Why GPT-6 Astra
burns quota so fast
On OpenAI's own credits table, Astra costs 250 (input) / 25 (cached) / 1,250 (output) credits per 1M tokens, while GPT-5.6 Sol costs 100 / 10 / 500. All three ratios are exactly 2.5×. The same workload therefore gets you roughly 40% as far on Astra. This page has the official numbers, the per-plan message allowances, and how to route by task.
Updated 2026-09-09
Four numbers that decide how long you last
Quota cost relative to Sol
Input 250 vs 100, cached 25 vs 10, output 1,250 vs 500 credits per 1M. All three ratios match exactly — this is not an estimate.
Plus messages per 5 hours
On the same official table Sol is 10–100, Terra 25–200, Luna 250–2,000. Astra is exactly half of Sol.
Max output, reasoning included
Reasoning tokens count toward output, and output credits cost 5× input. Budgeting on the visible reply alone badly understates it.
Input threshold for long-context pricing
Above this input size the API rates become $20 / $2 / $25 / $75. The pricing page does not say whether that reprices the whole request or only the overage.
It is not a bug, it is the list price
The first reaction after moving to Astra is usually "is my quota broken?". By the official credits table it is not: Astra's three rates are cleanly 2.5× Sol's. Within the same subscription allowance, Astra gets through roughly 40% of the work Sol would. If your workflow is a long-running agent, the 5-hour window will bottom out noticeably earlier — that is by design, not a fault.
What the official per-plan table says
OpenAI's pricing documentation gives ranges of local messages per 5 hours: Astra is 5–45 on Plus and Business, 25–225 on Pro 5x, and 100–900 on Pro 20x. On the same table Sol is 10–100 / 50–500 / 200–2,000 and Luna is 250–2,000 / 1,250–10,000 / 5,000–40,000. The ranges are wide because what one message costs depends on context length and reasoning volume.
Timeline
GPT-6 Astra ships with API model id gpt-6-astra. The system card publishes the same day, stating this is OpenAI's first model to reach the Critical cybersecurity level under its Preparedness Framework.
Astra enters Codex. The launch post says it rolled out that day to a limited set of organizations and over the following days to Plus, Pro, Business and Enterprise, plus the OpenAI API, Azure and AWS Bedrock.
This page was checked against the official credits table and pricing page that day: Astra 250 / 25 / 1,250 versus Sol 100 / 10 / 500, a 2.5× ratio. API list price for Astra is $10 input and $50 output per 1M tokens.
What is official and what is not
✅ From the official docs
The credit rates 250 / 25 / 1,250 and the per-plan message ranges come from OpenAI's pricing documentation. The API list price $10 / $1 / $12.50 / $50, and the $20 / $2 / $25 / $75 tier above 272,000 input tokens, come from the API pricing page. The 1,050,000 context, the 128,000 output ceiling including reasoning tokens, and the reasoning effort levels come from the official model documentation.
⚠️ Community observations, not official numbers
Various measured ratios circulate in the community — that a session with very high cache hit rates still drains a 5-hour window quickly, or that the real gap against Sol is wider than the list price suggests. These are user observations; OpenAI has not confirmed any specific multiple. To judge your own consumption, run the same real requests against both model ids and read the bill, rather than trusting a number someone repeated.
Same work, two ways to run it
Move everything to Astra
Genuinely stronger on long context, tool orchestration and computer use, but the same subscription allowance gets through roughly 40% of what Sol would. Suits a small volume of hard work.
Route by task
Send long context and complex tool chains to Astra; keep routine edits and test writing on a value tier. Most teams get more total output and a more predictable bill.
How to burn less
Three things help most. First, drop the reasoning effort: Astra supports low, medium, high, xhigh and max, and does not support none — if you were on none or minimal, start at low and compare. Second, control output: the 128,000 ceiling includes reasoning tokens and output credits cost 5× input, so making the model write less saves more than making it read less. Third, watch the 272,000 input threshold — crossing it moves the API to a more expensive tier.
On QCode
gpt-6-astra is in QCode's model table and has real traffic over the last 30 days. It shares one key and one quota with the three GPT-5.6 tiers, Claude and Gemini, so switching is a change to the model field. That means you can run the same real requests against both model ids and read the cost difference directly instead of guessing from someone else's benchmark.
Frequently asked questions
How much more does Astra cost than Sol?
On the official credits table, 250 vs 100 input, 25 vs 10 cached and 1,250 vs 500 output per 1M tokens — all three are 2.5×. Within the same subscription allowance, Astra gets through roughly 40% of the work.
Why does my 5-hour window drain faster than someone else's?
The official figure is a range, not a constant: 5–45 messages on Plus. What one message costs depends on context length and reasoning volume, so long-context sessions at a high effort setting land at the bottom of the range.
Does lowering the reasoning effort actually help?
Yes, but measure it yourself. Astra supports low, medium, high, xhigh and max, and not none. The official migration guide says anything previously on none or minimal should start at low and be compared. Reasoning tokens count toward output, and output costs 5× input, so lowering the effort saves on the most expensive part.
Above 272,000 input tokens, is the whole request repriced?
The pricing page lists a higher tier for input above 272,000 but does not say whether it reprices the entire request or only the overage. That distinction moves the bill noticeably on long-context work, so measure with a small request before scaling up.
Should I move everything to Astra?
For most teams, no. Routing by task is safer: long context, complex tool orchestration and computer use go to Astra, routine edits and test writing stay on a value tier. On QCode both sit behind one key and one quota, so you can benchmark them on the same real requests.
Can I call gpt-6-astra on QCode?
Yes. It is in the model table and has real traffic over the last 30 days. It works like any other model: same endpoint, same key, set model to gpt-6-astra. Pricing tracks the official rate; the live figure is on the models page.
Sources
Credit rates and per-plan message ranges: OpenAI pricing documentation. API list price and the long-context threshold: OpenAI API pricing page. Context and output ceilings and reasoning levels: OpenAI model documentation. Release date and Critical level: the GPT-6 Astra system card. All captured 2026-09-09. Anything marked as a community observation is not an official number.
Measure it once on your own requests
Astra, the three GPT-5.6 tiers, Claude and Gemini share one QCode key and one quota; switching is a model-id change. Plans start at ¥60/month and the key is live the moment you pay.
Related reading
GPT-6 Astra complete guide
Specs, pricing, four migration constraints, and what OpenAI still has not published.
Codex 5-hour limit explained
How the window is counted, when it resets, and what to do when you hit it.
Choosing between Sol, Terra and Luna
Positioning, price and fit for the three GPT-5.6 tiers side by side.
Figures on this page were captured 2026-09-09 from OpenAI's pricing documentation, API pricing page, model documentation and system card; OpenAI may change them, so check the official page before committing. Anything marked as a community observation is not an official number and is not a performance guarantee.