62.7%.
That’s GPT-6 Astra on ARC-AGI-3 with ARC Prize’s standard harness — about $26,000 for the run. Same model, same week, Provider Adapter harness: 99.9% for about $19,000.
That’s not a rounding error. That’s a 37-point gap on the same benchmark, depending on how you wrap the model.
OpenAI’s launch week is full of “AGI era,” “most intelligent and aligned,” and flashy computer-use demos (Blender, eBay, legal drafts, ordering food in one flow). Fine. The demo is real enough. The leaderboard asterisk is realer.
Two scores. One model.
ARC Prize is clear about what they’re measuring:
| Harness | Best observed (Semi-Private) | Rough cost |
|---|---|---|
| Standard (provider-neutral; model keeps notes it chooses) | 62.7% (Astra max) | ~$26K |
| Provider Adapter (keeps opaque reasoning state + compaction) | 99.9% (Astra high) | ~$19K |
Both are state-of-the-art under their own rules. The mistake is pretending those two numbers mean the same thing when you screenshot a chart.
OpenAI’s own “~3×” claim is the retained-reasoning + compaction lift — a different cut of the same story than the 37-point gap between harnesses above. Don’t mix the metrics.
ARC also notes the adapter runs were ~3.66× faster and used ~49% fewer tokens on the tasks both harnesses solved. So the “better” score isn’t just higher — it’s cheaper and faster because the wrapper remembers work.
Plain English: the wrapper moved the needle
OpenAI has already written the quiet part out loud. On ARC-AGI-3, turning on two production settings — retained reasoning and compaction — is what triples scores and cuts output tokens hard. That’s not “the model woke up.” That’s context management.
- Standard harness: closer to a fair, provider-neutral stress test. The model has to re-derive plans more often.
- Provider Adapter: closer to how ChatGPT / Codex actually run — keep hidden reasoning across turns, compress old history instead of deleting it.
If your takeaway from launch week is “Astra is 99.9% AGI,” you read the marketing deck. If your takeaway is “my agent stack is leaving money on the table if the harness is dumb,” you read the eval.
What to do Monday (before you switch stacks)
- Ask which harness any score used. Standard vs adapter. Notes vs retained state. If they won’t say, treat the number as theater.
- Stop comparing $/token in isolation. Astra API sits in the same band as Claude Fable 5.1 ($10 / $50 per 1M in standard mode, per public rollout coverage). Care about $/task and whether your wrapper retries, truncates, or keeps state.
- Audit your own agent loop before you chase the new model: Do you keep reasoning across tool calls? Do you compact, or do you wipe context every turn? A mid model with a good harness beats a frontier model with a goldfish memory.
- Treat computer-use demos as product demos, not proofs. “Anything you can do on a computer, Astra can do for you” is a pitch. Your CRM, your inbox, your permissions — that’s the real eval.
The AGI demo is real. The leaderboard asterisk is too.
You can believe Astra is a jump at desktop agents and refuse to launder a 99.9% number that depends on a provider-specific harness. Both can be true. Adults hold both.
This week’s feed wants you to pick a tribe: AGI believers vs doomers vs open-weights cope. Skip the tribe. Ship a checklist.
If you only remember one line: the model didn’t 3×. The harness did. Measure yours.
Sources
- ARC Prize — OpenAI’s GPT-6 Astra on ARC-AGI-3
- ARC Prize — GPT-6 Astra results
- OpenAI — how two settings tripled ARC-AGI-3 scores
- Launch context: The Verge, CNBC
Want the one-page “how to read a frontier launch” checklist? Reply HARNESS / join the waitlist.