The Short Answer
GPT-6 Astra's headline 98.6% on ARC-AGI-3 came from a harness OpenAI built for it; on the standard harness ARC Prize calls "apples-to-apples," it scored 62.7% against Claude Opus 5's 30.2%. Both numbers are real. When the comparison surface is vendor-shaped, the only durable answer is architecture that treats models as swappable — with ibl.ai you own all the code and the data, model-agnostic across any LLM.
Astra is a genuine leap. The point of this post is not that the number is fake — it is that there are three numbers, they differ by 36 points, and which one you quote depends on whose harness ran the test.
What did GPT-6 Astra actually score on ARC-AGI-3?
Three different figures, all published by ARC Prize for the same model on the Semi-Private set:
| Harness | Reasoning effort | Score | Cost of run |
|---|---|---|---|
| Standard (cross-provider) | — | 62.7% | $26,098 |
| Provider Adapter | High | 99.9% | $18,817 |
| Provider Adapter | Max | 98.6% | $17,332 |
The 98.6% that circulated is the third row.
Note the cost column, because it inverts the usual intuition: the adapter run scored far higher and cost roughly $8,000 less, and ARC Prize records the adapter runs as approximately 3.66x faster by aggregate elapsed time.
Efficiency and score moved together, which is what makes the harness difference substantive rather than cosmetic.
What is the difference between the standard harness and a provider adapter?
It is a difference in what the model is allowed to carry between steps.
The standard harness requires a model to carry forward whatever notes it chooses to keep as it moves through the environment, under a neutral interface that every provider meets on the same terms.
The provider adapter preserves opaque reasoning state between requests and uses compaction for longer conversations, so the model can reuse prior work rather than reconstructing it.
That is not cheating — it is the model using infrastructure its provider built for it, and it reflects how the model would actually be deployed.
ARC Prize is explicit that these are different evaluation conditions, and that the standard harness is the one providing "an apples-to-apples comparison across providers."
Both numbers describe something true. They just answer different questions: how good is this model in a neutral rig versus how good is this model with its vendor's own scaffolding.
How does Astra compare to other frontier models on the same benchmark?
On the standard harness, which is the like-for-like column:
| Model | ARC-AGI-3 |
|---|---|
| GPT-6 Astra | 62.7% |
| Claude Opus 5 | 30.2% |
| GPT-5.6 Sol | 7.8% |
| Claude Opus 4.8 | 1.5% |
| GPT-5.6 Terra | 0.8% |
| Gemini 3.1 Pro | 0.4% |
Two things are true at once, and enterprise readers need both.
Astra genuinely leads: 62.7% is more than double the next model, and the jump from GPT-5.6 Sol's 7.8% is enormous for six months. Anyone dismissing this as benchmark theater is not reading the standard-harness column either.
And the headline overstated the margin. Comparing Astra's adapter-harness 98.6% to Opus 5's 30.2% puts two different evaluation conditions in the same sentence. The real like-for-like gap is large; it is not near-saturation against a field stuck in the twenties.
Why does a benchmark footnote matter for enterprise architecture?
Because it means you cannot fully outsource model selection to a leaderboard.
If the scaffolding a vendor supplies can move its own score by 36 points, then published comparisons are partly a measurement of vendor tooling, not only of model capability.
That is not an accusation — ARC Prize published all three numbers precisely so the distinction is visible. But it does change what a procurement team can conclude from a headline.
The only reliable way to know which model is best for your workload is to run the candidates against your own data, your own prompts and your own evaluation set. That is a straightforward exercise — if switching models is cheap.
It usually is not.
Most enterprise AI deployments weld authentication, data pipelines, prompt engineering, safety guardrails and compliance documentation to one provider's API.
Under that architecture every re-evaluation is a rebuild, so the re-evaluation never happens, and the lock-in becomes self-enforcing: you stay because leaving is expensive, not because the model is still the best one.
We wrote about this pattern when GPT-5.6 shipped and model routing became the enterprise question; Astra is the same lesson at a larger magnitude.
What does GPT-6 Astra cost, and what else shipped with it?
Astra was released on September 3, 2026, priced at $10 per million input tokens and $50 per million output tokens, with cached input at $1 and cache writes at $12.50.
It also arrived with a classification worth noting: it is the first OpenAI model to cross the "Critical" threshold for cybersecurity under the company's Preparedness Framework, meaning it can identify and develop working exploits against hardened systems without step-by-step human direction.
Exploit-creation capability is gated behind a limited-access program. In OpenAI's own cyber-safety testing the model refused 91.5% of requests that should not receive cyber assistance, against 59% for GPT-5.6 Sol, as reported by CSO Online.
For a regulated institution, that is a second argument for the same architecture.
A model whose capability profile now includes autonomous exploit development is a model your security and compliance teams will want to reason about specifically — including the ability to route certain workloads away from it, or to keep sensitive work on a self-hosted open-weight model entirely.
Routing is only an option if the platform is model-agnostic to begin with.
What does model-agnostic architecture actually require?
Four things, and none of them is the model itself.
- A unified API layer that abstracts provider-specific integration, so swapping a model does not change application code.
- A context layer that connects agents to institutional systems — SIS, HRIS, CRM, ERP, EHR — independently of which model reasons over them.
- Memory and evaluation that persist across model changes, so an upgrade does not reset your accumulated context or invalidate your benchmark set.
- Routing, so sensitive queries can go to a self-hosted open-weight model while general queries go wherever is cheapest that week.
With ibl.ai, you own all the code and the data — the platform is deployed on your own infrastructure with full source code access, model-agnostic across any LLM, usage-based with no per-seat pricing, and deployable anywhere from your own cloud to on-premise, GovCloud or a fully air-gapped network.
The practical test is simple.
When the next frontier model ships — and on the current cadence that is weeks, not quarters — can you evaluate it against your own workload and switch if it wins, without a migration project?
If the answer is no, you are not choosing your model. Your first integration decision is still choosing it for you.
ibl.ai is family-owned and operated from New York, NY. For institutions weighing who will still be standing behind a deployment several model generations from now, that is not a small consideration.
Related reading: why vendor lock-in is the real risk in model-agnostic AI, and what published inference benchmarks reveal about the cost floor.
Sources: ARC-AGI-3 scores, harness definitions and run costs from ARC Prize's Astra report; pricing and release details from OpenAI; cyber-safety classification via CSO Online and The New Stack's analysis of the harness caveat.