Databricks defaults its coding to GLM-5.2, and Apple sues OpenAI

The week on the beat

1. Databricks makes GLM-5.2 its default coding engine

After benchmarking coding agents on its own multi-million-line codebase, Databricks found the Chinese open-weights model matched Opus 4.8 on pass rate at $1.28 per task against $1.94. This is the model whose third-party price could not sit still two weeks ago; cheap and good is winning enterprise defaults anyway.

Read the story →

2. Apple sues OpenAI over trade secrets

The complaint alleges OpenAI encouraged former Apple employees to bring confidential prototypes, roadmaps and supplier data to their new roles. Two companies most builders depend on somewhere in their stack are now in open legal conflict over hardware plans.

Read the story →

3. Half the Fortune 500 now uses Hugging Face

CEO Clem Delangue says companies start on frontier APIs and migrate to open models as costs scale, with production workloads increasingly self-hosted. If you are betting on open weights, the enterprise is quietly betting with you.

Read the story →

4. OpenAI builds GPT-Red, an automated red-teamer

An internal system that attacks OpenAI's own models with adversarial prompts and injections during training. If it works, safety testing starts scaling with the models instead of with the headcount doing it by hand.

Read the story →

5. DeepSeek is raising again, weeks after $7 billion

The FT reports the lab is considering more capital on top of its recent round. Read together with last week's custom-chip story: DeepSeek is spending like a company that intends to keep undercutting Western labs on price for years.

Read the story →

6. Apple Intelligence clears its biggest regulatory gate

The week's most-covered story (27 outlets): Apple registered its AI service with China's cyberspace regulator, the mandatory step before generative AI can ship there, with Alibaba and Baidu as local partners.

Read the story →

Model moves

• New: the GPT-5.6 family (OpenAI). Six models in one drop: Sol at $5/$30 per 1M, Terra at $2.50/$15, Luna at $1/$6, each with a Pro variant, all with a 1,050,000-token context window. Model page →
• New: Muse Spark 1.1 (Meta), and Meta's first paid model API. A multimodal agentic model at $1.25/$4.25 per 1M with a 1M-token context, sold through the new Meta Model API (public preview, US-only). Not on OpenRouter yet, so the tracker shows no floor price. Model page →
• New: KAT-Coder-Pro V2.5 and KAT-Coder-Air V2.5 (Kwaipilot). Two coding models at $0.74/$2.96 and $0.15/$0.60 per 1M, both with 256K context. Model page →
• Price change: Kimi K2.5's floor dropped a third. DigitalOcean now serves it at $0.375/$2.025 per 1M, down 31 and 25 percent. Moonshot's own list price is unchanged at $0.60/$3.00. Third-party floors swing; ride the discount with a fallback wired in. Model page →
• Price change: GLM-5.1's floor discount evaporated. DigitalOcean raised output from $3.08 to $4.30 per 1M (+40 percent), now a hair under Z.ai's $4.40 list price. Model page →
• Price change: gpt-oss-120b floor down 17 percent. OpenInference now serves it at $0.03/$0.15 per 1M. Model page →
• Benchmark: GLM-5.1's SimpleBench score was revised down from 58.7 to 55.1 on Epoch's latest run. Model page →
• Context: DeepSeek-V3.2 grew from 131,072 to 163,840 tokens. Model page →

Personal take

GLM-5.2 continues to be in the news and in discussions across developer communities. Two weeks ago, its 3rd-party price could not sit still, and I said not to build a margin on it without a fallback. This week, Databricks made it the default coding engine after it matched Opus 4.8 at two-thirds the per-task cost. Both things are true at once: open-weight models are good enough to win enterprise defaults, and their pricing still behaves like a spot market. The same providers cut Kimi K2.5 by a third and raised GLM 5.1 by 40% within 3 days. The playbook for developers remains the same: use the cheap models and engineer around the wobble. Meanwhile, OpenAI shipped 6 GPT-5.6 variants, and Meta dropped its first paid API. Picking a model is now a procurement decision, and the frontier labs know it.

Until next Thursday, Anmol