Databricks Rolls Out GPT-6 Astra to All 3,500 Engineers

Databricks has rolled out GPT-6 Astra to all of its 3,500 engineers, following a pilot involving around 200 users to assess coding quality and costs. 

The company used the pilot to assess model quality and coding costs before making Astra available across its engineering organisation.

“Astra unambiguously outperforms our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks, especially those related to high-level system design or long-range horizontal tasks,” Patrick Wendell, Co-founder and Vice President of Engineering at Databricks, said in a post on X.

Engineers with access to Astra increased their overall coding spend by around 60% compared with the baseline, according to Wendell. However, the company didn’t find much improvement on medium- and low-complexity coding tasks. 

Wendell said Databricks suspects existing models already perform those tasks near saturation, leaving less room for Astra to deliver meaningful gains.

Databricks used Unity Gateway, its internal model-access and experimentation layer, to run cohort-based tests during the pilot. The company uses the system to evaluate new models before wider deployment.

Rather than making Astra the default model for every task, Databricks is assigning engineers a separate Astra budget. 

The approach is intended to encourage selective use for complex work while retaining lower-cost models for routine coding. Engineers can mix and match tools and models within their overall budget, with additional budget increases available through internal mechanisms.

“We give engineers a sub-budget specific to Astra to encourage them to use Astra selectively on complex tasks while preferring lower cost models for everyday tasks,” Wendell said.

GPT-6 Astra is OpenAI’s best-performing model yet for complex reasoning, software engineering, computer use, and professional workflows.

OpenAI said Astra achieved 97.6% on FrontierMath Tier 4 research-level mathematical problems, 99.9% on ARC-AGI-3 that tests how AI agents explore unfamiliar environments, and 100% on cybersecurity benchmark ExploitBench—benchmarks the company said represented new highs among the competition. 

It also scored 57.9% on Terminal-Bench 4.0, compared to 37.3% for GPT-5.6 Sol, to assess how AI agents perform multi-step engineering work inside real command-line and terminal environments.