OpenAI’s GPT-6.2 announcement is as much a demo reel as a benchmark PDF. GPT-6.2 Astra builds playable browser games, turns architecture photos into CAD and physical lamps, and refreshes investor slides from 10-Q filings. GPT-6.2 Sol is the model OpenAI employees reportedly default to for mixed technical and non-technical work.
This list collects real artifacts OpenAI and partners cited on November 17, 2026, what each proves for builders, and how to stress-test similar workflows without treating marketing clips as production guarantees. Start with the launch pricing post and comparison tables when you pick Astra versus Sol.
Fan of demo classes from the GPT-6.2 launch post, from games to legal agents
1. Meridian Sound Racing — playable futuristic web game
Browser computer-use style interaction for GPT-6.2 generated 3D web demos
OpenAI links Meridian Sound Racing, a browser racing game GPT-6.2 Astra generated on a winding city track. It is the clearest end-user artifact in the launch: not a screenshot, but a URL you can load and drive.
What it shows: combined 3D layout, game logic, and front-end polish in one trajectory — the same lane computer-use agents struggled with when Astra was new.
Try something similar: ask Astra for a single-scene canvas game with keyboard controls, then measure output tokens and human bug-fix minutes versus GPT-6.2 Sol on the same prompt.
2. Marina City paper lamp — photo to CAD to physical build
OpenAI documented a Chicago Marina City balcony motif turned into a lamp: reference photos → CAD revisions → hand assembly, plus a site visitors can use to design their own paper lamp.
What it shows: visual reasoning + CAD + documentation, not just mesh pretty pictures. It aligns with embodied design agents themes without a robot arm.
Caution: physical fabrication needs safety review; treat STL exports as prototypes.
3. Chassis dynamics on a mountain road — Python plus Blender
GPT-6.2 Astra built a 3D car model, wrote a Python solver for tire and suspension dynamics, and animated results in Blender from a creative brief.
What it shows: multistep simulation + visualization beyond static meshes — useful for mechanical teams comparing against NVIDIA science-agent posts.
4. Earnings deck refresh from 10-Q and Q2 release
Side-by-side slides compare GPT-6 Astra versus 6.2 Astra updating a retailer earnings deck. OpenAI highlights headline takeaways, channel split, and management next steps with less filler.
What it shows: template-aware finance comms — closer to Semianalysis GDPx agent failure analysis workloads than coding leaderboards.
Builder takeaway: feed your prior deck PDF as the style anchor; log how many bullet edits remain.
5. Alderfen Field Systems valuation workbook
A fictional ag-tech case asks for a reconciled three-statement model, per-share DCF, and sensitivity tables. OpenAI contrasts layout choices: 6.2 Astra keeps bridges beside sensitivities; 6 Astra scatters bridges.
What it shows: spreadsheet cognition (segment drivers, EV-to-equity bridge) — the same class of work AA-AnalystAgent scores on Medicaid transport spend.
6. RC buggy assembly guide from public manuals
From photos and manuals, Astra produced component models, an animated assembly guide, and an illustrated parts catalog.
What it shows: technical documentation automation for hardware teams — pair with Blender AI MCP workflows if you already script Blender.
7. Harvey legal agent bench — memos and concessions
Niko Grupen (Harvey) reported 3× Legal Agent Benchmark throughput versus GPT-6 Astra with 34.5% fewer output tokens, plus fewer unwarranted concessions in filings.
What it shows: legal agent depth, not chat Q&A — compare against Claude Opus legal workflows on your matter types.
8. Legora research across tax, M&A, litigation
Jacob Lauritzen (Legora) said Astra won 8 of 10 research tasks versus alternatives, with 62% fewer tokens than GPT-6.1 Sol on complex legal-agent evals.
What it shows: multi-jurisdiction research routing — useful if you already benchmark remote labor indices.
9. ClickUp Brain — apps, dashboards, slides inside workspace
Jay Hack (ClickUp) described blind tests where 6.2 Astra beat production models with ~40% fewer tokens and ~30% fewer review rounds on in-workspace artifacts.
What it shows: embedded agent UX matters; token savings only count if review rounds actually fall in your workspace integration.
10. Higgsfield Miracle — real-time video editing
Alex Mashrabov (Higgsfield) credited 6.2 Astra with enabling Miracle real-time video editing and 73% one-shot success on internal 3D evals versus three or more iterations on GPT-6 Astra.
