OpenAI has officially launched its new GPT-5.6 model series, comprising the flagship Sol, balanced Terra, and cost-effective Luna. This update not only marks a significant performance leap but also signals the retirement of the standalone Codex application, with its core capabilities fully integrated into ChatGPT. This integration forms a new AI super-application, powered by GPT-5.6 as its central engine.
Regarding pricing, Sol is set at $5 per million input tokens and $30 per million output tokens. Terra's costs are halved at $2.5 for input and $15 for output, while Luna offers the most economical option at $1 for input and $6 for output. These models are now fully launched and will progressively roll out across ChatGPT, Codex (referring to its integrated capabilities), and the API within 24 hours.
At the launch event, Claude emerged as an implicit "main character." In the Agents’ Last Exam, a long-range agent workflow evaluation spanning 55 industries, GPT-5.6 Sol achieved a new high score of 53.6 points, surpassing Claude Fable 5 (adaptive inference) by 13.1 points. Even with only medium inference enabled, Sol still outscored Fable 5 by 11.4 points, at approximately one-quarter of the cost. The smaller Terra and Luna also exceeded Fable 5, costing about one-sixteenth as much.
On the broader Artificial Analysis Intelligence Index, Sol with full inference scored only 1 point less than Fable 5, yet it completed tasks 61% faster and at roughly half the cost. The core narrative of this release is clearly cost-efficiency. Sam Altman also emphasized on X that OpenAI heard enterprise clients' concerns about AI costs, and GPT-5.6 Sol represents a significant leap forward in "cost per task." To further reinforce this, new billing rules introduce a prompt caching mechanism: writing is charged at 1.25 times the standard input price, while reading receives a 90% discount. Developers can also manually set cache breakpoints, ensuring retention for at least 30 minutes, making billing more predictable.
Programming is a key battlefield for GPT-5.6. Sol achieved a new record of 80 points on the Artificial Analysis Programming Agent Index, surpassing Fable 5 by 2.8 points, with less than half the output tokens and time, and approximately one-third of the cost. It also set new best scores on Terminal-Bench 2.1 and DeepSWE. Terra and Luna also performed strongly: Terra slightly exceeded Fable 5, and Luna surpassed Opus 4.8, both completing tasks in roughly one-third of the time and at about one-quarter of the cost of their competitors. Early customer feedback is positive; for instance, a Lovable co-founder reported that the new models reduced user steps for application building by about 25%, tool calls by 35% to 48%, and stalled tasks by 15%. Developer Matt Shumer showcased a highly detailed voxel Manhattan autonomously generated by Sol over nearly a week.
OpenAI also highlighted GPT-5.6's significant advancements in design judgment and computer operation. The company stated that it achieved a "leapfrog improvement" in design capabilities, generating aesthetically pleasing and user-friendly interfaces from high-level directions. Crucially, it leverages enhanced computer operation abilities to self-inspect rendered outputs, identify visual and functional issues, and perform corrections. In knowledge work scenarios, Sol achieved a new record of 92.2% in the BrowseComp web browsing benchmark, and 62.6% in the OSWorld 2.0 computer operation evaluation, surpassing Opus 4.8 while reducing output tokens by 85%. When handling PPTs, documents, and spreadsheets, GPT-5.6 can infer and apply complete design systems, ensuring format and data accuracy.
Nonetheless, GPT-5.6 does not lead in all benchmarks. On SWE-Bench Pro, Sol scored 64.6%, trailing Claude Fable 5 at 80% and Anthropic's higher-tier Mythos 5 at 80.3%. For the most challenging FrontierMath Tier 4 math problems, Sol achieved 65.9%, even lower than its predecessor GPT-5.5's 72.5%, while Fable 5 scored 87.8%. In the GDPval-AA professional work evaluation, Sol's 1747.8 Elo was slightly below Fable 5's 1759.6. Notably, in the GeneBench Pro biology benchmark, OpenAI specifically mentioned that Claude Fable 5 was excluded because it "does not answer high-level biological questions and declined most problems in this evaluation," revealing a subtle competitive jab.
GPT-5.6 introduces the "ultra mode," which deploys up to sixteen agents in parallel, exchanging more tokens for stronger results and faster execution. This mode significantly boosts performance in benchmarks like BrowseComp, SEC-Bench Pro, and Terminal-Bench 2.1. Furthermore, "max mode" provides models with extended thinking time for reasoning and validation. On the API side, a new Programmatic Tool Calling feature allows models to autonomously write mini-programs to orchestrate tools and filter intermediate data, greatly reducing token transfers and enhancing the model's ability to execute complex tasks independently. Product-wise, OpenAI unveiled ChatGPT Work—an agent capable of acting across applications and files, executing projects for hours, and delivering finished products, leveraging underlying Codex technology. Additionally, a new desktop app and directly shareable Sites have been launched.
[AgentUpdate Depth Analysis]
The launch of GPT-5.6, especially the deep integration of Codex capabilities and the introduction of Programmatic Tool Calling, undoubtedly injects significant momentum into the AI Agent ecosystem. Compared to prevalent Agent frameworks like LangChain or CrewAI, GPT-5.6 offers more advanced native self-planning and tool orchestration capabilities. While traditional Agents require developers to explicitly define tools and manage state flows, GPT-5.6's ability to "self-write scheduling scripts" significantly reduces Agent development complexity and enhances its robustness and efficiency in complex tasks. The implementation of parallel multi-agent collaboration in ultra mode heralds a future where Agents can more effectively break down and solve large-scale problems, which is revolutionary for enterprise applications. In the future, we anticipate Agent architectures relying less on hard-coded logic and more on emergent model "intelligence," transitioning from "tool users" to "intelligent workflow orchestrators." This will accelerate AI Agents' evolution into true "digital employees," profoundly impacting automation in fields like programming, data analysis, and even creative design, pushing towards "AIGC process automation."
