OpenAI Unleashes GPT-5.6 and ChatGPT Work: A Direct Assault on Anthropic's Price and Speed

The artificial intelligence landscape has rapidly shifted from a race over raw parameter sizes to a grueling war of economic attrition and workflow integration. Just as competitor Anthropic began capturing enterprise mindshare with its highly publicized Claude Cowork platform and the White House discussions surrounding its Fable 5 and Mythos 5 models, OpenAI has delivered a massive counter-offensive. By launching a suite of three new celestial-themed models—GPT-5.6 Sol, Terra, and Luna—alongside a rebranded desktop-automation ecosystem called ChatGPT Work, OpenAI is making an aggressive, calculated play to undercut its rivals on price, execution speed, and practical office utility.

Quick summary

• Celestial Suite Debut: OpenAI has released GPT-5.6 Sol, Terra, and Luna, specifically optimized to beat Anthropic's flagship models on speed, performance, and operational cost.
• ChatGPT Work Launches: This new desktop automation assistant can access browsers and local files to manage multi-step, multi-hour projects for non-coding professionals.
• Price and Efficiency War: In benchmark tests, the midrange GPT-5.6 Sol model matched or exceeded Claude Fable 5 while operating at roughly 61% faster speeds and one-quarter of the estimated run cost.

Why it matters

This release signals a critical pivot point for the tech sector: AI is evolving from a passive assistant that answers questions into an active agent that works alongside humans. With ChatGPT Work gaining direct access to local file systems, calendars, databases, and communication channels like Slack and Microsoft Teams, the boundary between software and user agency is rapidly eroding.

For businesses, the primary implication is economic. By offering GPT-5.6 Sol at an estimated 25% of the cost of Anthropic’s Claude Fable 5, OpenAI is starting an aggressive price war. Enterprises that previously hesitated to deploy large-scale agentic workflows due to high API costs can now reconsider. However, this level of system-level integration introduces major security and psychological hurdles. Allowing an AI to manipulate local applications, edit spreadsheets, and reorganize private databases requires a degree of trust that many IT departments and individual users may not yet be prepared to yield.

Background

To understand this release, one must look at the recent regulatory and competitive friction in Washington and Silicon Valley. The launch of GPT-5.6 was initially delayed by government oversight committees concerned with safety and potential systemic misuse. This regulatory pause coincided with intense public focus on Anthropic’s White House engagements regarding its own advanced models, Fable 5 and Mythos 5.

Simultaneously, OpenAI faced criticism for its fragmented product design. Previously, its advanced desktop and agentic coding capabilities were siloed within the Codex application. This technical separation made agentic features inaccessible or overly complicated for ordinary business professionals. Recognizing that Anthropic had successfully captured the non-coder market with its intuitive "Claude Cowork" branding, OpenAI pivoted. The company has now merged its Codex and ChatGPT desktop frameworks into a single application, separating the coding-specific agentic tools from the broader, more approachable ChatGPT Work suite.

Breaking Down the Benchmarks: Sol vs. Fable 5

To justify its claims of dominance, OpenAI released performance data across several industry-standard evaluations, positioning GPT-5.6 Sol as a direct "Claude killer." On the Agents' Last Exam benchmark, which tests an AI's ability to handle complex, multi-step workflows across 55 professional fields, GPT-5.6 Sol scored 53.6, outperforming Claude Fable 5 by 13.1 points.

On the Artificial Analysis Intelligence Index, which measures generalized reasoning, Sol finished within a single point of Fable 5. Crucially, however, Sol completed these evaluation tasks in 61% less time than its competitor. For software engineering applications, Sol reached a score of 80 on the Artificial Analysis Coding Agent Index, outperforming Fable 5 by 2.8 points.

Independent evaluations support these efficiency gains. Qodo, an AI code-review platform, conducted internal PR benchmarks comparing the new architecture to GPT-5.5. According to Qodo CEO Itamar Friedman, GPT-5.6 demonstrated superior precision while consuming roughly three times fewer tokens and delivering approximately two times lower median latency. This drastic reduction in token consumption is what allows OpenAI to offer its medium-reasoning Sol model at a fraction of the market rate.

ChatGPT Work: Real-World Enterprise Application

The practical centerpiece of OpenAI’s announcement is ChatGPT Work. Unlike basic chatbots, this tool is designed to connect directly with external business systems, including Google Drive, SharePoint, email clients, CRMs, and project management boards. Users can assign complex tasks, monitor the AI’s step-by-step progress, intervene to redirect focus, and approve critical actions before they are executed.

Early enterprise adoption figures suggest significant potential. Angela Ferrante, head of enterprise marketing at Zapier, noted that her team used ChatGPT Work to build an automated lead-review pipeline. The agent traced customer interactions across Zapier's internal systems, identified communication gaps, and generated weekly dashboards that ultimately uncovered seven figures in previously missed sales opportunities.

However, the tool's deep operating system integration has also sparked concern. During official demonstrations, presenters showed ChatGPT Work accessing Apple Notes to entirely reorganize a user's personal files. While technically impressive, this level of invasive local access has raised red flags among data privacy advocates who worry about security vulnerabilities and accidental file modification.

Qnews24h insight

OpenAI's latest move highlights a broader structural change in how AI products are built and sold. The transition from technical version numbers (like GPT-4 to GPT-5) to distinct, task-oriented model identities (Sol, Terra, Luna) indicates that developers can no longer rely purely on raw scale to win clients. Corporate buyers want specialized, fast, and cost-predictable systems rather than massive, expensive, all-purpose models.

However, the real battleground for the next generation of AI is not benchmark metrics—it is user psychology and operating system permissions. Software capable of automatically editing databases, drafting slideshows, and shifting local files around is only as valuable as the guardrails that prevent it from making catastrophic errors. By branding this release as its "most robustly safeguarded" to date, OpenAI is attempting to soothe corporate anxieties. Yet, the company must prove in real-world scenarios that its agentic tools can perform these invasive local operations without compromising sensitive enterprise data or corrupting user files.

Sources

• Information and metrics sourced from ZDNET.