01
Build & debug
Coding, tool use, and complex reasoning in one workflow.
What to check
- Does the main user flow work?
- Does it work on a narrow screen?
- Are errors handled clearly?
THE ASTRA FIELD GUIDE / SEPTEMBER 2026
Big capabilities.
Real things to try.
GPT-6 Astra is in the cabin. Explore the documented capabilities, choose a task, and take a clear test prompt into your own session.
1.05MAPI context window
128KMaximum output tokens
Text + imagesInput · text output
5 levelsLow through max reasoning
Vendor documentation · Checked September 5, 2026 · Model specifications ↗. Features depend on the app, tools, and access available to you.
CAPABILITY MAP
Choose the work. Then test it.
02
Research & verify
Research with web search and connected tools.
What to check
- Do the sources support each claim?
- Are facts separated from inference?
- Are source and event dates clear?
03
Create & analyze
Document creation and long-context reasoning.
What to check
- Are the source facts preserved?
- Do the calculations reconcile?
- Is the result ready for its audience?
04
Inspect & explain
Text and image input, with text output.
What to check
- Is each finding visible in the image?
- Are assumptions clearly labeled?
- Can someone act on the advice?
05
Work across tools
Computer use, MCP, and asynchronous tool calling.
What to check
- Did it use the available tools?
- Did it respect the task boundaries?
- Can the work be checked and repeated?
06
Steer as you go
Mid-turn steering and adjustable reasoning effort.
What to check
- Did it incorporate the correction?
- Did completed work stay useful?
- Did it still finish the original goal?
Capability descriptions: OpenAI model card and Astra guide. Test ideas and success checks are TonkaToyXL’s suggested exercises, not published benchmark results.
THE PROMPT WORKBENCH
Give Astra a job worth doing.
A useful test needs a real task, a defined output, and a way to tell if it worked.
01 / SET UP YOUR TEST
This builds a prompt in your browser. It does not send your text to a model or run a paid API call.
02 / TAKE IT FOR A TEST
Build & debug
Copy this into a session where you have Astra access. Save the output and compare it against the success checks.
NEW FINDINGS / SEPTEMBER 4–5
What happened in the cabin.
Results reported in the latest TonkaToyXL videos. Follow each one back to the test and its limits.
5 / 5BOUNDED FUNCTIONAL CHECKS
A rescue game, five checks.
The published Cabin Rescue Run report records five passes on the first attempt, with no code repairs. One task, one fixed seed; the capture does not measure real-time performance.
Watch the rescue test ↗1 patchVISUAL REFINEMENT DISCLOSED
A storm cabin, with the patch shown.
Storm Watch passed five bounded checks. One Astra-authored patch fixed a label overlap. It is an illustrative simulation, not a weather forecast or engineering model.
Watch the storm test ↗3 buildsONE ATTEMPT + TWO REFINEMENTS
Four models. Uneven results.
The latest comparison covers a kart game, skybridge, and Blender cabin. It preserves regressions and unavailable outputs. Estimated API-equivalent costs are not actual subscription invoices.
Inspect the full comparison ↗WHAT’S NEW
More control over work in progress.
Async tools let supported applications keep independent work moving. Mid-turn steering lets you revise instructions during a task. Reasoning effort can change during a conversation while preserving the prompt cache.
See feature requirements ↗THE CABIN STANDARD
Capabilities are a starting point.
Our latest videos report bounded task results. Record the exact model, prompt, tools, effort, retries, time, and output before making a comparison. A task result is not a universal ranking.
Our existing benchmark comparison is a dated August 29 snapshot. It does not include an Astra ranking.
Inspect the benchmark snapshot →How we label evidence →