# GPT-6 Astra’s balloon demo stops at 6.5 despite three review rounds

**GPT-6 Astra can use a critic agent to improve a build, but a critic’s score cannot certify the result.** In a demonstration account reviewed as of September 11, 2026, a water-balloon simulation received scores of 4.7, 6 and 6.5. All three fell short of the prompt’s eight-point target.

The useful result is the record of unfinished work. The builder made revisions, the critic continued finding problems, and the loop stopped when its budget of three rounds ran out. Developers evaluating Astra should preserve that distinction between finishing a run and passing acceptance.

A balloon bursting around a projectile makes an appealing visual test. It also exposes a measurement problem: realistic reflections, plausible water motion and numerically accurate fluid simulation require different evidence.

![Official GPT-6 Astra model identity](https://s4.tenten.co/learning/content/images/2026/09/landing-page-3-10.png)

#### GPT-6 Astra supplies capacity; the environment supplies access

OpenAI lists a 1,050,000-token context window, a maximum input of 922,000 tokens and a maximum output of 128,000 tokens. Those are capacity limits, not guarantees that every detail will survive a long project correctly. Astra accepts text and images and produces text; image creation and software operation depend on tools supplied to the model. [Official model specifications](https://developers.openai.com/api/docs/models/gpt-6-astra)

The surrounding workspace matters. A browser needs access to the running application. The agent needs persistent files and an appropriate development environment. A Blender-to-Unreal workflow needs installed software, an actual project and a working tool connection. Naming a tool in a prompt does not install or configure it.

The desktop app documentation describes projects, folders and artifact inspection. It also explains the Chat and Work toggle in the ChatGPT interface. For development, check which files and tools the chosen environment can actually access. [Desktop app documentation](https://learn.chatgpt.com/docs/app)

![Chat and Work toggle](https://s4.tenten.co/learning/content/images/2026/09/landing-page-2-11.png)

The official Work entry-point illustration; available features depend on the account interface.

Ultra requires a separate clarification. OpenAI describes it as a mode that uses subagents. The API’s `reasoning.effort` setting instead accepts `low`, `medium`, `high`, `xhigh` and `max`. It has no `ultra` value. Record both orchestration and reasoning settings when comparing runs. [Model and Ultra documentation](https://learn.chatgpt.com/docs/models)

#### Prompt one: make the critic inspect the balloon

A builder agent owns implementation. A separate critic opens the result, captures evidence and ranks defects. That separation gives the critic a concrete job beyond reading the builder’s completion message.

The following prompt preserves the demonstration’s requirements and its three-round limit.

```text
Build a ray-tracing simulation of a bullet piercing a water balloon. The balloon should pop, and the water should burst outward, deform from its original shape and eventually fall to the ground. Make it look as realistic and physically accurate as possible.

The animation must meet these requirements:
Let me freeze the animation at any point and view the scene from any angle.
Add sliders for bullet speed, balloon size, lighting, gravity and other parameters.
Do not use Three.js or any external web libraries. Build everything from scratch.

For each iteration, use a separate critic agent to capture screenshots from different angles and evaluate realism and physical accuracy.
Score the result from 0 to 10. A score of 8 or above means it looks convincing, like a real shot.
Continue until the critic scores it 8 or higher. Below that threshold, give the builder a ranked issue list and try again, for a maximum of three rounds.
```

The account reports about 32 minutes of execution and the following ratings. These are scores reported within the demonstration, not results from an independent Tenten test.

| Review round | Score out of 10 | Eight-point target |
| --- | ---: | --- |
| Round 1 | 4.7 | Not reached |
| Round 2 | 6 | Not reached |
| Round 3 | 6.5 | Not reached; round limit exhausted |

The defensible conclusion is narrow: the critic rated later versions more highly. Without a fixed rubric, source code and independent reruns, the 1.8-point gain cannot be converted into an improvement in physical accuracy. The record also cannot establish consistent grading across rounds.

For a visual demo, inspect refraction, surface continuity and visible clipping separately. For a physics teaching tool, define the motion model and check numerical behavior. Ray tracing describes how light contributes to an image. It does not, by itself, supply a physically valid water simulation.

#### Prompt two: character animation exposes the dependencies

The next task asks for a third-person character sprinting and jumping between rooftops in Unreal Engine. That creates requirements for animation, rigging, control, collision and asset import. A polished screenshot covers only part of the job.

The character URL, environment reference and Blender MCP endpoint were not supplied with the account. Replace the explicit placeholders below with valid resources before using the prompt. MCP, the Model Context Protocol, connects agents to external tools; the endpoint must correspond to an installed, running Blender integration.

```text
You are a world-class game developer. Build a procedural 3D game in Unreal Engine. The user controls this character from a third-person viewpoint:
<CHARACTER_SKETCHFAB_URL>

The character can sprint and jump very high, leaping across rooftops like a ninja.
Rig the character and add appropriate animations. Find relevant animations on Mixamo and retarget them to the character.
If downloading an asset requires login, use the Playwright Chrome extension to open my current Chrome session, where I am already signed in.

Build a procedural environment inspired by ancient Chinese imperial architecture. Use the attached image for visual inspiration:
<ENVIRONMENT_REFERENCE_IMAGE>

Generate all assets in Blender through MCP, then render them procedurally in Unreal Engine.
Blender MCP is installed and running at:
<BLENDER_MCP_ENDPOINT>

Make the game as detailed, realistic and grand as possible, with AAA-style graphics and effects. It must be playable inside Unreal Engine.

After each completed attempt, a separate critic agent must take its own screenshots from several viewpoints and zoom levels. Score game design and aesthetics from 0 to 10.
A score above 8.5 represents the requested AAA quality. Passing requires a score above 8.5 and zero errors.
Otherwise, give the builder a ranked issue list and try again, for a maximum of four rounds.
```

The prompt’s AAA label expresses an ambition. An 8.5 score is not an industry certification. Likewise, zero errors needs an explicit test scope: the project opens, the required route works, and the character does not pass through the ground on landing. One successful playthrough cannot establish the absence of every possible defect.

