GPT-5.6: Three models, one goal — get the job done, not just write back

On July 9, OpenAI released GPT-5.6 to all users and split it into three separate model lines: Sol (flagship), Terra (balanced performance at a lower price), and Luna (the cheapest variant). This isn't marketing — it's a fundamental architectural shift. Models are no longer a single universal window, but a spectrum of specialized engines, each optimized for a different price-to-performance ratio.

The numbers are impressive. GPT-5.6 Sol scored 53.6 points on Agents' Last Exam, which is 13.1 points more than Claude Fable 5. On the Artificial Analysis Coding Agent Index, it scored 80 points — a new record. And most interestingly: Terra and Luna outperform Fable 5 at roughly one-sixteenth the cost. For the first time, OpenAI has explicitly bet on the "performance per dollar" metric rather than absolute scores.

The most technically significant innovation is Programmatic Tool Calling — the model's ability to write and run its own programs that coordinate tools, filter intermediate results, and autonomously decide the next step. Instead of the developer having to code every action and send every response back to the model, GPT-5.6 handles it itself with fewer tokens and queries. According to Cisco CTO Arjuna Sambamoorthy, "GPT-5.6 consistently stays focused on long tasks and arrives at quality solutions with minimal guidance."

The new Ultra mode then takes the entire philosophy to the extreme: it coordinates four agents in parallel, working on different parts of a task simultaneously. On Terminal-Bench 2.1, it achieves 91.9% — that's 3.9 points more than Claude Mythos 5.

Pricing: Sol costs $5 per million input / $30 per million output tokens, Terra $2.50 / $15, and Luna $1 / $6.

GPT-Live and ChatGPT Work: A voice that hears you even while it speaks, and an agent that works for hours

Just a day earlier, on July 8, OpenAI launched GPT-Live — a new generation of voice models built on a fully duplex architecture. Unlike the previous Advanced Voice Mode, which worked on a "I speak — you speak" principle, GPT-Live can listen and speak simultaneously. During a conversation, it can interrupt you, say "mmhmm," or stay silent while you think.

Even more important is the architectural change under the hood: GPT-Live separates continuous conversation from deep thinking. When you ask a complex question, it delegates it to GPT-5.5 in the background while continuing to talk with you. In the future, the background model will be continuously updated to the latest version. For ChatGPT users, deployment is free — Free users get GPT-Live-1 mini, while paying users get the full GPT-Live-1 version. Czech is not yet among the primarily optimized languages, but voice control in English works even from the Czech Republic.

ChatGPT Work, also launched alongside GPT-5.6, completes the transition from a chat window to a production tool. It can work across connected applications (Slack, Notion, Google Drive, Microsoft 365), stay on a project for hours, and create editable documents, spreadsheets, presentations, and interactive websites. According to Shopify, "ChatGPT Work functions as an operating layer for AI — it pulls context from Slack and active projects every day, turns it into tasks, and helps run repeatable work."

Grok 4.5 and Muse Spark 1.1: Competition pushes on price and capabilities

Elon Musk's SpaceXAI (formerly xAI) released Grok 4.5 the same week — a model focused on coding, agentic tasks, and knowledge work. The new version brings the ability to generate entire applications and complex product outputs. Grok's key weapon is an aggressive pricing policy that pushes the entire market toward cheaper execution. Grok is available via API and the X platform, but does not yet have a localized interface in the Czech Republic.

Meta, meanwhile, introduced Muse Spark 1.1 — the second version of its multimodal model, which combines a million-token context with visual perception, coding, computer control, and multi-agent orchestration. A notable feature is active context management: the model can compress long sessions without losing important state and decides on its own whether to program an action or execute it directly in the interface.

Strategically, Meta has for the first time monetized its model through the Meta Model API. It's no longer just about releasing open-weight models — Meta wants to sell measured intelligence as a service, just like OpenAI or Anthropic.

What this puzzle means for the average user — and for the Czech Republic

The individual announcements add up to a coherent strategy. Frontier labs are vertically integrating models, voice interfaces, agents, browsers, desktop environments, and layers for artifact creation. They are fighting for ownership of the loop between intent and outcome.

For the average user, this means one simple thing: AI will soon not be something you "chat with," but something that actually works for you. You state a goal — you get a finished presentation, analyzed data, a prepared report, a programmed application. Chat fades into the background and becomes mere infrastructure, much like you don't think about TCP/IP protocols when you open a web page today.

But with that come new risks. A hallucinated paragraph in a chat is annoying. A hallucinated workflow that runs through your CRM, file system, or financial model — now that's an incident. Long-running agents need permissions, an audit trail, checkpoints, and the ability to roll back changes. OpenAI and Anthropic are working on it, but it remains an unsolved problem for now.

For Czech companies, there are two takeaways. First, the cost accessibility of agentic AI is dropping radically — Luna from GPT-5.6 costs $1 per million input tokens, opening the door even for smaller firms. Second, the language barrier remains — none of the new models have Czech among their priority languages, although basic understanding of Czech is at a decent level across all of them. Czech companies that want to deploy agentic AI into customer-facing processes will have to reckon with the fact that tuning Czech prompts and outputs will require their own engineering effort.

The era of chatbots isn't ending. It's just compiling itself into the infrastructure on which a whole new generation of applications will run — and whoever builds it first will dictate the rules of the game.