GPT-6.2 Astra is OpenAI's flagship reasoning/agentic model (API: gpt-6.2-astra), a proprietary autoregressive text/agent model with a 1,050,000-token context window, 128K max output, five selectable reasoning-effort levels, and frontier results across Terminal-Bench 4.0, FrontierMath, HLE-Diamond, Harvey LAB-AA and AA-3D at $8/$40 per 1M tokens.
GPT-6.2 Astra is the flagship of the GPT-6.2 series, announced by OpenAI on 9 September 2026 alongside the smaller GPT-6.2 Sol, with third-party trackers re-indexing it on 17 November 2026. It is positioned explicitly for "real-world engineering, professional work and scientific research," i.e. long-horizon agentic execution rather than chat throughput, and it is marketed as roughly 20% cheaper per token than its direct predecessor GPT-6 Astra ($8/$40 per 1M input/output vs. $10/$50). Verified operational envelope: 1,050,000-token context window, 128,000-token maximum completion, cache reads at $0.40/1M and cache writes at $10/1M, web-search tooling at $10 per 1,000 calls on OpenRouter, five reasoning-effort settings (low, medium, high, xhigh, max), and a published knowledge cutoff of 18 May 2026.
On capability, GPT-6.2 Astra posts SOTA-class numbers on the coding/agentic and scientific-research axes: 64.1% on Terminal-Bench 4.0, 45.4% on FrontierCode 2.0, 78.1% on Terminal-Bench Science, 71.5% on CritPt v2 Hard, 72.2% on HLE-Diamond, 100% on FrontierMath Tier 4 and 92.6% on FrontierMath Tiers 1-3. It is equally strong on professional/legal work (Harvey LAB-AA 93.8%, though the adversarial Vals HLAB all-pass rate is only 15.0%, showing that all-or-nothing multi-document legal tasks remain a hard failure mode) and on human-preference Elo evaluations, where it leads AA-3D at 1318 Elo (Claude Fable 5.2 at 1271, GPT-6.2 Sol at 1259) and the AA-Robotics v0.1 preview at 79. Artificial Analysis places it second overall in the Intelligence Index at 62.1 (62 at max and xhigh effort, 60 at high effort), roughly 5 points above GPT-6 Astra, but still behind Claude Fable 5.2 on factual-knowledge measures (AA-Omniscience v1.1) and on HLE-Diamond per AA's own commentary — its AA-Omniscience Index of 50.0 against an AA Cyber Index of 37.3 and 0.0% on CyberGym shows that broad parametric recall and real-world exploit-loop autonomy are not uniform strengths.
Architecturally, OpenAI has not published a technical report, parameter counts, MoE routing details, or attention variants (MLA/GQA) for this model. It should therefore be treated as a proprietary dense-or-MoE autoregressive transformer with an undisclosed parameter scale; only its serving envelope and reasoning-budget controls are officially disclosed. Serving behaviour is characterised by heavy test-time reasoning: Benchmark Heaven measures ~104 tokens/sec, while OpenRouter reports a lower median generation speed of ~43 tokens/sec across routed providers, so throughput and TTFT are strongly effort- and provider-dependent. Buyers should benchmark at the exact reasoning effort they intend to ship, budget for 128K-token output ceilings on agent loops, and never assume uniform recall across the full 1.05M-token window, since no independent MRCR v2 long-context figures have been published for this exact variant.
Architecture Type: Proprietary autoregressive transformer (OpenAI GPT-6.2 series flagship); dense vs. sparse mixture-of-experts topology, total/active parameter counts, layer widths and attention variant (MLA/GQA) are undisclosed
No OpenAI technical report or standalone architecture disclosure exists for gpt-6.2-astra; the official model page publishes only serving and capability facts. Confirmed: text/agent modality (the OpenAI model page also lists image input, while the OpenRouter endpoint is text-oriented), 1,050,000-token context window, 128,000-token maximum output, knowledge cutoff 18 May 2026, and configurable reasoning effort at low / medium / high / xhigh / max with no published per-request thinking-token budget. Reported capability surfaces include long-horizon terminal and repo work, computer use, document workflows and multi-agent delegation, plus native tool/web-search integration. Pricing: $8/1M input, $40/1M output, $0.40/1M cached input, $10/1M cache write, $10 per 1,000 web searches (OpenRouter). Because MoE status is genuinely unknown, isMoe is recorded as false (unconfirmed rather than positively dense).
Scientific and engineering research pipelines: research-grade physics, multi-step experiment reasoning and scientific workflow automation (Terminal-Bench Science 78.1%, CritPt v2 Hard 71.5%, HLE-Diamond 72.2%)
Legal and professional knowledge work: long-document analysis, contract/issue-spotting and client-ready drafting where Harvey LAB-AA reaches 93.8%
Cyber defense and secure-code review: CWE-class vulnerability detection and remediation review (CWE-Bench-AA 65.0%, DeepSecBench-AA 46.8%, AA Cyber Index 37.3)
Spatial/3D and embodied planning: 3D experience generation and robotics instruction planning (AA-3D 1318 Elo leader, AA-Robotics v0.1 79)
High-assurance math and formal reasoning workloads: FrontierMath Tier 4 saturation at 100% and Tiers 1-3 at 92.6% for verification-heavy pipelines
Operational Caveats & Boundaries
•Reasoning-effort cliff: headline scoring is at max/xhigh effort (AA Intelligence Index 62), but at high effort the index drops to 60 and per-task latency/cost rise sharply — pin and validate the exact effort level you ship.
•Factual recall is not uniformly best-in-class: Artificial Analysis ranks it below Claude Fable 5.2 on AA-Omniscience v1.1 and HLE-Diamond despite its 50.0 Omniscience Index, so retrieval grounding is still required for citation-critical work.
•Cyber autonomy is uneven: strong static/benchmark vulnerability work (CWE-Bench-AA 65.0%) coexists with 0.0% on CyberGym and only 37.3 on the AA Cyber Index — do not assume reliable end-to-end exploit reproduction or unattended offensive-security loops.
•Throughput and TTFT are provider- and effort-dependent: ~104 tps on Benchmark Heaven vs. ~43 tps median via OpenRouter routing, with long time-to-first-token when max-effort reasoning tokens are enabled; plan concurrency and streaming UX accordingly.
•Context/output asymmetry: 1,050,000-token input but a 128,000-token output ceiling, and no independently published MRCR v2 long-context recall figures for this exact variant — validate recall near the window edge before relying on whole-corpus stuffing.
•Cost surface beyond headline pricing: cache writes at $10/1M, web search at $10 per 1,000 calls, and agent loops that re-read large contexts can make realized cost per task far exceed the $8/$40 sticker; budget per completed task, not per token.
ARMES Zero Data Retention
Calls to GPT-6.2 Astra are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by OpenAI.