GPT-6.2 Sol
GPT-6.2 Sol (openai/gpt-6.2-sol) is OpenAI's efficient, versatile mid-tier reasoning model from the November 17, 2026 GPT-6.2 release — positioned below flagship GPT-6.2 Astra and above GPT-6.1 Sol. It pairs a 1,050,000-token context window (≈922K usable input tokens), 128K max output, a June 1, 2026 knowledge cutoff and $2/$10 per-million-token pricing with adjustable reasoning effort (low→max). Notably it matches GPT-6 Astra on GPQA Diamond (no tools) at 96%, scores AA Intelligence Index v5.0 = 57.3 at max effort, and is SOTA-class across terminal, science, legal and cyber evaluations in Benchmark Heaven telemetry.
Empirical Evaluation Results
Architectural Profile & Capabilities
GPT-6.2 Sol is the workhorse member of OpenAI's GPT-6.2 generation, announced November 17, 2026 alongside GPT-6.2 Astra. OpenAI markets it as "our go-to model for everyday work" — an upgrade to GPT-6.1 Sol that is "smarter, more efficient, and lets you complete more tasks" — with Astra reserved for the hardest multi-step professional projects. The variant is served through the OpenAI API (model ID gpt-6.2-sol), ChatGPT Work, Codex and OpenRouter; a separate "GPT-6.2 Sol Pro" listing is the same underlying model with different serving characteristics, so integrations should pin the canonical slug. Verified specification (variant-specific, OpenAI developer docs / system card / OpenRouter): 1,050,000-token context window with up to 922,000 input tokens per request; 128,000 max output tokens; text + image input, text output; knowledge cutoff June 1, 2026 (GPT-6.1 Sol was April 30, 2026); pricing $2/1M input and $10/1M output; selectable reasoning effort (low / medium / high / xhigh / max, with OpenAI recommending medium as the default starting point for new integrations); reasoning tokens billed as output and supported via the Responses API. OpenAI also ships an "Ultrafast" serving option at launch. Architecture: GPT-6.2 Sol is a closed-weight autoregressive reasoning transformer. OpenAI has not published parameter counts, layer/head configuration, attention variant (MHA/GQA/MLA) or an explicit dense-vs-MoE topology statement in any material surfaced for this release; the system card and model docs cover capability, safety and deployment-specification data rather than training topology. It is therefore catalogued here as a dense/non-MoE transformer only by default — no mixture-of-experts routing has been confirmed, and no neural-architecture claims should be derived from this dossier. What is documented is the inference-time behaviour: test-time reasoning with a six-position effort ladder, where Artificial Analysis measures the Intelligence Index moving 41 (low) → 53 (medium) → 55 (high) → 56 (xhigh) → 57 (max), and a max-effort run averaging ~40,360 output tokens per index task (22,800 reasoning + ~17,561 answer tokens) — roughly 40% fewer tokens than GPT-6.1 Sol at the same setting, which is the concrete basis of the "more efficient" claim. Capability profile: on the aggregate AA Intelligence Index v5.0 it records 57.3, level with GPT-6 Astra and two points above GPT-6.1 Sol, but below GPT-6.2 Astra. It is unusually strong for a non-flagship tier on hard science (HLE-Diamond 64.0%, Terminal-Bench Science 67.6%, CritPt v2 Hard 62.5%, GPQA Diamond 96% no-tools — matching GPT-6 Astra outright), on legal knowledge work (Harvey LAB-AA 91.5%), and on agentic/terminal coding (Terminal-Bench 4.0 60.1%, FrontierCode 2.0 44.4%). Cyber results are strong on secure-code evaluation (CWE-Bench-AA 61.7%, DeepSecBench-AA 51.6%, AA Cyber Index 37.7) though CyberGym resolution is reported at 0.0% — a benchmark-contamination/mismatch caveat rather than a clean capability signal. Its weakest documented areas are strict all-pass legal orals (Vals HLAB strict all-pass 11.1%) and healthcare depth (HealthBench 53.9%, MLCR-AA Medical 37.2%) relative to its science ceiling. Artificial Analysis places it at 1,259 Elo on AA-3D (Astra 1,318; Claude Fable 5.2 1,271) and 78 on AA-Robotics v0.1 (Astra 79). Landscape position (October/November 2026): Sol is the price-performance layer of the OpenAI stack, not a frontier-absolute model. Against Claude Opus 5 / Sonnet 5, Gemini 3.8 Flash / Pro, DeepSeek V4 Pro / V4.1 Flash, xAI Grok 4.6, Qwen 3.8 and Kimi K3, its defensible advantages are the 1.05M context, the 2:10 price ratio and its unusually flat science/legal profile for a mid-tier SKU; its documented liabilities are a coding-agent regression versus its own predecessor on SWE-Atlas-QnA-AA and a stalling aggregate coding trajectory reported at launch. Operational read: benchmark Sol at the effort level you intend to deploy. The 41→57 Index spread across the effort ladder is larger than the gap between Sol and Astra, so most "is Sol good enough?" questions are really effort-configuration questions.