What it shows: generative media control for studios — adjacent to Kling and image arena competition but video-first.
11. Basis accounting — 50% faster long-horizon work
Eros Wang (Basis) reported 50% average and up to 73% time drop on long-horizon accounting trajectories, citing better Bash planning.
What it shows: ops finance automation — echoing Revolut Codex fincrime stories but on accounting close tasks.
12. Jump Trading hardware optimization
Adam Alnasser (Jump Trading) described Astra saturating a three-year-old benchmark in under a day on complex optimization — treat as anecdote, not a verified public score.
What it shows: quant / HPC scripting interest in frontier models beyond SWE-bench.
What developers are saying about the demos
HN users montenegrohugo (50535307) praised 3D site performance on Linux without GPU acceleration; matt3210 (50535350) noted broken graphs on the marketing page. Both are reminders to judge your hosted artifact, not OpenAI’s landing-page JS.
Mapping demos to API effort settings
OpenAI’s post mixes max reasoning scores with everyday Sol anecdotes. When you reproduce a demo, declare which SKU and which reasoning tier you used. A Meridian Sound Racing-scale web build is an Astra job; a LifeVal-style kindergarten shortlist is a Sol job. Mixing them in one trace and blaming the model for “over-engineering” usually means you reached for Astra on a Sol-shaped prompt.
Document tool access too: computer-use demos assume browser and filesystem affordances you may not expose in a stripped API integration. Start with read-only web fetch, then add write tools once you have approval logs.
Failure modes HN already flagged
Beyond partner success stories, the launch thread surfaced UI broken charts and heavy GPU use on the marketing site — reminders that landing-page polish ≠ model quality. Separate product marketing from your staging eval. If graphs fail to render in OpenAI’s post, your internal Grafana for token usage matters more than their front-end WebGL.
How to build your own shortlist
Reproduce three demo classes this week:
- Slide refresh from a real PDF template.
- Single-file web game with input handling.
- Spreadsheet model with explicit reconciliation checks.
Log results beside GPT-6.1 Sol runs.
Checklist: publishing demo results responsibly
When you tweet or present internal replications of OpenAI’s launch demos, include:
- Model ID and date (
gpt-6.2-astraon 2026-11-17 or later). - Tools enabled (browser, Blender export, spreadsheet engine).
- Human edit time after the first model pass.
- Failure screenshot when the demo breaks — HN’s honesty about broken charts is a good standard.
Investor decks should not cite IMC trading briefs or Jump Trading optimization quotes as your own results. Attribute them to OpenAI’s partner section or rerun the workflow under NDA with your quant team.
Demos and cyber guardrails
Several launch demos flirt with cyber-capable trajectories (secured code, vulnerability discussion). Consumer API keys run with tighter refusals than the charts with classifiers disabled. If your demo is “find bugs in this binary,” expect friction until Daybreak hosts 6.2. Keep offensive demos on approved ranges described in Vals cyber offense coverage, not on a default production key.
Workshop tie-in: reproducing one demo in a lunch session
Teams testing 6.2 should pick one public artifact — the earnings deck refresh or the USB-C shopping flow is too Google-specific; the spreadsheet reconciliation or single-scene game maps cleanly to a 45-minute working session. Bring a anonymized template file, run 6.2 Sol first for cost, escalate to Astra only if Sol fails acceptance checks, and record token counts in a shared spreadsheet. That exercise produces more useful routing data than debating whether OpenAI’s marketing graphs use linear axes. Point interested teammates at explainx.ai’s live AI workshops if they need guided agent evals rather than ad-hoc lunch experiments. Save HN thread links in your runbook so engineers can read skepticism alongside vendor demos. If 6.2 Sol wins on tokens with equal quality, route default traffic there; reserve 6.2 Astra for the jobs in sections 2–6 above.
Partner quotes: how to treat them in your eval plan
Harvey, Legora, ClickUp, Higgsfield, Basis, and Ramp supplied launch-day quotes with token deltas and blind-test anecdotes. Those are useful for hypothesis generation — which vertical might see the largest savings — but they are not substitutes for your repo or your slide templates. Copy their workflow shape (legal research loops, in-workspace artifact review, studio video pipelines) into a three-case internal benchmark before you cite their percentages in a leadership deck.
If a partner names 34.5% fewer tokens or 40% fewer review rounds, demand the same metric definition from your telemetry: count human edit minutes, not just model turns.
