GPT-6 Astra
Explore benchmark scores, genre strengths, weaknesses, and recent examples for GPT-6 Astra on Orivel.
Model Overview
Released
2026-09-03
Input
$10.00 / 1M
Output
$50.00 / 1M
Max output
128k tokens
OpenAI's newest flagship, released September 3, 2026, and the model OpenAI's president called a possible arrival of AGI. On Orivel it takes the OpenAI flagship slot from GPT-5.6, which moves to the balanced position. An Astra Pro variant exists; Orivel uses the standard gpt-6-astra, as it does for every generation.
Specs and pricing
| Item | GPT-6 Astra |
|---|---|
| Context window | 1.05M tokens |
| Max input / output | 922k / 128k tokens |
| Input / 1M | $10 |
| Output / 1M | $50 |
| Cached input / 1M | $1 |
| Over 272k input tokens | 2x input, 1.5x output, for the whole request |
| Reasoning effort | low through max |
API behaviour we measured
| Item | Behaviour |
|---|---|
| Structured output (strict json_schema) | Works |
| temperature | Rejected (400 error) |
| Streaming, function calling | Supported |
| Realtime, Assistants, fine-tuning, embeddings, audio, image generation | Not supported |
Is the price the whole story here?
Astra costs two and a half times the model it replaces in this slot, and the same as Anthropic's flagship. That is a real jump, and it lands on a site where the previous OpenAI flagship was the cheapest frontier option available. If your work does not need frontier reasoning, GPT-5.6 sits one tier down at a fraction of the rate and remains the better default.
The part worth planning around is the safety posture rather than the benchmarks. OpenAI shipped Astra with an explicit warning about its cyber capabilities and a public build that declines some prompts. We checked before adopting it: coding, system design, an explanation of SQL injection with vulnerable and fixed code, and a latency post-mortem all came back answered. The restrictions are real but narrower than the announcement suggests, so treat refusals as something to verify for your own workload rather than assume in either direction.
What changed
- Released September 3, 2026; OpenAI describes it as a generational leap in cybersecurity, professional work, software engineering and science
- Rolled out first to organisations in OpenAI's Daybreak cybersecurity programme, then to paid ChatGPT plans, the API, Azure and AWS Bedrock
- Context window 1.05M tokens; up to 922k input and 128k output
- Pricing: $10 input / $50 output per 1M tokens; cached input $1; requests over 272k input tokens billed at 2x input and 1.5x output
- reasoning.effort supports low, medium, high, xhigh and max
- Rejects `temperature` with a 400; structured outputs, streaming and function calling all work
- The public build declines some cybersecurity prompts; OpenAI warned about the model's cyber capabilities at launch
Specs across the current lineup
Published specs and pricing, comparable before any benchmark data exists.
| Model | Context | Max output | Input / 1M | Output / 1M |
|---|---|---|---|---|
| openai GPT-6 Astra | — | 128k | $10.00 | $50.00 |
| google Gemini 3.8 Flash | — | 66k | $0.75 | $3.75 |
| anthropic Claude Fable 5.1 | 1M | 128k | $10.00 | $50.00 |
| google Gemini 3.5 Flash-Lite | — | 66k | $0.30 | $2.50 |
| openai GPT-5.6 | 1M | 128k | $4.00 | $20.00 |
| anthropic Claude Opus 5 | 1M | 128k | $5.00 | $25.00 |
| openai GPT-5 mini | 400k | 128k | $0.25 | $2.00 |
| google Gemini 3.1 Flash-Lite | — | 66k | $0.25 | $1.50 |
| anthropic Claude Sonnet 5 | 1M | 128k | $2.00 | $10.00 |
Overall Performance
Overall Rank
#6
Overall win rate
Average Score
Wins
1
Sample Count
2
Win Rate by Model
| Model | Wins | Losses | Draws | Win Rate | Detail |
|---|---|---|---|---|---|
| Google Gemini 3.8 Flash | 1 | 0 | 0 |
100%
|
View Gemini 3.8 Flash vs GPT-6 Astra Comparison & Evaluation |
| Anthropic Claude Opus 5 | 0 | 1 | 0 |
0%
|
View Claude Opus 5 vs GPT-6 Astra Comparison & Evaluation |
Compare by Genre
Strength by Evaluation Criteria
Average score by criterion (out of 10)
Latest Tasks
Explanation
Explaining Eventual Consistency to a Relational Database Developer
Explain the concept of 'eventual consistency' in distributed systems to a junior software developer who is only familiar with traditional relational databases a...
Business Writing
Announcing a Delayed Product Launch to Enterprise Customers
You are the Director of Customer Success at Northwind Analytics, a B2B software company with about 400 enterprise customers. Your team must announce that the la...