OpenAI Launches GPT-5.6 Series: Sol, Terra, and Luna
On July 10, OpenAI officially released the GPT-5.6 series, comprising three models: the flagship Sol, the balanced Terra, and the speed/cost-focused Luna. All three are now rolling out to ChatGPT, Codex, and OpenAI API users.
Pricing & Performance
Sol costs $5/$30 per million tokens (input/output), Terra $2.5/$15, and Luna $1/$6.
The GPT-5.6 series introduces Max reasoning intensity and Ultra mode. In Ultra mode, the system coordinates 4 agents in parallel to handle complex tasks. According to OpenAI's benchmarks, GPT-5.6 Sol scores 88.8% on Terminal-Bench 2.1, climbing to 91.9% with Ultra mode enabled.
ChatGPT Work and Desktop App Consolidation
Alongside the new models, OpenAI launched ChatGPT Work, an agent product designed for long-running workflows. It can pull information from apps and files, conduct research and analysis, and generate documents, spreadsheets, presentations, reports, and websites.
On the product side, the Codex desktop app has been upgraded to a new ChatGPT desktop application with three entry points: Chat, Work, and Codex. The original ChatGPT desktop app is now called ChatGPT Classic and will continue to receive model updates and security maintenance. The new app is available on macOS and Windows.
OpenAI reports that Codex has over 5 million weekly active users, with more than 1 million using it for non-software-development purposes. Internally, OpenAI uses ChatGPT Work and Codex for finance, data整理, and document production.
The ChatGPT Atlas browser will be discontinued on August 9, with its browser agent capabilities moving to ChatGPT and Codex. ChatGPT Work is initially available to Pro, Enterprise, and Edu users, with Plus and Business users gaining access within days.
xAI Releases Grok 4.5: First Programming Agent Model
On July 9, SpaceXAI released Grok 4.5, its first model specifically trained for programming and agent tasks. The model was developed jointly with Cursor.
Grok 4.5 is designed for real-world engineering scenarios, handling large codebases and long-running tasks spanning multiple repositories, tools, and hundreds of skills. The model was trained on tens of thousands of NVIDIA GB300 GPUs. Data processing involved deduplication, quality scoring, and domain focusing; reinforcement learning covered hundreds of thousands of tasks with a focus on multi-step software engineering.
Elon Musk called it an "Opus-level model," claiming a good balance between intelligence, speed, and cost efficiency.
ByteDance Launches Seedream 5.0 Pro
ByteDance also released Seedream 5.0 Pro, an multimodal image generation model and the latest upgrade to the Seedream series. Technical specifications and availability details are yet to be fully disclosed.