GPT 5.6 സോൾ

OpenAI's frontier model for complex professional work and the flagship of the GPT-5.6 family, corresponding to the unsuffixed tier of earlier GPT-5 generations. It is built for demanding reasoning, coding and agentic workflows, and is particularly strong at command-line and multi-step coding tasks. Reasoning effort spans none through max, so one model covers a quick lookup and a long deliberate build.

Sol is the top of a three-tier family that OpenAI shipped together after a limited preview, and the naming is deliberate: the number is the generation, while Sol, Terra and Luna are durable capability tiers that the lab says can advance on their own cadence. Sol is the tier it puts in front of work that has to be right - long professional tasks, real codebases, agents that keep going without a person re-steering them every few steps.

The claim OpenAI leads with is not a peak score, it is work per token. On Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, the lab reports Sol setting a new high of 53.6, ahead of the strongest competing model by 13.1 points, and says that even at medium reasoning effort it stays ahead by 11.4 points at roughly a quarter of the estimated cost per task. Its published table puts Sol at 52.7% there against 46.9% for GPT-5.5. The same argument runs through the coding results: OpenAI reports new state-of-the-art scores on Terminal-Bench 2.1 at 88.8% and on DeepSWE at 72.7%, with 64.6% on SWE-Bench Pro, and says it reaches them while spending less than half the output tokens and less than half the time of the model it displaces.

The effort dial is the feature that makes those numbers readable. Reasoning effort runs from none through low, medium, high and xhigh to max, and OpenAI describes max as giving the model more time than xhigh to reason, explore alternatives, run checks and revise its approach. Above that sits ultra, which coordinates four agents in parallel by default and trades higher token use for both stronger results and faster time to an answer. Programmatic Tool Calling in the Responses API lets Sol write and run small programs that coordinate tools and filter intermediate data, so tool-heavy work advances with fewer round trips through the model.

Away from code, OpenAI reports 90.4% on BrowseComp and 62.6% on OSWorld 2.0, the latter with 85% fewer output tokens than the model it beats, plus 70.6% on BenchCAD and 83.4% with a Python tool available. It also puts weight on design judgement: stronger computer use lets the model inspect the rendered result rather than only emit the code behind it, catch visual and functional problems, and finish the work before handing it back. On documents, decks and spreadsheets the lab says the biggest gain is in following a supplied template or reference file faithfully, including rules carried in a slide master.

OpenAI is specific about where the model stops. It reports GPT-5.6 as more capable than its earlier models in both biology and cybersecurity while crossing the Critical threshold in neither, and says its testing shows the model is better at finding and fixing vulnerabilities than at reliably carrying out autonomous end-to-end attacks against hardened targets, and that in biology it can support legitimate research without providing the end-to-end capability needed to create a dangerous novel threat. It also warns about the cost of that caution: Sol's cyber safeguards block roughly ten times more potentially harmful activity than previous models, which creates friction for benign work, so ChatGPT and Codex offer a one-press retry on a lower-capability model.