Databricks benchmarked the Chinese open-source AI model GLM 5.2 against Anthropic's Opus 4.8 on real coding tasks from its own codebase and found them statistically tied in performance, with GLM 5.2 costing $1.28 per task versus $1.94 for Opus. The company is now deploying GLM 5.2 as its primary coding engine, joining Coinbase, Lindy, and Snowflake in a broader shift toward cheaper Chinese models that cost 60 to 90 percent less than Western alternatives while maintaining comparable quality.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Databricks tested the Chinese open-source model GLM 5.2 on its own codebase and found it tied with Anthropic's Opus 4.8 in performance while costing $1.28 per task versus $1.94 for Opus. The company now plans to make GLM 5.2 a daily workhorse for developers. Other companies—Coinbase, Lindy, and Snowflake—have similarly switched to cheaper Chinese models including GLM 5.2 and Deepseek v4.
Why it matters
Cost pressures are shifting development workflows away from Western AI providers. On OpenRouter, Chinese models topped 30 percent of weekly traffic since February 2026, up from 11 percent last year, at 60 to 90 percent lower cost than Western alternatives. For engineering teams and smaller companies, this could mean substantial savings without sacrificing code quality—though token efficiency and routing smarter work to cheaper tiers matters as much as the model's headline price.
What to watch
Databricks found that its developers handle mostly medium-complexity tasks (61 percent), so the company plans to route more work to cheaper performance tiers based on complexity. The top-performing cluster includes Opus 4.8, GLM 5.2, and GPT 5.5 in certain configs, each hitting 82 to 90 percent pass rates; the Pareto frontier for best quality-to-cost ratio now spans models from OpenAI, Anthropic, and open-source providers.
The shift Databricks documents reflects a widening performance parity between Chinese open-source models and Western proprietary ones, combined with substantial cost advantages. On OpenRouter, Chinese models have grown from 11 percent of weekly coding traffic last year to 30 percent since February 2026, all while trading at 60 to 90 percent lower cost. This is not a story of Chinese models overtaking Western ones across the board—Databricks' own analysis shows no single lab dominates across all three performance tiers, and the top cluster still includes Opus, GLM 5.2, and GPT 5.5 in certain configurations. Rather, it is a story of cost-performance trade-offs becoming transparent and actionable at scale.
Databricks' methodology matters here: by building a benchmark from real pull requests in its own multi-language codebase rather than relying on public datasets, the company sidestepped a known problem—solution leakage into training data—that even OpenAI has cautioned against. The finding that token efficiency (how much context a model sends, measured like fuel economy) varies widely by software environment may be more practically important than headline token prices. Token efficiency can offset headline cost differences: Databricks found that its Pi harness sent about three times less context than Claude Code, making Opus 4.8 at "high effort" 2.08× cheaper per task at comparable pass rates.
For engineering teams, the implication is not that one model is universally cheaper, but that routing decisions matter. Databricks found that 61 percent of its developers' tasks are medium complexity, with only 12 percent high; the company now plans to match model tiers to task difficulty rather than defaulting to the most expensive option. This multi-tier approach, combined with the demonstrated parity at the top between Western and Chinese models, suggests that cost pressure will increasingly force Western AI providers to either lower prices or specialize in use cases where their models demonstrably outperform cheaper alternatives.
No comments yet. Be the first to share your thoughts!
Log in to join the discussion
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
