OpenAI Dropped GPT-6 Astra. Here's What the Benchmarks Actually Say (and Don't Say)
OpenAI released GPT-6 Astra on September 3, 2026 — the first full-number release since GPT-5. It hits 99.9% on ARC-AGI-3 with the Provider Adapter harness, improves token efficiency by 70% over GPT-5.6 Sol, and costs $10/$50 per million tokens. But the story is more nuanced than a headline score.
This article covers the benchmarks, pricing changes, architecture hints, community skepticism, and where Astra actually lands relative to Fable 5.1 and GPT-5.6 Sol.
What GPT-6 Astra Costs
The pricing change is the first thing you need to know because it changes the value calculus.
| Metric | GPT-6 Astra | GPT-5.6 Sol | Delta |
|---|---|---|---|
| Input | $10/M | $4/M | 2.5x |
| Output | $50/M | $20/M | 2.5x |
| Cache read | 90% discount | 90% discount | Same |
| Cache write | 25% premium | 25% premium | Same |
Astra costs 2.5x more than GPT-5.6 Sol across the board. That's a steep increase for a model that doesn't clearly dominate on every axis. The same cache discount structure applies — if your workload is cache-heavy, the effective price delta narrows, but the raw per-token cost is significantly higher.
Sources: Artificial Analysis benchmarking article, confirmed by HN discussion (1340 pts).
ARC-AGI-3: The 99.9% Headline Needs Context
The ARC Prize blog published results showing GPT-6 Astra scoring 99.9% for $19K on ARC-AGI-3 Semi-Private using a Provider Adapter harness, and 62.7% for $26K using the Standard harness.
The harness difference matters:
- Standard harness — the model carries forward notes it chooses to keep through the environment.
- Provider Adapter harness — preserves opaque reasoning state between requests and uses compaction for longer conversations, allowing the model to reuse prior work.
The Provider Adapter harness gives Astra a massive advantage by preserving reasoning context across calls. Without it, the score drops to 62.7% — still impressive, but not the 99.9% number in headlines.
Astra also surpassed the human baseline in action efficiency — it used fewer actions to solve tasks than the median human participant. That's a genuine milestone regardless of harness choice.
Source: ARC Prize blog — Greg Kamradt, Sep 3, 2026.
HN commenters pointed out the scorecard is potentially misleading: GPT-5.6 Sol's 7.8% shown on the chart would jump to ~30% with the same Provider Adapter harness, making the comparison less dramatic. Source: HN comment on scorecard comparison.
Coding and Agent Benchmarks
Artificial Analysis Coding Agent Index
GPT-6 Astra scores 67 in the Coding Agent Index — approximately equal to Claude Opus 5 and Fable 5 in Claude Code. Fable 5.1 leads at 70. On token efficiency, Astra uses one third the tokens of GPT-5.6 Sol (max) in the Codex harness and one fifth of Claude Opus 5 (xhigh).
This is Astra's strongest selling point for developers: equal coding agent capability at lower token consumption than its predecessor, even if it doesn't lead the leaderboard.
Source: Artificial Analysis Coding Agent Index benchmarks.
SRE-Bench: Binary Reverse Engineering
OpenAI tested Astra on SRE-Bench, which measures whether models can reverse engineer software binaries without source code access. Results:
| Attempts | GPT-6 Astra | GPT-5.6 Sol |
|---|---|---|
| Single attempt | 88.0% | 55.9% |
| Within 4 attempts | 99.2% | 68.7% |
Astra nearly doubles Sol's single-attempt success rate. This is relevant for security analysis, legacy system migration, and any scenario where source code is unavailable.
Source: HN comment citing OpenAI's system card, referencing SRE-Bench paper.
Intelligence Index and Other Evaluations
Artificial Analysis found distinct stories across their two flagship indices:
- Intelligence Index: Astra uses fewer tokens than GPT-5.6 Sol for similar performance, but this is "outweighed by higher prices" — the 2.5x cost increase negates the token efficiency gain.
- Analytical Quality Elo: Significant increase in both rubric scores and Analytical Quality in AA-Briefcase.
- Presentation Quality Elo: GPT-5.6 Sol (max) still leads all models — Astra regressed here.
- Humanity's Last Exam: 6-point gain, a long-standing evaluation of general knowledge.
Source: Artificial Analysis article.
