OpenAI throws Astra into the top-tier model ring
- ID
- 21287
- Status
- summarized
- Published
- 04 Sep 2026, 6:42 AM
- Fetched
- 04 Sep 2026, 7:17 AM
- Provider
- The Register
- Category
- technology
- Original URL
- https://www.theregister.com/ai-and-ml/2026/09/03/openai-throws-astra-into-the-top-tier-model-ring/5294378
- Source URL
- https://www.theregister.com/headlines.atom
Summary
- Score
- 7.5
- Created
- 04 Sep 2026, 7:17 AM
- Tags
- Audience
- developersai_ml_learnersai_agent_userssaas_founders
What happened
OpenAI released GPT-6 Astra after a training pause over AI security concerns, initially to Trusted Access Program participants with broader rollout to paid subscribers in days. It is priced at $10/million input tokens and $50/million output tokens—matching Anthropic's Fable 5.1 released two days prior. OpenAI claims hallucination dropped to 2% (from 9.4% on GPT-5.6 Sol), out-of-scope actions fell to 0% (from 48%), and it is the first model rated 'Critical' for autonomous cybersecurity exploitation capabilities.
Why it matters
At $10/$50 per million tokens, Astra is priced at the top tier—builders shipping agent workloads need to model cost carefully before defaulting to it. The claimed 0% out-of-scope rate and 2% hallucination rate are concrete enough to benchmark against your own evals before trusting it in production agent pipelines, especially given the 'Critical' cyber capability rating.
Discussion angle
Compare Astra's $10/$50 pricing and claimed benchmarks against what you're currently paying for Claude or Gemini—does the hallucination and out-of-scope improvement justify the cost for your specific agent use cases, or is this still too expensive for high-volume workloads?