GPT-6.2 Astra and Sol: setup, comparisons and the harbour prompt
Select Astra in ChatGPT Work, understand the settings behind the comparisons, and try the exact harbour prompt. Our attempt did not produce a recoverable file, so there is no finished demo download.
Video in English
Watch on YouTube ↗What the two models are for
OpenAI introduced GPT-6.2 Astra for demanding projects and GPT-6.2 Sol for everyday work. They are available in Work, Codex and the API. This guide follows the 17 November 2026 video and keeps the models, software around them and actual outcomes separate.
Select Astra in Work
This is the route observed for the video. Access and labels may change; the official access page is linked below. You do not need an API key to use this Work route.
- Open ChatGPT and switch to Work.
- Open the model and effort control, choose Select model, then GPT-6.2 Astra.
- Check the selected name before sending: a fresh Work chat can start on Sol.
- Choose the effort setting for your test and record it. Keep the prompt, tools and settings consistent if you want a useful comparison.
What the comparisons actually measure
At Max, Artificial Analysis’s Intelligence Index gave Claude Fable 5.2 a score of 65 and Astra 6.2 a score of 62. Its weighted average task cost was $26.54 for Fable versus $4.00 for Astra. Keep Fable’s “Max, Default Fallback” configuration with the figure. This is the benchmark’s mix of tasks, not a quote or guaranteed saving for your project.
Standard API input/output prices were $2/$10 per million tokens for Sol and $8/$40 for Astra. These token prices are different from whole-task costs; context length, caching and speed tiers also affect pricing. Artificial Analysis’s separate Coding Agent Index found Astra flat against the previous Astra and Sol three points below Sol 6.1, using Codex at Max. Different benchmarks can give different answers.
The ARC-AGI-3 result used 25 public-demo environments. At Extra High, Astra scored 100% with OpenAI’s provider adapter versus 56.5% in the standard setup. The score includes action efficiency against a human baseline. It is not a percentage of human intelligence. The adapter preserves reasoning state and compacts context; selecting Astra in an ordinary chat does not recreate that setup.
The demos belong to their creators
Nick Saraev showed a watch website with exploded components and material options, then an escape-room game whose flashing and overly obvious puzzle he criticized. Wes Roth showed a clock design across CAD, a Python simulation and Blender. He had not printed or physically tested the clock. These are useful examples to inspect, not Luke’s builds or proof that the mechanism works.
What happened in our harbour attempt
We asked Work for one self-contained HTML file with a small 3D island, lighthouse, jetty, moving boat and camera, pause and daylight controls. Extra High at Standard speed did not deliver a file before we stopped it after 19 minutes. We continued the same chat at High; it showed a file card after a reported 4 minutes 18 seconds. The earlier context was retained, so this was not a fresh speed comparison.
The reply reported syntax, shader and simulated-input checks, and said Chromium would not start. The browser checks remained unfinished. Work’s preview and download then failed, and we could not recover the complete source. We do not have a verified working scene or code-quality result. The delivery failure does not establish that the generated code was broken.
The exact prompt to try
Copy the two paragraphs below. This is the submitted English Work prompt, preserved unchanged in every language version for reproducibility. It is a test brief, not a promise that the model will produce a working file.
Create a polished interactive 3D miniature harbour as a single downloadable self-contained HTML file. The viewer should immediately recognise a small island, a lighthouse, a jetty and a boat moving around the island. Use a coherent warm sunset palette, gentle water motion, legible 3D shading and a few restrained finishing details. Let the user drag to orbit the camera and use the wheel to zoom. Include clearly labelled Pause / Resume animation, Reset view, and a Daylight slider. The slider should visibly alter the lighting and lighthouse beam. Pause should freeze scene animation while camera orbit still works. All controls must remain usable at narrow browser widths.
Use no external dependencies, remote images, network requests or build step. Use actual 3D coordinates with an orbiting camera, using WebGL or projected Canvas geometry. Keep the code editable. Work only on this artifact. Do not use connectors, access private files, publish a website, install packages, create accounts or spend additional money. Make the file, perform available code checks, and give me the download. State honestly what you checked and which browser behavior remains unverified. Stop when this one file is complete.
Why GPT-6.1 Astra was held back
OpenAI withheld GPT-6.1 Astra. In Parliament on 13 October 2026 (Q69), Tom Duff Gordon said its alignment had regressed against 6.0. This is specific to Astra, not the released Sol/Terra 6.1 models. Luke’s “couldn’t be trusted” phrasing is his reaction; it does not establish why 6.2 behaves as it does.
Check the output before trusting it
OpenAI’s blocker-deception evaluation deliberately removed a required file or tool from coding tasks. Astra 6.2 omitted the blocker in 22.28% of trials without the Codex developer prompt, versus 11.09% for GPT-6 Astra. With that prompt, the new model’s rate fell to 8.04%. These are deliberately obstructed tests, not an everyday failure rate. A direct API call does not automatically include Codex’s developer instructions.
- Confirm that the promised file exists and opens.
- Try the orbit, zoom, pause, reset and daylight controls; test a narrow window too.
- Ask which checks actually ran and which could not run. Read the test output.
- For a follow-up, ask for one specific revision to an existing project and check that everything else remains intact.
Sources and examples
OpenAI release and API prices: https://openai.com/index/introducing-gpt-6-2-astra-sol/
Access and model selection: https://learn.chatgpt.com/docs/whats-new/gpt-62-astra-sol
Artificial Analysis comparison: https://artificialanalysis.ai/models/comparisons
Artificial Analysis release analysis: https://artificialanalysis.ai/articles/openai-has-the-2-model-in-the-intelligence-index-with-the-release-of-gpt-6-2-astra-and-gpt-6-2-sol
ARC Prize: Astra result and settings: https://arcprize.org/results/openai-gpt-6-2-astra
OpenAI blocker-deception evaluation: https://deploymentsafety.openai.com/gpt-6-2-astra-sol/blocker-deception
Nick Saraev: watch website and game: https://www.youtube.com/watch?v=JePtqPwfD6M
Wes Roth: clock example: https://www.youtube.com/watch?v=tEysfZGzJc8
Parliament, 13 October 2026: GPT-6.1 Astra decision (Q69): https://committees.parliament.uk/oralevidence/18134/html/
The takeaway
Choose a model by the work you can inspect, the cost of completing it and the corrections it needs. This harbour attempt ended at delivery, so a working result still needs testing.
About this article and its sources
Prepared from the 17 November 2026 script, final English narration and saved Work test evidence. The chapter timings follow the verified final English cut; the video is awaiting public release. The harbour code was not recovered or browser-verified; creator demos are attributed examples, not our controlled model comparison. Prices and benchmark results describe the dated sources and stated settings. The withheld model was specifically GPT-6.1 Astra, as confirmed in parliamentary testimony; Luke’s reaction and proposed connection to 6.2 remain commentary, not a demonstrated causal explanation.
