OpenAI said GPT-5.6 focuses on improving the amount of useful work completed per token, aiming to deliver stronger performance while reducing compute requirements and cost. According to the company, GPT-5.6 Sol achieved a score of 53.6 on Agents’ Last Exam, outperforming Claude Fable 5 by 13.1 points. On the Artificial Analysis Intelligence Index, OpenAI said GPT-5.6 Sol with maximum reasoning came within one point of Fable 5 while completing tasks in 61% less time at roughly half the estimated cost.
The company also highlighted gains across coding benchmarks. OpenAI said GPT-5.6 Sol reached a score of 80 on the Artificial Analysis Coding Agent Index, surpassing Claude Fable 5 while using fewer output tokens, lower latency, and lower estimated cost. It also reported state-of-the-art results on Terminal-Bench 2.1 and DeepSWE, benchmarks focused on command-line workflows and long-horizon software engineering tasks.
A major addition accompanying the model family is Programmatic Tool Calling in the Responses API. OpenAI said the feature allows GPT-5.6 to write and execute lightweight programs that coordinate tools, process intermediate results, monitor progress, and determine subsequent actions without routing every intermediate step back through the model. Developers can also access a multi-agent beta in the Responses API to build experiences similar to the new ultra reasoning mode.
OpenAI is also expanding GPT-5.6 beyond software development into broader workplace tasks. The company said the model can generate editable presentations, documents, spreadsheets, and interactive visualizations while drawing context from services including Slack, Notion, Microsoft 365, and Google Drive. It also said GPT-5.6 demonstrates improved computer-use capabilities, allowing it to inspect rendered outputs and refine visual or functional details before returning completed work.
For cybersecurity, OpenAI said GPT-5.6 delivered its strongest results to date across ExploitBench, ExploitGym, and SEC-Bench Pro. The company said the model is intended to support defensive tasks such as secure code review, vulnerability validation, threat modeling, malware analysis, and patch verification. Those capabilities are available through OpenAI Daybreak’s Trusted Access for Cyber program for qualified individuals and organizations.
OpenAI also reported improvements in scientific research, stating that GPT-5.6 Sol achieved stronger performance than GPT-5.5 across biology, life sciences, chemistry, and genomics evaluations while using fewer tokens and less time.
Alongside the performance improvements, OpenAI said GPT-5.6 launches with what it describes as its most robust safety system to date. Before general availability, the company said it conducted its largest evaluation program yet, including human red teaming, external expert testing, automated assessments totaling approximately 700,000 A100e GPU hours, and additional monitoring during the limited preview period. The company said GPT-5.6 includes layered safeguards that combine model training, real-time checks, monitoring, reasoning-based safety systems, and account-level enforcement.
OpenAI said GPT-5.6 is available immediately across multiple products with different access levels. ChatGPT Plus, Pro, Business, and Enterprise users can access GPT-5.6 Sol, while free and Go users receive GPT-5.6 Terra through ChatGPT Work and Codex. Developers can access all three models through the OpenAI API.
The company also announced new pricing for the model family. GPT-5.6 Sol costs $5 per million input tokens and $30 per million output tokens. Terra is priced at $2.50 per million input tokens and $15 per million output tokens, while Luna costs $1 per million input tokens and $6 per million output tokens. OpenAI said the models also introduce updated prompt caching with explicit cache breakpoints and a minimum cache lifetime of 30 minutes.
This analysis is based on reporting from OpenAI.
Image courtesy of The Hacker News.
This article was generated with AI assistance and reviewed for accuracy and quality.
