{"slug": "gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test", "title": "GPT-6.2 Astra vs. Sol in Kilo: The Great Pyramid Test", "summary": "OpenAI released GPT-6.2 Astra and GPT-6.2 Sol, priced at $8 per million input tokens and $40 per million output tokens for Astra and $2 and $10 for Sol, with Astra's input and output prices 20% lower than GPT-6 Astra's and a 60% discount on cached input. In a single-run comparison through Kilo, both models built an interactive Three.js, Vite and TypeScript pyramid scene placing 204 of 204 stones across all 8 courses, with Astra producing more detailed visuals and Sol a simpler scene with a clearer crew-status panel. OpenAI claims Sol comes within 0.5 points of Astra on FrontierCode 2.0 at about one-fifth the cost, and both models are available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users and in the API as gpt-6.2-astra and gpt-6.2-sol.", "body_md": "**OpenAI just released [GPT-6.2 Astra and GPT-6.2 Sol](https://openai.com/index/introducing-gpt-6-2-astra-sol/). Astra is the new frontier model for the hardest projects. Sol is the fast, versatile model for everyday work.**\n\nWe wanted to see the difference on a real task, not just on a chart. So we gave both models the same big prompt: *build an interactive 3D scene where workers construct a pyramid, with boats bringing stones along a river.* Then we put the results side by side.\n\nThis post covers what OpenAI announced, what we saw, and how you can run the same test yourself in Kilo Desktop, including on the ChatGPT plan you already pay for.\n\n## **Meet Astra and Sol**\n\nAstra costs [$8 per million input tokens and $40 per million output tokens](https://openai.com/index/introducing-gpt-6-2-astra-sol/). Sol costs $2 and $10. Both are built to get better results on demanding tasks with fewer output tokens, so you get more done within your usage limits.\n\nThese are standard API rates. Higher rates apply above 272K input tokens.\n\n- **Astra’s price dropped.** Its input and output token prices are 20% lower than GPT-6 Astra’s, with a 60% discount on cached input.\n- **Sol’s price held.** It keeps GPT-6.1 Sol’s rates.\n- **Where to find them.** Both are in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and in the API as gpt-6.2-astra and gpt-6.2-sol. Ultrafast is available for both at launch.\n\n## **What OpenAI says they can do**\n\nThe headline claim: Sol gets close to Astra at a fraction of the cost. On FrontierCode 2.0, which checks whether changes to real codebases meet maintainers’ merge standards, Sol comes within 0.5 points of Astra at about one-fifth the cost (both at max).\n\nFor Agents’ Last Exam, OpenAI notes that its Fable 5.2 results leave out nine tasks that used a fallback model.\n\nThe AA-3D result is the one that made us curious. OpenAI says Astra and Sol can turn a written brief into an interactive browser experience, and that Astra builds frontends that look and work better, particularly for websites, gaming and 3D design. A pyramid-building scene is a good way to test that.\n\n## **The test: build a pyramid**\n\nOne of my colleagues, Marius Wichtner, built a pyramid benchmark for fun. It gives each model the same prompt and shows the finished apps side by side.\n\nThe prompt asks for a polished, interactive 3D scene of ancient Egyptian workers building a pyramid, with boats delivering materials along a river. Workers have to coordinate heavy loads. Boats have to dock, unload and keep materials flowing. The model picks the tech, the style and the building method.\n\nWe ran it with GPT-6.2 Astra and GPT-6.2 Sol, both through Kilo. Both models chose Three.js, Vite and TypeScript, and both built a pyramid of 204 stones.\n\nThis is one run per model, so treat it as a first vibe look.\n\n## **Astra looks like a film set. Sol looks like a strategy game.**\n\nBoth models finished the job: 204 of 204 stones placed, all 8 courses built. The difference is in how they got there.\n\n### **Astra pyramid building in progress**\n\n### **Sol pyramid building in progress**\n\nOn looks alone, Astra wins clearly. Up close, the sledge, the ropes and the dock planks hold up. Sol’s scene is simpler, but its panel tells you exactly what every crew is doing, which makes the work easy to follow.\n\nThat fits what OpenAI says about the two models: Astra for the work where quality matters most, Sol for everyday use. It is one run each, though. Run it a few times on your own tasks before you pick.\n\n## **Run the test yourself in Kilo Desktop**\n\n[Kilo Desktop is now generally available](https://blog.kilo.ai/p/kilo-desktop-is-generally-available), and it is a good place for this kind of comparison. Every task gets its own session, and each session can run in its own git worktree, so two models never fight over the same files.\n\n1. [Download Kilo Desktop](https://kilo.ai/install) and open an empty project.\n2. Start a session and give it its own git worktree. Pick GPT-6.2 Astra.\n3. Start a second session in a new worktree. Pick GPT-6.2 Sol.\n4. Paste the prompt below into both.\n5. When they finish, open each result in the built-in browser and compare.\n\n**Here is the prompt:**\n\n```\nBuild a polished, interactive 3D scene that simulates ancient Egyptian workers constructing a pyramid, supported by boats delivering materials along a river.\n\nChoose the technical approach, visual style, layout, construction methods, and animation techniques yourself. Focus on meeting the requirements below rather than following a prescribed design.\n\nCore Requirements\n\n- Show construction from the beginning through to a finished, explorable pyramid.\n- Workers must perform meaningful work and coordinate intelligently. Heavy loads should be handled collectively, with believable assignments and movement.\n- Boats should travel convincingly, stop at suitable docking or anchor points, unload, and support a functional flow of materials to the builders.\n- Include different useful wares, not only stones. Their delivery and use should matter to construction.