OpenAI Announces Public Launch of GPT-5.6 Sol, Terra, and Luna as AI Lab Rivalry Heats Up

OpenAI Announces Public Launch of GPT-5.6 Sol, Terra, and Luna as AI Lab Rivalry Heats Up

OpenAI confirmed today that GPT-5.6 Sol, along with two companion models named Terra and Luna, will launch publicly this Thursday. The company is also expanding preview access globally ahead of the launch.

The release marks the first time OpenAI's next-generation model family will be available to the general public. GPT-5.6 Sol was originally previewed in late June. A variant called Sol Ultra was already made available in Codex, OpenAI's coding agent. Thursday's launch opens the full model family to a wider audience.

What We Know About Sol, Terra, and Luna So Far

OpenAI has not published detailed specs for each model. Context from the company's earlier preview and from independent evaluations gives a clearer picture.

GPT-5.6 Sol is the flagship model. METR, an independent research nonprofit that evaluates frontier AI systems, published a predeployment evaluation in late June. The report found that Sol is highly capable at autonomous task completion. Its 50 percent Time Horizon, the task duration it can complete half the time, sits at roughly 11 hours. That puts it in the same league as Anthropic's most advanced models.

Early testers on Hacker News describe Sol as incredibly capable. One user said it "fixed all the problems I had with GPT-5.5." Another said the model is "incredibly determined" and will run tasks for a full day without needing explicit goals. A third said Sol is faster and cheaper per task than Anthropic's Fable.

Terra and Luna appear to be smaller or specialized models. They are likely positioned to serve different price and performance tiers. This multi-model strategy mirrors what Anthropic has done with its Claude lineup, where different models target different use cases and budgets.

Anthropic's Countermove: Claude Cowork Goes Mobile

The OpenAI announcement arrived on the same day Anthropic made a significant product play of its own. Claude Cowork, Anthropic's agentic AI tool for general knowledge work, is expanding to mobile and web.

Cowork was previously only available as a desktop app. It can now run tasks across devices and continue working in the background, even when a laptop is closed. The expansion opens the tool to users who could not install the desktop app.

Anthropic released usage data from 1.2 million Cowork sessions across more than 600,000 organizations. The numbers challenge the common assumption that AI agents are primarily for coding. Software development accounted for only 8.7 percent of Cowork usage. The two largest categories were business process operations at 33.4 percent and content creation at 16.4 percent.

This data suggests that AI agents are being adopted fastest for the administrative and operational work that keeps companies running. Reconciliation, reporting, drafting, updating. These tasks consume hundreds of hours per month in most organizations.

The product expansion positions Cowork as a general workplace tool. Anthropic calls it a tool for "the work around the work." The timing is notable given OpenAI's Thursday launch. Both labs are betting the AI market will be won by whoever owns the platforms where work gets done, not by whoever has the best chatbot.

A Security Warning for Agentic AI

Today also served as a reminder that pushing AI agents into workplace workflows brings real security risks.

Noma Security published a vulnerability called GitLost. It allows attackers to trick GitHub's Agentic Workflows into leaking private repository data. The attack uses prompt injection. An attacker posts a crafted GitHub issue in a public repository. When an AI agent processes the issue, hidden instructions cause it to fetch data from private repositories belonging to the same organization and post it publicly.

The vulnerability works because the AI agent treats user-controlled content as trusted instructions. Noma Security's researchers demonstrated that GitHub's guardrails failed to prevent the leak. This is a textbook agentic AI security problem. The agent's context window is also its attack surface. Any content it reads can be weaponized if the model treats that content as instructions.

GitHub was notified and the vulnerability was responsibly disclosed. But the finding should give any organization pause before deploying AI agents with broad access to internal systems.

How to Act on Today's Developments

The public launch of GPT-5.6 Sol, Terra, and Luna means more model choices and likely lower prices. Competitive pressure between OpenAI and Anthropic has been driving costs down across the industry. If you have been evaluating which model family to standardize on, the next week will give you more data points.

The Claude Cowork data confirms something we have seen with our own clients at Triweb AI. The highest-value AI use cases in most organizations are not cutting-edge coding applications. They are the routine business processes that consume the most time. This is where AI agents can deliver the fastest return on investment.

The GitLost vulnerability is a concrete example of a risk we raise with every client deploying AI agents. When an agent has access to internal data, the security model shifts. Traditional access controls are not enough. You need to audit what content the agent reads, what it can do with that content, and how instructions from untrusted sources are handled.

Thursday's model launch, Anthropic's mobile expansion, and the GitLost disclosure share a common thread. AI agents are moving from experimental tools to mainstream workplace infrastructure. The companies that treat them as such, with the right security practices and realistic use-case planning, will get the most value.