Short Summary
OpenAI’s GPT-6 Astra went live on OpenRouter on September 4, 2026, priced at $10 per million input tokens and $50 per million output tokens — roughly 5x more expensive than the previous flagship, GPT-5.6 Sol. In exchange, it offers a 1.05-million-token context window and a 94.6% GPQA Diamond score. My take: it’s worth the premium only for genuinely long agentic work, not general chat.
What GPT-6 Astra Actually Is, and Why OpenRouter Prices It So High
OpenAI positions GPT-6 Astra as its “flagship model for demanding end-to-end work,” built for advanced analysis, software engineering, deep research, scientific work, and document creation. The clearest jump over the previous generation is in long-horizon agentic tasks — multi-step work that involves controlling a computer or browser on the user’s behalf.
On the technical side, Astra supports up to 1.05 million tokens of context and can return up to 128,000 completion tokens per request. It accepts multimodal input — PDFs, images, and text files — and supports tool calling plus structured output via JSON schema. Those extra capabilities are also why it costs more per request than the previous generation: every one of them adds tokens and compute time.
How Much More Expensive Is It Than GPT-5.6 Sol, in Practice
On OpenRouter, GPT-6 Astra costs $10 per million input tokens and $50 per million output tokens. Cache reads run $1 per million tokens, cache writes $12.50 per million tokens, and web search adds $10 per 1,000 calls if enabled.
Compare that to GPT-5.6 Sol, the previous flagship released July 9, 2026: $2 per million input tokens and $10 per million output tokens. In other words, Astra costs roughly 5x more on both input and output than Sol, despite sharing the same 1.05-million-token context window. Any workload that ran comfortably on Sol will see its monthly bill jump by that same ratio if moved to Astra wholesale.
Speed, Latency, and Reliability, as Measured on OpenRouter
OpenRouter’s usage data puts Astra’s peak throughput at around 45 tokens per second, with P50 latency at 2.44 seconds. On reliability, the OpenAI Flex provider reports 100% uptime and 99.22% availability over a trailing 3-day window.
On quality, Astra scores 94.6% on GPQA Diamond, ahead of GPT-5.6 Sol’s 92.4%. OpenRouter hasn’t published a TAU-Bench score for Astra yet, so there’s no direct agentic-benchmark comparison available — Sol’s TAU-Bench score sits at 76.0%.
Astra vs. GPT-5.6 Sol, Side by Side
Here’s what OpenRouter has published for both models, laid out together so it’s clear what the higher price actually buys.
| Factor | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Release date | Sep 4, 2026 | Jul 9, 2026 |
| Context window | 1.05M tokens | 1.05M tokens |
| Input price | $10 / 1M tokens | $2 / 1M tokens |
| Output price | $50 / 1M tokens | $10 / 1M tokens |
| Cache read price | $1 / 1M tokens | $0.20 / 1M tokens |
| GPQA Diamond | 94.6% | 92.4% |
| Strength | Long agentic tasks, computer/browser control | Command-line coding and multi-step reasoning |
Who’s Actually Using Astra on OpenRouter, and When to Adopt It
OpenRouter’s usage data shows Codex is the top caller of Astra by volume at 61.6 billion tokens, followed by Hermes Agent at 61.2 billion tokens and Claude Code at 16.1 billion tokens. That mix points to coding agents and automated workflows running long, continuous sessions — not general-purpose chatbots.
My recommendation: adopt Astra specifically for work that genuinely needs the long context, like reviewing an entire codebase or research that requires reading dozens of documents in one pass. For short Q&A or general drafting, GPT-5.6 Sol still wins on value — it costs five times less while trailing on benchmarks by less than 3 points.