\n- Build the entrance, tunnels, and burial spaces as part of the simulation. Workers must construct them, rather than simply revealing a prebuilt interior.\n- The finished pyramid should function as a believable virtual tomb with a connected, walkable interior.\n- Draw inspiration from plausible ancient Egyptian engineering. Use judgment where historical methods are uncertain.\n\nExperience and Quality\n\n- Make the building, hauling, unloading, and underground work animations convincing, including when viewed up close.\n- Different wares should have appropriate handling for their material, weight, and purpose. Avoid generic or cheap-looking motion, floating loads, unexplained transformations, and obvious clipping.\n- Provide transparency and first-person buttons that remain available during construction.\n- Transparency should make internal and underground work easy to inspect.\n- First-person mode should offer a pleasant, responsive walking experience around accessible areas and through the finished interior.\n- Keep the ordinary HUD minimal, retaining compact speed, reset, transparency, and first-person controls.\n- Make underground information and navigation attractive and useful without cluttering the view.\n- Ensure stable geometry and rendering, particularly inside the pyramid. Avoid flickering walls, visible gaps, and clunky movement.\n- Use free tools and assets without requiring paid services.\n\nWorkflow and Delivery\n\n- Build from scratch and stay in the current worktree.\n- Implement and test the solution yourself, including the complete construction flow and exploration.\n- Keep testing and iteration fast. Avoid endless analysis, repeated long benchmarks, or waiting through construction in real time for every check.\n- Finish with a working application, start it, open the preview, and briefly report the results and any remaining limitations.\n```\n\n**A few tips for a fair comparison:**\n\n- Use the same Kilo Desktop version and settings for both models.\n- Run each model at least three times. One run can flatter or hurt either side.\n- Write down time and plan usage per run, so you can compare cost as well as looks.\n- Check the prompt’s own requirements: boats docking, workers hauling together, a walkable interior, working transparency and first-person buttons.\n\n## **Bring your ChatGPT plan to Kilo**\n\nYou don’t need an API key to try these models. With [Sign in with ChatGPT](https://blog.kilo.ai/p/sign-in-with-chatgpt), you pick “Continue with ChatGPT” in Kilo, and your requests draw on the Codex and ChatGPT Work usage included in your plan instead of per-token API billing.\n\n- **Works across Kilo.** Desktop, VS Code, JetBrains, the CLI, Cloud Agents, Code Reviews and the mobile app.\n- **You stay in control.** Per OpenAI, connecting an app adds no new allowance, and you can cap how much of your weekly plan Kilo uses in ChatGPT settings under Usage.\n- **Your chats stay yours.** Using your plan does not give Kilo access to your ChatGPT conversations or memories.\n- **Switch freely.** Use your plan's[OpenAI models](https://kilo.ai/models/by/openai) next to the 500+ models in the Kilo Gateway, from the same model picker.\n\n### **For teams**\n\nKilo now supports [ChatGPT Business plans](https://blog.kilo.ai/p/announcing-business-plan-support) too. Admins enable “Use ChatGPT plan” for Kilo under Tenant settings, then External access in the OpenAI admin console, and can set a workspace-wide limit on how much of each member’s included allowance approved services can use. Usage shows up in the OpenAI dashboard like any other connected app.\n\nOpenAI’s [Sign in with ChatGPT docs](https://learn.chatgpt.com/docs/sign-in-with-chatgpt) list eligible plans (Go, Plus, Pro and Business) and the full admin steps.\n\n## **Which one should you use?**\n\nStart with Sol for everyday work. It costs a quarter of Astra per token, and OpenAI’s own FrontierCode result puts it within 0.5 points of Astra at about one-fifth the cost. Reach for Astra when quality matters most: 3D and visual work, long projects, and the hardest problems. In our pyramid test, that is where it pulled ahead.\n\nThe best answer depends on your tasks, and Kilo makes it easy to find out. Run both on the same job, compare the diffs, and keep the one you would merge. If you’d rather not choose, [Auto Efficient](https://blog.kilo.ai/p/auto-efficient) picks the least expensive model proven capable of the task.\n\n[Download Kilo Desktop](https://kilo.ai/install), connect your ChatGPT plan, and try the new Astra and Sol today. Then tell us what you build.", "url": "https://wpnews.pro/news/gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test", "canonical_source": "https://blog.kilo.ai/p/gpt-62-astra-vs-sol-in-kilo-the-great-pyramid", "published_at": "2026-11-19 15:07:40+00:00", "updated_at": "2026-11-19 15:14:20.264234+00:00", "lang": "en", "topics": ["large-language-models", "ai-products", "generative-ai", "ai-tools", "developer-tools"], "entities": ["OpenAI", "GPT-6.2 Astra", "GPT-6.2 Sol", "Kilo", "Kilo Desktop", "Marius Wichtner", "FrontierCode 2.0", "Three.js"], "also_reported_by": [], "takeaway": "In a single Kilo run, GPT-6.2 Astra and GPT-6.2 Sol both completed the same Three.js pyramid scene at 204 of 204 stones, with Astra producing more detailed visuals and Sol a simpler scene with a clearer crew-status panel.", "recommendation": {"advice": "Run the same prompt several times per model on your own tasks before choosing between Astra and Sol; this comparison used only one run each.", "playbook": "evals-testing", "playbook_url": "https://wpnews.pro/playbooks/evals-testing"}, "alternates": {"html": "https://wpnews.pro/news/gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test", "markdown": "https://wpnews.pro/news/gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test.md", "text": "https://wpnews.pro/news/gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test.txt", "jsonld": "https://wpnews.pro/news/gpt-6-2-astra-vs-sol-in-kilo-the-great-pyramid-test.jsonld"}}