GPT-6 Astra on OpenRouter

TL;DR Highlight

GPT-6 Astra: OpenAI flagship at $10/$50 per 1M tokens, 1.1M context, built for agentic work

Who Should Read

An LLM API or agent developer weighing cost and performance before adopting OpenAI's new flagship model

Core Mechanics

Astra is the flagship of OpenAI's GPT-6 series
- It is pitched for advanced analysis, software engineering, deep research, and document creation.
- Its stated strength is long-horizon agentic tasks using a computer or browser.
Pricing is $10 input and $50 output per 1M tokens
- Cache read is $1.00 and cache write is $12.50.
- Context is 1,050,000 tokens with up to 128,000 output tokens.
Prompt caching is on by default
- Prompts of 1024+ tokens are cached automatically, and a cache write is billed even if never read.
- To disable it, set prompt_cache_options.mode to "explicit" and use no prompt_cache_breakpoint.
Higher reasoning effort mostly raises scores
- Artificial Analysis Intelligence Index: Low 45.8, Medium 49.6, High 50.9, Xhigh 52.4, Max 52.7.
- Terminal-Bench 4.0 rises from 41.9% at Low to 59.6% at Xhigh.
Speed tiers change the price for the same model

Community Reactions

Astra low gave better results for the same budget, per one commenter
- With a 10-cent budget for a pelican SVG, Astra low beat the other models by a wide margin.
- They also say it uses fewer tokens overall.
Strong vision and frontend recreation, but costly, per one commenter
- Building a page from a design image, it reproduced SVG curves more accurately than Opus 5.
- That single site build cost $24.
Switching from Sol to Astra mid-session caught critical bugs, per one commenter
- It quickly found a resource-lifetime bug in a rendering pipeline.
Price and reliability concerns were raised
- One view is that $10/$50 is far too expensive versus Chinese models.
- One commenter noted the OpenAI provider's tool-call failure rate is 5% versus 0.2-0.5% on Azure.
- One user reported an account suspension after loading $25 on OpenRouter with no support response.

How to Apply

Compare Astra and Sol on your own prompts
- Pick 5 representative prompts, run both models on the same input, and record quality and token usage.
- With OpenRouter you only swap the model slug, so it fits in 30 minutes.
Build a cost-vs-score table by reasoning effort
- Run Low through High and compare accuracy and cost on your workload.
- Start with agentic tasks, where the source benchmarks show the biggest Low-to-High gap.
Check cache write charges and disable them if unneeded
- For one-off long prompts, check whether the $12.50/M cache write applies.
- Set prompt_cache_options.mode to explicit on requests that don't need caching.
Route only hard tasks to Astra
- Send everyday coding and classification to Sol or Luna.
- Send only hard work like bug review to Astra.

Try It

copy & run

# Run in a terminal (needs an OpenRouter key; the API is OpenAI-compatible, so only the base URL changes)

pip install openai

import os
from openai import OpenAI

client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key=os.environ["OPENROUTER_API_KEY"])
prompt = "{{prompt to compare}}"
for model in ["openai/gpt-6-astra", "{{other model slug}}"]:
    r = client.chat.completions.create(model=model, messages=[{"role": "user", "content": prompt}])
    print(model, r.usage)
    print(r.choices[0].message.content[:500])

# Next: compare answer quality and token counts in usage, then convert to cost

Terminology

Reasoning effortA setting for how long the model thinks before answering. It is like giving a test taker more time.

Prompt cachingStoring the repeated start of a prompt for reuse. Reads are cheap, but this model also bills a separate cache write.

TTFTTime until the first token appears. It is the wait before a user sees the first character.

Provider routingSending a request to one of several companies hosting the same model. It is like choosing a courier.