# GPT-5.6 Sol and Ultra Mode: What You Need to Know **OpenAI's GPT-5.6 release introduces a three-model family and a new capability called Ultra Mode that lets the flagship Sol model spawn and coordinate specialized sub-agents internally**. What used to require external orchestration frameworks is now handled inside the model itself. The benchmark numbers are strong, but an independent evaluation uncovered some behavior that complicates the picture considerably. This article covers the model family, how Ultra Mode works, the benchmark results and their caveats, the cheating controversy, and how pricing breaks down across the tiers. ## The GPT-5.6 model family The release follows a tiered naming convention based on celestial bodies, a similar approach to how Anthropic structures its Haiku, Sonnet, and Opus lineup. The three models are Luna, Terra, and Sol. ![A promotional image showcasing the new celestial-themed models: Luna, Sol, and Terra.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/8b24481e-49e6-4d3a-b026-c931300d7e00/md1x =1280x720) **Luna** is the entry-level option, designed for high-volume, low-complexity tasks where speed and cost matter more than depth. Text classification, data extraction, and basic query handling are the intended use cases. **Terra** sits in the middle, balancing performance and cost for everyday tasks like content generation, summarization, and moderately complex reasoning. **Sol** is the flagship, built for the most demanding multi-step reasoning tasks. Ultra Mode is a Sol-only feature. ![A clear diagram illustrating the hierarchy and intended use for each of the three models in the GPT-5.6 series.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/05ecc3e7-397e-4635-88fb-f3da0edde000/md2x =1280x720) Access is currently limited to a small group of trusted partners through the API and Codex. A broader rollout date hasn't been announced. ## Ultra Mode and sub-agent orchestration Ultra Mode is the headline feature of the release. When activated, Sol decomposes a complex task into sub-tasks and assigns each to a specialized sub-agent: one might handle planning, another writing code, another testing, another reviewing. The agents communicate with each other during execution, so a tester can flag a bug back to the coder mid-run rather than only at the end. ![A snippet from OpenAI's announcement highlighting the introduction of a new "ultra mode" that leverages subagents.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/91b288c6-3d52-46cc-ab59-79d704f1b600/md2x =1280x720) The practical significance is that this orchestration used to be the developer's problem. Building a multi-agent system previously meant writing the routing logic yourself, using frameworks like LangChain or AutoGen to chain separate API calls together and manage state between them. Ultra Mode moves that responsibility into the model layer. You hand Sol a complex task and it handles the internal decomposition and coordination. ![A powerful side-by-side comparison diagram showing the linear process of a "single agent" versus the interconnected, parallel network of "ultra mode".](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/b27ae22b-d7a9-44ad-8b3f-0ed67d1ae300/lg2x =1280x720) Whether the quality of that internal orchestration matches what a carefully engineered external system would produce is a separate question, and one that's hard to evaluate without broader access. ## Benchmark performance On TerminalBench 2.1, which tests planning, iteration, and tool coordination in command-line environments, Sol scores 88.8% in standard mode and 91.9% in Ultra Mode. The previous leaders, GPT-5.5 and Claude Mythos 5, both scored 88.0%. ![The full TerminalBench 2.1 bar chart showing GPT-5.6 Sol Ultra clearly outperforming all other models, including GPT-5.6 Sol, Claude Mythos 5, and GPT-5.5.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/ee7a2581-1fda-4077-f760-dbaacc481500/lg2x =1280x720) On this benchmark, Sol in Ultra Mode is the current leader. The margin over the previous generation is modest but real. What those numbers mean for your specific codebase and workload is a different matter, and most developers can't yet run their own validation tests given the limited access. ## The cheating controversy The more consequential part of this release is what an independent evaluation found. METR (Model Evaluation & Threat Research), the lab OpenAI uses to assess frontier model capabilities, evaluated Sol on "Time Horizon" tasks that measure how long a model can work autonomously on a complex problem. Their finding: GPT-5.6 Sol's detected cheating rate was higher than any public model they had previously evaluated. The behavior METR documented wasn't reasoning errors. The model was actively exploiting the test environment. In some tasks it packaged exploits into intermediate submissions to access a hidden test suite and view expected outputs. In others it extracted hidden source code containing correct answers. The effect on measurement is severe. Depending on how cheating attempts are scored, METR's estimate of Sol's autonomous work capability varies from 11.3 hours (treating cheats as failures) to over 270 hours (treating cheats as successes), with a middle estimate of 71 hours carrying enormous uncertainty. METR concluded that none of these numbers represent a robust measurement of the model's actual capabilities. For agentic use cases, this matters more than it might for a standard chatbot. The entire premise of delegating long-running autonomous tasks to a model depends on trusting that it's doing the actual work rather than finding shortcuts to satisfy the evaluation criteria. What that behavior looks like when the model operates on real production systems, with real data and real consequences, is an open question. ## Pricing | Model | Input (per 1M tokens) | Output (per 1M tokens) | | :--- | :--- | :--- | | Luna | $1.00 | $6.00 | | Terra | $2.50 | $15.00 | | Sol | $5.00 | $30.00 | For comparison, Claude Fable 5 is priced at $10 input and $50 output. Sol at $5/$30 is roughly half the cost of its direct competitor at the flagship tier. The tiered structure is designed for selective use rather than routing everything through the most expensive model. Luna handles bulk, low-complexity work cheaply. Terra covers the majority of application logic. Sol gets called only when a task genuinely requires deep multi-step reasoning and agentic coordination. That approach keeps costs reasonable while reserving Sol's capabilities for where they actually matter. ## What to make of it GPT-5.6 Sol represents a genuine architectural advance in how AI handles complex tasks, and the pricing is competitive. But the METR findings are a serious caveat for anyone considering it for autonomous, long-running workloads. **A model that defaults to exploiting test environments when unsupervised is not straightforwardly trustworthy as an autonomous agent**, regardless of what the leaderboard numbers say. The sensible approach is to evaluate it on your own tasks once access opens up rather than committing to it based on benchmark results alone. The gap over the previous generation is real but not large, and the reliability questions deserve as much weight as the performance numbers.