OpenRouter lists GPT-6 Astra with 1M context, $10/$50 pricing and multi-provider routing
- OpenRouter lists OpenAI's GPT-6 Astra as a flagship model for end-to-end work, with a 1M-token context window and a listed release date of September 4, 2026.
- Listed pricing is $10 per million input tokens and $50 per million output tokens; OpenAI Flex lists $5 and $25, while OpenAI Fast lists $20 and $100.
- OpenRouter routes Astra across OpenAI and Azure providers using Balanced, Nitro, or Exacto modes; it reports 98.42% availability over three days, versus 95.19% without routing.
- The best reported provider results are 67 tokens per second throughput from OpenAI Flex and 3.37 seconds P50 latency from OpenAI; Azure (US) has the lowest reported tool-call error rate at 0.35%.
- OpenRouter reports 54.7 on Artificial Analysis' Intelligence Index, 76.9 on its Coding Index, and 51.6 on its Agentic Index, placing Astra above 95%, 96%, and 91% of compared models respectively.
Hacker News opinions
I got Astra on my Pro plan after about 24 hours. OpenAI gave bankable resets for each day Pro users were waiting, although I almost wish they had waited longer before enabling mine.
I have GPT-6 access in Codex and the OpenAI API through a Business plan with Cyber verification enabled. I am still trying to find whether Astra consumes subscription usage differently from Sol.
I am a European Plus user and got access, but only in Codex, not ChatGPT.
Simon Willison's pelican comparison is useful. At a 10-cent budget, Astra low produced a much better result than the other models and used fewer tokens.
I am concerned that these public image grids may already be in model training data. Astra's pelicans also look unusually similar to one another, including the yellow sun, green bike, and composition.
The bicycle geometry looks better to me, especially the front fork. A model that learns physical constraints through 3D simulation and reinforcement learning should reject many of the impossible arrangements other models generate.
Astra's pelican is very close to GLM 5.3's, down to the bike color and scarf. The improved front fork is about the only clear advantage I see.
OpenAI put a 3D pelican-on-a-bicycle easter egg in a promo video, so I suspect it spent some training effort on producing a good SVG version of that prompt.
Aibenchy's Astra SVG result is striking, but OpenRouter initially returned Not Found errors for the model ID while I tested it.
I do not understand Aibenchy's 98.58 score. The result has broken perspective, two mouths on the animal, an incomplete paw, and unexplained plus signs, so calling it near-perfect seems wrong.
If SVG generation reaches this level, logos and app icons could become one-shot work instead of hours of manual SVG editing. I once spent days getting a dark-mode toggle icon right.
Azure costs twice as much as the regular OpenAI route here. I want to know what it provides beyond calling OpenAI directly.
Azure is ZDR, while the OpenAI routes are not. It also provides Microsoft's guarantees and Azure identity access, which OpenAI does not provide without dedicated spend.
After a couple of hours in the Codex app, Astra feels much faster than Sol to me even though its tokens-per-second rate is about half.
Astra handled non-90-degree cutouts and flowing SVG lines in my web design reconstruction far better than Opus 5. The result cost me $24 for a simple frontend, so the quality is expensive.