GPT-6 Astra Gained 37 Points From Its Harness, Not Its Model.

The most useful number OpenAI published at launch is not on its chart. On ARC-AGI-3, GPT-6 Astra scored 99.9% under OpenAI's own Provider Adapter harness and 62.7% under ARC Prize's standard one. Same model, same tasks, same day.

The difference was scaffolding. And the run that scored 37 points higher was also the cheaper one.

On ARC-AGI-3, Astra scored 99.9% under OpenAI's own Provider Adapter harness and 62.7% under ARC Prize's standard harness. Same model, same tasks. The difference is that the Provider Adapter preserves opaque reasoning state between requests and compacts long contexts, so the model reuses prior work instead of rebuilding it.

The cost detail is the part to internalise: the run that scored 37 points higher was also 3.66x faster, used 49% fewer tokens, and cost $7,281 less. Better scaffolding was not a tax on performance. It was the performance.

Take the operator lesson rather than the AGI argument. A frontier lab found 37 points of capability by changing what the model carries between calls. Nobody has done that work on your workflows. Whatever you are running today is on the equivalent of the standard harness, and the gap between your current results and your achievable results is probably larger than the gap between two model versions.

Concretely, the three things the Provider Adapter did are three things you can do:

Carry state between calls instead of restarting the context each time. Compact rather than truncate when the context fills, so earlier work survives in summary instead of falling off the end. Let the model reuse its own prior reasoning rather than forcing it to re-derive. That is not exotic infrastructure. It is a week of plumbing on most stacks.

The three mechanisms, and why they are yours to copy

Nothing in the Provider Adapter is exotic. It does three things, and all three are available to anyone running an agent on any model.

Carry state between calls instead of rebuilding context each time. Compact rather than truncate when the window fills, so earlier work survives in summary instead of falling off the end. Let the model reuse its own prior reasoning rather than forcing it to re-derive what it already worked out.

That is a week of plumbing on most stacks, not a research programme. And the same three moves are what Anthropic runs in its own production on-call: standing instructions committed to git, a lessons log the agent appends to and re-reads at the start of every investigation, and an orchestrator that hands work to subagents instead of restarting a conversation.

The operator conclusion is uncomfortable in a useful way. A frontier lab found 37 points by changing what the model carries between calls. Nobody has done that work on your workflows. Whatever you run today is on the equivalent of the standard harness, and the gap between your current results and your achievable results is probably larger than the gap between two model versions.

Stop shopping for a better model until you have compacted the one you have.

What ARC Prize actually said

Worth reading the benchmark's own framing rather than the coverage of it. The standard harness is provider-neutral by design, so numbers can be compared across labs; under it a model carries forward only the notes it explicitly chooses to keep, and everything it is thinking stays visible and portable.

The Provider Adapter is a legitimate product capability, shipped in the Responses API, and OpenAI says the changes were not made for this benchmark. Both of those can be true while the comparison is still uneven, because the other models in the chart were not given the same scaffold. ARC Prize was careful about the conclusion, and the caution belongs next to the number: "saturating the benchmark would not represent proof of achieving AGI."

For scale, and this is the part that survives either reading: on the same neutral harness six months earlier, GPT-5.6 Sol scored 7.8% and Claude Opus 5 scored 30.2%. The jump is real whichever rig you use. It is the headline that is not reproducible outside OpenAI's own.

Turning it into topology

The engineering version of this lesson has a name and a set of rules. Treat the agent system as an explicit graph of nodes and data-contract edges rather than a chain of prompts, and three of the rules do the work here:

Keep durable state and pass artifact references, not transcripts. Handing a downstream step a whole conversation is how context rot spreads; handing it a pointer to a result is how state stays bounded.

Put verification on the edge, not at the end. Bad data should not be allowed to travel between nodes, because a long-horizon run compounds an early error for hours before anyone sees it.

Only add cycles with a completion test, a max round count and a token budget. An unbounded loop is not a design, it is a bill, and this model is priced to make that expensive quickly.

One line from that framing is the whole argument compressed: "the model is only one node. The product is the system around it."

Part 2 of Working with GPT-6 Astra: The Operator Guide.