An existing login also has boundaries. Use an authorized session and let the user complete authentication challenges or missing access requirements. Confirm the permitted use of downloaded assets. A successful download does not establish unrestricted rights.

The game account reports roughly one hour and 26 minutes of execution, but the supplied material omits the complete outcome. It cannot establish that the game became playable, passed its reviews or met the requested visual standard.

![Work artifact inspection interface](https://s4.tenten.co/learning/content/images/2026/09/landing-page-1-12.png)

The official Work interface shows an inspectable presentation artifact. It is not a screenshot from the balloon or game demonstration.

#### Give the critic evidence and keep the acceptance gate fixed

A second role can catch omissions, but agents using the same model can share blind spots. More review also has a cost. OpenAI’s subagent documentation explains that each agent performs its own model and tool work, generally consuming more tokens than a comparable single-agent run. [Subagent workflows](https://learn.chatgpt.com/docs/agent-configuration/subagents)

A useful critic report identifies the action, time or camera view that exposes a defect. Keep fixed timestamps and camera positions for the balloon. Keep a repeatable input route for the game. These are proposed acceptance practices, not tests demonstrated by the supplied account.

Start with read-only access for the critic and let the builder own project changes. Require the critic to obtain its own screenshots and inspect relevant test output. It should not make a build pass by relaxing the acceptance conditions. Save the artifact version and findings from every round so that a disappearing defect can be investigated.

Public benchmarks have their own execution conditions. OpenAI reports 99.9% for Astra on ARC-AGI-3, with a footnote identifying its Responses API harness. That setup retains reasoning state and compacts context. ARC Prize Foundation distinguishes the Standard harness from a Provider Adapter that preserves more state. [Astra evaluation report](https://openai.com/index/gpt-6-astra/), [ARC Prize leaderboard methodology](https://arcprize.org/leaderboard)

A benchmark result belongs to that tested configuration. When sharing it, include the evaluation version, tools and context handling. None of those scores automatically validates a particular fluid simulation or game project.

#### GPT-6 Astra costs should follow accepted work

Standard API pricing is $10 per million input tokens and $50 per million output tokens. Requests exceeding 272,000 input tokens use twice the input and cache rates and 1.5 times the output rate for the entire request. [Official pricing](https://developers.openai.com/api/docs/models/gpt-6-astra)

For a simple arithmetic example, 100,000 uncached input tokens and 20,000 billable output tokens at Standard rates cost $2 in text tokens, before tool charges. This is not the balloon demonstration’s bill. Actual review rounds can consume very different amounts.

Subscription allowances need their own measurement. OpenAI says model choice, context, reasoning, tools and caching affect consumption. Two hour-long tasks can therefore use different amounts of an allowance. The demonstrator’s estimated weekly percentages are unsuitable as a general budget rule. [Usage and allowance documentation](https://learn.chatgpt.com/docs/pricing)

Teams should measure total cost per accepted artifact: builder and critic usage, tools, failed attempts and human review. Extra inference may pay for itself if it reduces manual correction. Repeated cosmetic revisions can raise spending without improving acceptance. Task records can distinguish those outcomes; elapsed time alone cannot.

#### Frequently asked questions

##### Is GPT-6 Astra Ultra the same as API max reasoning?

No. OpenAI describes Ultra as a mode involving subagents. The API has separate `reasoning.effort` values, including `max`. Record the model, reasoning setting, agent count and tools when comparing workflows.

##### Does an eight-point critic score prove accurate water physics?

No. Eight points is a subjective threshold in this prompt. Physical accuracy needs explicit simulation assumptions, numerical checks and repeatable tests. Visual plausibility is only one part of that assessment.

##### Why can a three-round loop stop without passing?

The prompt sets both a quality target and a round limit. The balloon demonstration ended at 6.5, below eight, after exhausting three rounds. Report the failed acceptance state and remaining defects when that happens.

##### How much weekly allowance does an hour of GPT-6 Astra use?

There is no universal percentage. Model selection, task complexity, context, reasoning, tools and caching affect usage. Estimate from the account’s usage dashboard and task records rather than a single demonstration’s duration.

#### Sources

- [OpenAI: GPT-6 Astra model specifications and pricing](https://developers.openai.com/api/docs/models/gpt-6-astra)
- [OpenAI: Model selection and Ultra](https://learn.chatgpt.com/docs/models)
- [OpenAI: Subagent workflows](https://learn.chatgpt.com/docs/agent-configuration/subagents)
- [OpenAI: Desktop app](https://learn.chatgpt.com/docs/app)
- [OpenAI: Usage and allowances](https://learn.chatgpt.com/docs/pricing)
- [OpenAI: Astra release and evaluations](https://openai.com/index/gpt-6-astra/)
- [ARC Prize Foundation: Leaderboard methodology](https://arcprize.org/leaderboard)

#### Author Insight

The most useful number here is the unsuccessful 6.5. It preserves evidence that the ratings improved while the original target remained unmet. A team willing to retain that result can investigate what its agents actually fixed. A team that accepts only confident completion messages may simply pay another agent to produce one.

#### Terminology

- Builder agent: the agent responsible for implementation and revisions.
- Critic agent: a separate agent that inspects artifacts and ranks defects.
- Harness: the environment coordinating the model, tools, state and evaluation.
- Context window: model processing capacity, distinct from the maximum permitted input.
- MCP: Model Context Protocol, used to connect agents with external tools.