changing the core weights.

Terminal-bench 3.0 evaluates multi-step terminal tasks in a simulated environment. DeepSWE v1.1 probes real engineering problem solving from open-source code bases. Moving from 4.6 to 28.3 and from 46.2 to 66.9 lifts the model from barely above threshold to a level near the category leaders. z.ai describes glm-5.3 as a large-scale reasoning model built specifically for complex software engineering and long-term agent tasks.

The model accepts text-only input and output and supports a context window of one million tokens. Reranking runs continuously and cannot be turned off. Three effort levels are offered—low, high and max—with max as the default. Pricing on openrouter is $1.40 per million input tokens and $4.40 per million output tokens. Reported throughput is 24 tokens per second (p50, optimal among providers) and median latency is 6.32 seconds.

z.ai notes that the improvement balances raw scores with token-efficiency, a critical point because reranking always consumes an internal token budget. The three effort levels let developers trade off: low for simple tasks that save tokens, max for hard tasks where quality outweighs cost. Using the same base as glm-5.2 demonstrates how far an existing model can be pushed with focused continued training.

Openrouter records availability of 99.85 % and 99.60 % across its metrics, with an automatic routing mechanism that switches to a healthy provider if the primary provider errors, provided request filters permit it. Per-provider availability data are accessible via the endpoints api. In practice, developers building on glm-5.3 receive relatively stable access without managing fallback between infrastructure providers.