Databricks has announced the implementation of Zhipu AI’s open-source GLM 5.2 as the default coding engine across its engineering division.
This decision comes after extensive internal assessments revealed that the Chinese-developed model closely rivals the quality of Anthropic’s Opus 4.8 while offering a more economical price point at $1.28 per task—34% less than Opus’s $1.94 [1].
A blog post dated July 8, authored by a team led by co-founder and CTO Matei Zaharia, highlights this significant shift, marking one of the most notable transitions from a U.S.-based frontier model provider to a Chinese alternative [2].
In conducting its evaluation, Databricks constructed a proprietary benchmark derived from actual pull requests within its own extensive multi-million-line codebase, which includes programming languages such as Python, Go, TypeScript, and Scala.
This approach avoided reliance on public evaluation suites like SWE-Bench [2]. To ensure the integrity of the testing, the company truncated Git history to prevent models from surfacing existing solutions.
In this assessment, GLM 5.2, Opus 4.8, and OpenAI’s GPT 5.5 all achieved high performance, with pass rates ranging from 82% to 90%. However, GLM’s significantly lower token pricing rendered it the victor in terms of cost-effectiveness [1].
This strategic pivot positions Databricks alongside other prominent U.S. tech firms—such as Coinbase, Lindy, and Snowflake—that have recently transitioned their production workloads to Chinese open-weight models.
This trend is intensifying pricing pressures on industry giants like Anthropic and OpenAI within the enterprise sector [4].
The Benchmark
In its evaluation process, Databricks scrutinized eight distinct models against genuine coding tasks extracted from its production codebase. These tasks were classified by complexity: 61% categorized as medium, 19% low, and 12% high [1].
The models were assessed on their capability to generate functional code modifications that accurately reflected the intentions of actual engineer-authored pull requests.
The top performance echelon—comprising GLM 5.2, Opus 4.8, and GPT 5.5—garnered pass rates oscillating between 82% and 90%.
A middle tier, including Sonnet 4.6, Sonnet 5, and GPT 5.4, achieved scores between 71% and 82%. Meanwhile, GPT 5.4-mini and Haiku 4.5 lagged behind, ranging from 51% to 60% [1].
The economic analysis unveiled that the raw token pricing does not provide a straightforward reflection of task-level economics.
Variations in token efficiency per model and software environment necessitate that Databricks’ per-task cost metric—$1.28 for GLM 5.2 in contrast to $1.94 for Opus 4.8—serves as a more pragmatic evaluation for enterprise stakeholders [1].
The Model: Zhipu AI’s GLM 5.2
GLM 5.2 is a robust model comprising 753 billion parameters, employing a mixture-of-experts architecture with 40 billion active parameters per token. Released on June 13 by Zhipu AI, a Beijing-based entity, this model operates under an MIT license [3].
It boasts a one-million-token context window—approximately five times larger than its predecessor, GLM 5.1—and can generate outputs up to 131,072 tokens [3].
Pre-training was conducted on 28.5 trillion tokens, employing a novel asynchronous reinforcement-learning mechanism dubbed ‘slime’ to enhance learning from prolonged, multi-tool interactions [3].
An innovative feature known as IndexShare, which facilitates the sharing of a lightweight indexer across every four sparse-attention layers, yields a substantial reduction in per-token compute by approximately 2.9x at the full million-token context length [3].
On the API front, GLM 5.2 operates at around $1.40 per million input tokens and $4.40 per million output tokens.
In contrast, Opus 4.8 charges approximately $5 for input and $25 for output, making it 3.6 times more cost-effective for input and 5.7 times cheaper for output, prior to considering task-level efficiency variations [5].
A Broader Enterprise Migration
Databricks is not an outlier in this trend. Brian Armstrong, CEO of Coinbase, recently revealed that his company halved its AI expenditure by transitioning engineering tasks to GLM 5.2 and Moonshot AI’s Kimi K2.7 Code [4].
Meanwhile, Lindy CEO Flo Crivello shifted his startup’s entire API traffic from Anthropic’s Claude to DeepSeek v4 earlier in June, citing AI expenses that had previously surpassed payroll [6].
Snowflake has also assessed GLM 5.2, finding it to be competitively priced against Opus for routine workloads, albeit not on par with Claude for every task type [4].
Currently, Chinese-produced AI models account for 30% to 46% of enterprise API token traffic traversing U.S. developer platforms, a significant increase from just 4.5% in early 2025 [7].
The open-weight characteristic of these models—enabling companies to download weights, self-host, and modify them while incurring costs solely for compute—affords enterprises enhanced control over data residency and infrastructure expenditures, thus alleviating concerns about transmitting sensitive code through Chinese-hosted APIs [3].
Why It Matters
The decision by Databricks illuminates a critical insight derived from its benchmarking studies: no sole AI provider reigns supreme across all coding endeavors.
The company further advocates that enterprises should cultivate proprietary benchmark assessments tailored specifically to their unique codebases, rather than depend on public evaluation methodologies [2].
For Anthropic, the ramifications are pronounced. While Opus 4.8 remains within the upper performance tier, its costs—estimated at four to six times greater per token compared to GLM 5.2—are increasingly difficult to substantiate, particularly as open-weight alternatives demonstrate equivalent efficacy on real-world coding tasks [1][5].
This development echoes an earlier trend witnessed in the commoditization of cloud infrastructure providers, where enterprise clients gravitated towards open-source solutions as quality discrepancies diminished.
Valued at $62 billion during its latest funding round, Databricks serves over 10,000 enterprise clients and operates one of the most extensive data and AI platforms worldwide.
Its endorsement of GLM 5.2 as a viable production-grade coding framework carries considerable weight in enterprise AI procurement decisions [2].
What’s Next
Databricks has announced its commitment to enhancing the inference performance of GLM 5.2 on its platform, with Zaharia also emphasizing efforts to facilitate model deployment at scale [8].
The company asserted that its benchmark will remain an evolving evaluation, with model rankings subject to change as providers release new updates.
In response, U.S.-based AI research labs are likely to engage in aggressive pricing strategies. Anthropic and OpenAI have already implemented price reductions multiple times in 2025 and 2026; however, the disparity with open-weight Chinese models endures.
The ability of U.S. firms to distinguish themselves based on reliability, safety, or tailored enterprise features may ultimately dictate the extent of market share retention in this shifting landscape.
Source link: Mlq.ai.
