OpenAI introduces the GPT-5.6 Sol trial – AI News – #1 July 2026

With the launch of the GPT-5.6 generation, OpenAI is introducing a cleaner, more predictable naming system. Instead of complex suffixes, users are presented with three distinct capability tiers that can evolve independently:

Breaking the architecture down into three segments allows developers and enterprises to precisely optimize their AI infrastructure costs depending on the complexity of the problem at hand.

The most exciting innovation in the GPT-5.6 Sol architecture is its approach to inference-time compute, or reasoning effort. OpenAI is moving away from rigid, instantaneous response generation, opting instead to give the model “time to think deeply” about a problem.

The newly introduced max mode gives the Sol model the latitude for profound, multi-step analysis before delivering its first output token. Meanwhile, the ultra mode goes a step further, pushing beyond the constraints of a single LLM instance. GPT-5.6 Sol Ultra can autonomously spin up specialized sub-agents, delegate smaller sub-tasks to them, and then coordinate their work to synthesize results for highly complex, multi-layered engineering problems.

This advanced reasoning capability translates directly into massive performance leaps across demanding engineering and scientific tests. In programming tasks, the Sol model running in Ultra mode set a new record on the challenging Terminal-Bench 2.1, which evaluates planning and tool iteration in a command-line environment:

Remarkable progress was also recorded in GeneBench v1 (an evaluation for genomics and quantitative biology analysis requiring long-chain reasoning). GPT-5.6 Sol not only outclasses its predecessors in accuracy but does so while consuming significantly fewer output tokens than GPT-5.5. Similar efficiency gains are evident in cybersecurity tests (ExploitBench and UC Berkeley’s ExploitGym), where Sol competes with top market rivals while using only about a third of the output tokens previously required.

The rise of model autonomy and advanced agentic capabilities naturally raises concerns about digital security. In response, OpenAI has equipped the GPT-5.6 series with its most sophisticated multi-layered adversarial defense system to date.

If the real-time system detects high risk or an obfuscated jailbreak attempt, generation is halted instantly. The entire conversation context is then escalated to a larger reasoning model, which determines whether to block the content or permit it—ensuring legitimate defensive research, like code reviews or patch creation, can still proceed.

To ensure these defenses hold up against evolving, real-world threats, OpenAI leveraged its own AI to stress-test the new architecture. The company dedicated over 700,000 GPU hours on NVIDIA A100 clusters to automated red-teaming.

The primary objective was uncovering “universal jailbreaks”—exploits capable of working across various contexts and prompts rather than a single, narrow loophole. This automated process was supplemented by extensive human red-teaming from external experts, mapping out thousands of non-linear abuse scenarios before the trial even opened.

While these models are slated for a wider rollout across ChatGPT and the API in the coming weeks, the current restricted pilot phase has caused a stir. As part of an ongoing collaboration with the US government, OpenAI shared the models’ capabilities and risk profiles with the administration prior to release. The current closed trial is a direct result of coordinating with officials as they draft new executive frameworks for cybersecurity.

Interestingly, OpenAI explicitly stated in its announcement that it does not believe pre-release government vetting should become the industry standard, noting that it delays access to critical tools for developers, enterprises, and cybersecurity professionals globally.

For engineers and LLM system architects, infrastructure cost and predictability remain pivotal. The pricing structure for GPT-5.6 per 1 million tokens is as follows:

GPT-5.6 also introduces more predictable prompt caching with a minimum 30-minute cache time-to-live (TTL). Cache writes are billed at 125% of the base input price, while cache reads enjoy a substantial 90% discount.

For raw throughput needs, the true game-changer arrives this July with the launch of GPT-5.6 Sol on the Cerebras platform. Powered by dedicated wafer-scale hardware, the model is expected to achieve speeds of up to 750 tokens per second, unlocking entirely new horizons for real-time, zero-latency AI applications.

Want to receive the most critical updates from the world of AI, search engine algorithms, and modern web technologies delivered straight to your inbox? Sign up for the Delante newsletter and never miss an industry breakthrough!

Source: https://openai.com/pl-PL/index/previewing-gpt-5-6-sol/