Architecture: What We Know
OpenAI has not released detailed architecture specs, but the system card is available at deploymentsafety.openai.com/gpt-6-astra. Key signals from community analysis:
- Recurrent/looped transformer architecture — multiple HN commenters noted this structural similarity to latent reasoning approaches. "Recurrent/looped transformers themselves aren't a new concept, but it's interesting to finally see this approach show up in a frontier production model." Source: HN comment.
- Always-on reasoning — inferred from the 2.5x pricing and the model's ability to carry reasoning context across calls (necessary for ARC-AGI-3 performance).
-
No forced tool use — following the same pattern as Fable 5.1, forced tool calls (
tool_choice: {"type": "any"}) likely return errors because they would skip reasoning and degrade output quality.
Community Reaction and Open Questions
The HN thread (1340 points, 1070 comments as of writing) covers three main areas of concern:
1. Is this AGI or another point release? Francois Chollet's 2019 paper "On the Measure of Intelligence" was cited by multiple commenters who argued that frontier-model progress still looks like "skill acquisition optimization" — broader coverage and higher scores within existing benchmark distributions, rather than genuinely novel generalization capability. Source: HN comment.
2. The pricing-competitiveness gap. Astra costs 2.5x Sol but delivers modest gains on most benchmarks except ARC-AGI and SRE-Bench. On the Coding Agent Index, it ties models that cost less. One commenter's take: "Every other benchmark seems to be a relatively modest improvement, comparable with any of the 'point' updates from AI labs." Source: HN comment.
3. Improved prompting and collaboration. Some users reported Astra handles ambiguous prompts better — it asks clarifying questions rather than one-shotting assumptions. "If I give a poorly constrained/ambiguous prompt, I don't want the model one-shotting assumptions left and right." Source: HN comment.
Limitations Worth Calling Out
- ARC-AGI-3 99.9% is harness-dependent. Without the Provider Adapter, it drops to 62.7%. The comparison to other models needs the same harness applied to be fair.
- Coding Agent Index doesn't lead. Score of 67 ties existing models; Fable 5.1 leads at 70. Better token efficiency is real, but raw capability parity isn't a breakthrough.
- Pricing increase is steep. 2.5x across the board with modest benchmark gains is a hard sell for cost-conscious teams.
- Presentation quality regressed. GPT-5.6 Sol (max) still leads on this axis.
- Architecture details are sparse. No paper, no parameter count, no training cost disclosure. The system card covers safety, not specs.
- "AGI era" framing is marketing. OpenAI's blog and CEO statements claim "welcome to the AGI era," but the benchmarks show incremental gains, not a discontinuous leap. Skillful PR timing matters, but it doesn't change the data.
Should You Switch?
For ARC-AGI-3 style reasoning tasks and SRE-Bench binary analysis: Astra is the clear leader. The 99.2% within-four-attempts on SRE-Bench is genuinely hard to argue with.
For coding agent workloads: Benchmark it against what you currently use. It ties Fable 5 and Opus 5 on the Coding Agent Index but costs more. Token efficiency is better, but at 2.5x pricing, the per-task cost may be higher or comparable depending on your context length.
For high-throughput, cost-sensitive tasks: Stay on GPT-5.6 Sol or Fable 5.1. The 2.5x premium doesn't buy you proportional gains on general-purpose workloads.
For the "is this AGI" question: The data doesn't support a discontinuity claim. Better ARC-AGI scores, better binary reverse engineering, better token efficiency — these are real improvements on a continued trajectory, not a regime change. The ARC-AGI-3 benchmark specifically measures generalization to novel tasks, and Astra's performance there (especially with the harness that preserves reasoning context) is genuinely impressive. But the rest of the benchmark suite shows incremental improvement consistent with previous point releases.
Run your own evals on your own data. The headline scores are directional, not dispositive, and the pricing change means the cost calculation is different for every workload.
Sources: Artificial Analysis — Benchmarking GPT-6 Astra, ARC Prize blog — GPT-6 Astra on ARC-AGI-3, HN discussion (1340 pts), OpenAI system card, SRE-Bench paper (arXiv). ARC-AGI-3 scores from ARC Prize Standard and Provider Adapter harnesses. Pricing from Artificial Analysis. Community quotes from HN comments with references provided inline.
Top comments (0)