GLM-5.3-Flash reportedly matches Claude Opus on benchmarks at ~1/40th of the per-token price

A cheap open model tied with a flagship

A post on dev.to by Van Massey makes a claim that should concern anyone managing an LLM API budget: Zhipu AI's GLM-5.3-Flash, an MIT-licensed open-source mixture-of-experts model with 320 billion total parameters and 18 billion active ones, scores the same as Claude Opus 4.8 on the Artificial Analysis Intelligence Index while charging roughly one fortieth of Opus's official per-token rate.

According to the post, the model earned attention before it was even formally identified. It launched anonymously on OpenRouter under the name "Ox Alpha," took the top spot on the platform's leaderboard on day one, and served approximately 62 trillion tokens within six days.

The price table

The dev.to post lays out Zhipu's published rates for GLM-5.3-Flash per million tokens:

• International: $0.30 for input and $1.20 for output, with output dropping to $0.60 during a limited-time half-price promotion.
• Domestic (China): ¥0.8 for input and ¥2.8 for output.

The author places international pricing in the same low-cost band as DeepSeek V4 Flash, which lists around $0.14/$0.28 per million tokens, but with a higher intelligence score: 57 on the AA index versus 53 for DeepSeek V4 Pro. Domestically, Flash is said to cost about a tenth of the full GLM-5.3 flagship tier, or a twentieth during the promotional window. Zhipu also claims, per the post, that the total bill undercuts the adjusted DeepSeek V4-Flash.

Beyond price, the model differentiates with a context window of roughly 1.04 million tokens and native multimodal support for text, images, video and files.

The post runs a concrete example: an agent session consuming one million input tokens and one million output tokens costs $1.50 at standard rates, or $0.90 during the promotion. A workload that re-reads a large codebase and burns ten million input tokens costs $3.00. The author's point is that cheap input tokens are what make long-context agents economically viable, since accumulating and re-sending state stops dominating the bill.

Caveats the post itself raises

The dev.to analysis is notably cautious about the headline ratio, and its warnings matter as much as the numbers:

• The 1/40 figure compares against Opus 4.8's list price. Teams with negotiated enterprise or volume discounts will see a smaller real-world gap.
• Per-token price is not per-task price. If the cheaper model needs more retries, larger context re-sends or repeated tool calls to finish a job, the effective advantage shrinks. The direction is real; the magnitude depends on the workload.
• The half-price output rate is temporary, so budgets should be built on the standard $0.30/$1.20 rates.
• Identical index scores do not guarantee identical behavior on specific tasks, and the post notes that DeepSeek's ecosystem is more mature. Teams already invested in that tooling should run their own evaluations before switching on a single benchmark number.

Practical recommendations

The author suggests a measured rollout: use the model's free tier (200 requests per day) to evaluate output quality against a current model on real task mixes; route a slice of production traffic to the paid tier and compare cost per completed task rather than cost per token, watching retries and context growth; keep a higher-capability model as a reviewer for code review, financial or legal outputs where subtle errors are expensive; and, since GLM-5.3-Flash speaks the OpenAI-compatible protocol, front it with a gateway alongside Claude, GPT, DeepSeek and Kimi so models can be swapped per route without code changes.

Why it matters

If a model at this benchmark level genuinely costs a fraction of flagship rates, the default architecture of many AI products changes. Agentic workloads, which are token-hungry by nature, become viable to run far more aggressively, and expensive frontier models shift from being the workhorse to being a final quality gate. The MIT license adds a self-hosting option that closed models cannot offer. But the honest read is that per-token ratios are a starting point, not a conclusion: benchmark ties are not task-level guarantees, promotions expire, and the real measure is cost per completed task on your own workload. As the post frames it, a cheap tier like this is better treated as the new baseline to beat than as an automatic replacement.