GPT-6 Sol

GPT-6 Sol is the middle tier of OpenAI's third-generation GPT-6 family. It was announced and rolled out on 22 September 2026 in ChatGPT Work, Codex, and the API, roughly three weeks after GPT-6 Astra (3 September 2026) and alongside the cheaper GPT-6 Luna. OpenAI's own model guidance frames the family explicitly by workload: "Start with GPT-6 Astra for complex reasoning and coding, choose GPT-6 Sol to balance intelligence and cost, or use GPT-6 Luna for cost-sensitive, high-volume" work. Sol therefore inherits the Astra generation's post-training recipe and tool-use behavior while being re-optimized for token efficiency and price. ARCHITECTURE: OpenAI has published no architecture paper, parameter count, attention scheme, or positional-encoding detail for GPT-6 Sol; there is no public weight release and no system card with an architectural appendix at launch. On the balance of evidence — OpenAI's sustained MoE lineage from GPT-4-era onward, the three-way Astra/Sol/Luna tiering that mirrors capacity-and-cost segmentation rather than capability-era segmentation, and the presence of user-selectable reasoning effort with hidden reasoning tokens — the defensible classification is an autoregressive, decoder-only sparse Mixture-of-Experts transformer with a multimodal (image) input encoder and test-time compute scaling via reasoning effort levels. Total/active parameter counts are undisclosed and should be treated as unknown, not assumed. The model is not a diffusion LLM and is not non-autoregressive. Inference optimization claims center on "more efficient caching and inference" rather than architectural disclosure. INTERFACES AND SPECIFICATIONS: The documented context window is 1,050,000 tokens with a maximum output of 128,000 tokens (the predecessor GPT-5.6 Sol published a 922,000 maximum input against the same 1,050,000 window, and that asymmetry is expected to carry over). Reasoning effort supports none, low, medium (default), high, xhigh, and max — the same ladder introduced with GPT-5.6 — with reasoning tokens billed as output tokens. Codex ships tuned to a 272,000-token working context and the full 1.05M window requires explicit configuration, a recurring source of confusion in the developer community. Reported API pricing is $2 per 1M input tokens and $10 per 1M output tokens, consistent with OpenAI's announcement of roughly 50% lower API prices than GPT-5.6 promotional pricing (GPT-5.6 Sol was $5/$30 at launch, later promo-cut). Independent routing telemetry (OpenRouter) reports image as the declared modality alongside text. BENCHMARKS AND VERIFICATION STATUS: Because GPT-6 Sol shipped the same day as this dossier's research window, no independent, reproducible scorecard exists yet. The strongest figures currently circulating (GPQA Diamond ~90.9, SWE-bench Verified ~80.2, Terminal-Bench 2.1 ~70.8) read as first-party model-card claims pending third-party reproduction by benchmark hosts. For calibration, the immediately preceding generation is well documented: GPT-5.6 Sol reported ~94% GPQA Diamond, ~88% Terminal-Bench 2.1, 72.7% DeepSWE v1.1, ~32% CritPt, and an Artificial Analysis Coding Agent Index of 80 (max reasoning, briefly SOTA). GPT-6 Astra, the tier above, reports 96.0% GPQA Diamond, 74.1% DeepSWE v1.1, 58.2% Terminal-Bench 4.0, and 64.6% Terminal-Bench Science — with a dramatic Astra-over-Sol separation on science-terminal work (64.6% vs 22.4% for the older Sol). Expect GPT-6 Sol to land somewhat below Astra on agentic/science suites and materially above GPT-6 Luna, with the efficiency story (lower output tokens and cost per task) being the actual product claim rather than raw peak scores. Treat every non-published number as provisional and re-verify against the official model page and Artificial Analysis within 1-2 weeks of launch. OPERATIONAL PROFILE: Sol is a reasoning-by-default model whose cost is dominated by output tokens, not input. With input at $2/1M and output at $10/1M — a 5:1 ratio — and hidden reasoning tokens billed at output rates, the effective cost of a task is governed almost entirely by chosen reasoning effort and loop depth. OpenAI's own framing for the Astra generation ("more efficient caching and inference") is a caching claim as much as a speed claim: prompt-cache hit rates on stable prefixes are the single largest lever on real-world spend. Throughput is workload-dependent rather than intrinsic: at none/low effort the model behaves like a fast mid-tier reasoner, while at xhigh/max it becomes a deliberate, thinking-token-heavy model where wall-clock time scales with effort. Time-to-first-token grows with context length and is materially worse once the working set exceeds the 272k Codex default and approaches the 1.05M ceiling. RECOMMENDED FIT: Sol's economic niche is "Astra-quality judgement without Astra pricing" — long-context professional knowledge work (contracts, discovery, whole-repo comprehension), large agentic coding loops where output efficiency compounds across thousands of turns, high-volume structured extraction and report generation sitting in a batch tier, and multimodal screenshot/document understanding for support and QA automation. It is the wrong choice for deep scientific reasoning where Astra's Terminal-Bench Science gap is enormous, and the wrong choice for ultra-high-QPS trivial classification where Luna's price floor wins.