GPT-6.2 Sol is the model OpenAI built to be used, not admired. It launched on November 17, 2026 at exactly the price of GPT-6.1 Sol — $2 input, $10 output, $0.10 cached input — and exists for a specific economic argument: most production AI work does not need the frontier, and charging frontier prices for it is the fastest way to make a pipeline unprofitable.

What you get for $2 and $10

Sol shares the 1,050,000-token context window and 128K output ceiling with Astra. Multimodal input is text and image; output is text. The knowledge cutoff is June 1, 2026 — 32 days later than GPT-6.1 Sol’s April 30 cutoff, and later than Astra’s. Reasoning effort runs low through max, and unlike the original GPT-6 Sol, none is rejected.

The tool surface is identical to Astra’s: web search, file search, image generation, code interpreter, hosted shell, apply_patch, skills, computer use, MCP, tool search. Sol is not a stripped-down endpoint; it is the same harness at a quarter of the rate.

Benchmarks where Sol is the rational default

On the Artificial Analysis Intelligence Index, Sol scores 57 — matching the original GPT-6 Astra’s performance at roughly 15% of the cost per evaluated task. On FrontierCode 2.0, Sol lands within 0.5 points of Astra at about one-fifth the per-task cost. In Bito’s high-effort coding sessions, Sol scored 76.5% overall at **$2.22 per hard session** against Astra’s 77.5% at $8.19.

LifeVal — OpenAI’s evaluation of challenging everyday tasks like shortlisting schools or finding housing within a budget and commute — places Sol at 0.84 at max effort, #2 on the published chart. PatchBench, which tests security patches that preserve program behavior, puts Sol at 0.76 at max effort, also #2. These are closer to real production work than most headline benchmarks.

The regression that matters

Sol is not strictly better than GPT-6.1 Sol. On Artificial Analysis’s Coding Agent Index, GPT-6.2 Sol loses 3 points relative to GPT-6.1 Sol, driven by a 14-point decline on SWE-Atlas-QnA-AA, where Sol is more likely to omit key facts.

That is a specific, reproducible weakness: Sol may produce a correct-looking answer that leaves out a fact the downstream pipeline needed. For chat assistants and RAG with human review, manageable. For fully automated agent chains where omission propagates silently, a real cost. The mitigation is not to abandon Sol — it is to add a fact-completeness check at the output boundary before routing high-stakes work without review.

The cost math that decides the default

The rate card is half the story. At max effort, GPT-6.2 Sol spends 40,360 output tokens per Artificial Analysis index task, where GPT-6.1 Sol spends 67,824 — 40% less generated work on the same meter. OpenAI’s positioning rests on this: better results with fewer output tokens, so a given budget covers more work.

For a pipeline running thousands of tasks daily, effective cost per completed task determines the bill. If Sol’s token-efficiency gain holds on your workload, the same budget buys roughly 1.6× more tasks than GPT-6.1 Sol — before accounting for Sol already costing a quarter of Astra.

Where Sol is the wrong call

Sol is wrong when the task combines long-horizon planning, expensive failure, and a domain where Astra’s independent benchmarks lead: scientific research, interactive 3D, robotics planning, sustained computer-use. It is also wrong if your coding-agent pipeline cannot tolerate the SWE-Atlas-QnA-AA omission regression without an added verification layer.

For everything else — high-volume chat, classification, retrieval, routine review, normal-sized diffs, multi-step agents that do not need frontier reasoning — Sol is the model that makes the economics work. The default-model question has shifted: the premium tier is now the exception, not the starting point.