GPT-6.1-Sol
GPT-6.1 Sol is OpenAI's mid-tier GPT-6-series flagship (released Sept 29, 2026, ~7 days after GPT-6 Sol), delivering "near-Astra" agentic coding, computer-use and document-heavy performance at $2/$10 per 1M tokens with a 1.05M-token context window and native image input. It replaces GPT-6 Sol in ChatGPT Work/Codex and the API slug gpt-6.1-sol.
Empirical Evaluation Results
Architectural Profile & Capabilities
GPT-6.1 Sol (OpenRouter slug openai/gpt-6.1-sol; OpenAI API id gpt-6.1-sol) is an OpenAI GPT-6-series model launched September 29, 2026, superseding GPT-6 Sol (released roughly seven days earlier) and sitting directly beneath the flagship GPT-6 Astra. OpenAI positions it as delivering near-Astra capability for agentic coding, computer use, and document-heavy professional workflows at a fraction of Astra cost. erified API/product facts: 1,050,000-token context window, up to 128,000 completion tokens, text+image (vision) input with text output, $2/1M input, $10/1M output, $0.10/1M cached input (OpenRouter also lists $2.50/1M cache-write). It is available in ChatGPT Work and Codex for Plus/Pro/Business/Enterprise/Edu users and via the OpenAI API (and reported Azure/Bedrock channels), and shipped day-one into GitHub Copilot. A "Pro" configuration exposing reasoning.mode = 'pro' is offered on OpenRouter and is described as the same underlying model served at higher reasoning effort rather than a distinct parameterized checkpoint. Architecture: OpenAI does not publish parameter counts, attention variant, or routing for this model. Its system-card addendum states GPT-6.1 Sol uses the same data and training types as GPT-6 Astra, but no dense-vs-MoE, active-parameter, or attention-mechanism (MHA/GQA/MLA) disclosure exists. It is an autoregressive multimodal reasoning model with a test-time reasoning/'thinking' budget that scales via effort modes; per Artificial Analysis, GPT-6.1 Sol uses roughly 10-30% more output tokens than GPT-6 Sol across effort levels, indicating a reasoning-heavy decode profile rather than a non-autoregressive or diffusion paradigm. Benchmarks: No official numeric scorecard (SWE-bench Verified, GPQA Diamond, MATH-500, AIME, LiveCodeBench, ARC-AGI, TAU-bench, OSWorld) has been published for GPT-6.1 Sol at launch; OpenAI's own materials and the safety addendum stay qualitative ("near-Astra," outperforming prior OpenAI models on production cybersecurity evals). Independent third-party telemetry is thin and primitive: Artificial Analysis places it at roughly Claude Opus 5.5 Medium intelligence tier at under a third of the price, and OpenRouter's own provider row reports ~39 tokens/s throughput with ~2.88 s latency. Treat any specific coding/reasoning benchmark number circulating for this model as unverified until an official system card table or independent harness publishes it.
OpenAI discloses no parameter count, active-parameter count, sparse routing, or attention variant (MHA/GQA/MLA) for GPT-6.1 Sol. The system-card addendum only confirms it shares GPT-6 Astra's data and training regime. Observably it is an autoregressive text+image-in / text-out model with a variable test-time reasoning budget (standard vs 'Pro' reasoning mode) and a long-context design reaching 1.05M tokens with 128K max output. It emits ~10-30% more output tokens than GPT-6 Sol across effort levels, consistent with a reasoning-heavy frontier decode profile. No evidence supports a diffusion/LLM or non-autoregressive classification.
Recommended Workloads & Primary Use Cases
- •Capability benchmark opacity: no official SWE-bench Verified, GPQA Diamond, MATH-500, LiveCodeBench, ARC-AGI, TAU-bench or OSWorld figures were published at launch — validate with your own harness before committing to front-line duties.
- •Token-amplification cost: it consumes ~10-30% more output tokens than GPT-6 Sol on equivalent tasks, so effective cost per completed task can exceed the headline $10/1M output rate, especially in Agent/Pro reasoning modes.
- •Throughput/TTFT tradeoff: observed provider performance is only ~39 tokens/s with ~2.88 s latency; with reasoning tokens enabled, time-to-first-meaningful-output and overall wall-clock latency can be substantially higher than non-reasoning models, making it a poor fit for sub-second interactive UX.
- •Context caveat: the 1,050,000-token window is a nominal limit — no published needle-in-a-haystack or long-context recall curve exists, so assume retrieval/quality degradation and rising latency/cost at extreme context lengths; also confirm the 128K completion cap applies uniformly across OpenAI, Azure and OpenRouter endpoints.
- •Model-churn risk: it replaced GPT-6 Sol after ~7 days, and the 'Pro' tier is an inference-configuration (reasoning effort) rather than a separate checkpoint — pin model/version strings and reasoning-effort settings in production to avoid silent behavior drift.
- •Prompt/formatting sensitivity: as a reasoning model, it can over-deliberate or wrap answers in reasoning traces; constrained-output, structured-JSON and strict tool-calling workloads should be regression-tested against reasoning-mode on/off.
Calls to GPT-6.1-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.
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 Sol is OpenAI's mid-tier, cost-efficient high-end GPT-6 model (launched 22 Sept 2026, ~three weeks after flagship GPT-6 Astra), exposing a 1,050,000-token context window, 128K max output, image input, selectable reasoning effort, and hidden test-time reasoning tokens at $2/$10 per 1M tokens — positioned between GPT-6 Astra (quality) and GPT-6 Luna (throughput/price).
GPT-6 Luna foundation intelligence model.