Vercel says GPT-6 Astra leads DeepSecBench in 49 minutes

• GPT-6 Astra now leads Vercel's DeepSecBench, where rauchg's DeepSecBench post reports a higher vulnerability-finding score than GPT-5.6 Sol and Claude Opus 5.
• The end-to-end run took 49 minutes rather than roughly four hours for Sol, according to rauchg's result, while Vercel's benchmark update characterizes Astra's per-task cost as only slightly higher than the prior leader's.
• A separate DeepSWE run posted by haider1 put Astra at 30,000 output tokens and 29 steps, versus Sol's 60,000 and 61, at nearly identical cost and pass rate.
• Access is splitting by surface: OpenAIDevs' rollout starts with a limited group before ChatGPT, API, and AWS expansion, while synthwavedd's Daybreak report puts advanced cyber access first with approved defenders.

One detail in Vercel's AI Gateway changelog is that Astra can continue independent work while waiting for clarification. The OpenAI release page lists the paid ChatGPT plans, API, Azure, and Bedrock as the broader rollout surfaces.

Vercel's top three DeepSecBench entries were:

• GPT-6 Astra xhigh: 37.79
• GPT-5.6 Sol xhigh: 35.44
• Claude Opus 5 max: 32.44

Vercel's DeepSecBench leaderboard combines recall and precision into its score, while also recording cost, runtime, completion rate, and task counts. The 49-minute number is therefore an end-to-end agent-run result, rather than a raw model-latency measurement.

DeepSWE is a separate software-engineering benchmark, but its posted run data shows a second efficiency signal:

• Astra xhigh: 74% plus or minus 3%, $6.52, 30,000 output tokens, 29 steps.
• GPT-5.6 Sol max: 73% plus or minus 3%, $6.46, 60,000 output tokens, 61 steps.

The rows use different effort labels, xhigh for Astra and max for Sol, so the table does not normalize effort mode. It does show that the reported run used roughly half the output tokens and action steps.

OpenAI separately said Astra sets records on Agents' Last Exam, AutomationBench, and ScreenSpot Pro, three benchmarks for professional computer workflows.

In sama's launch post, the company also cited 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench. Those are different task suites from DeepSecBench, so the Vercel result supplies a distinct code-security measurement rather than a confirmation of OpenAI's own cyber score.

Vercel exposes the production model as openai/gpt-6-astra through its AI Gateway integration. OpenAI president Greg Brockman called Astra a “generational leap” in an Axios report excerpted by synthwavedd.

The cyber policy boundary is narrower than general distribution. OpenAI says the model can assist authorized security work investigating unfamiliar software and potential vulnerabilities in OpenAIDevs' security post, while synthwavedd's access report says Daybreak defenders received the advanced cyber route first.

Commenter cramforce said the public non-cyber version performs defensive tasks without frequent refusals, then added in a follow-up that requests beyond a defined path still receive refusals.

SRE-Bench is a separate reverse-engineering evaluation built from 262 binary instances based on 19 private programs. The targets span C, C++, Go, and Rust, and each instance requires all six task-specific objectives for a full solve.

The reported SRE-Bench result was 88.0% for Astra at pass@1 and 99.2% at pass@4, versus Sol's 55.9% and 68.7%; the pass@4 figure aggregates four independent trials. ValsAI noted that OpenAI used a custom harness without step limits in the same SRE-Bench thread.

The benchmark is now seeking private programs and anti-analysis tooling as harder targets, according to ValsAI's call for contributions and its contribution guide.