OpenAI has published deployment-level specifications but not training topology for GPT-6.2 Sol. Disclosed: model ID gpt-6.2-sol; 1,050,000-token context window with a ~922,000-token per-request input ceiling; 128,000 max output tokens; text + image input, text output; June 1, 2026 knowledge cutoff; adjustable reasoning effort (low / medium / high / xhigh / max) with reasoning tokens billed as output. NOT disclosed: total or active parameter count, layer/head configuration, attention mechanism (MHA/GQA/MLA/MQA), tokenizer details, or an explicit dense-vs-MoE statement — isMoe is recorded as false only because no mixture-of-experts routing has been confirmed in any public material, not because a dense design has been affirmed. Empirical reasoning behaviour is documented: Artificial Analysis measures Intelligence Index 41 / 53 / 55 / 56 / 57 at low / medium / high / xhigh / max, with a max-effort run consuming ~40,360 output tokens per index task (22,800 reasoning + ~17,561 answer), about 40% below GPT-6.1 Sol at the same setting. Measured output throughput on Benchmark Heaven is ~91 tokens/sec, consistent with a model that streams a substantial reasoning trace before answering.
Recommended Workloads & Primary Use Cases
- •Headline context ≠ usable context: the 1,050,000-token window is capped at ~922,000 input tokens per request, and prompts in the high-hundreds-of-thousands range push time-to-first-token and input cost up sharply — chunk-and-retrieve rather than relying on the full window for latency-sensitive paths.
- •Effort configuration dominates outcomes: the AA Intelligence Index moves from 41 at low effort to 57 at max, a spread larger than the gap to GPT-6.2 Astra, while max effort consumes ~40,360 output tokens per index task. Budget reasoning tokens as output ($10/1M) and benchmark at the exact effort level you will deploy — never compare a max-effort score to an Astra low-effort score.
- •Documented coding-agent regression versus its own predecessor: reporting at launch noted the Coding Agent Index running about three points below GPT-6.1 Sol at max compute, driven partly by a ~14-point decline on SWE-Atlas-QnA-AA. Re-run your own eval harness before migrating existing Codex/agent pipelines, and do not assume a version bump is a coding upgrade.
- •Hard-science strength does not transfer uniformly to regulated domains: Vals HLAB strict all-pass sits at 11.1% and HealthBench at 53.9%, so strict-scored legal orals and clinical decision support still require human-in-the-loop review. Treat 91.5% on Harvey LAB-AA as knowledge-work competence, not as a compliance attestation.
- •Ecosystem and metering traps: an OpenRouter "GPT-6.2 Sol Pro" listing is the same underlying weights with different serving behaviour, and reasoning tokens are billed as output — pin the canonical slug (openai/gpt-6.2-sol) and set explicit max_output_tokens / effort caps per route. Also note CyberGym is recorded at 0.0%, a harness/contamination artefact that makes single-number cyber rankings unreliable.
Calls to GPT-6.2 Sol are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by OpenAI.
More Models from OpenAI
Compare other engines available in this laboratory.
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.
OpenAI's flagship September 2026 frontier model with a 1.05M context window, MoE architecture, and SOTA reasoning across coding, formal mathematics, and long-horizon agentic automation.
GPT-6.1 Sol (openai/gpt-6.1-sol) is OpenAI's September 29, 2026 mid-tier reasoning model, an upgrade to GPT-6 Sol positioned one step below the flagship GPT-6 Astra. It targets agentic coding, computer use, and document-heavy professional automation at roughly one-fifth of Astra's per-token cost, with a 1.05M-token context window, text+image input, and a measured output throughput of ~53 tokens/sec.