Key Takeaways
- OpenAI unveiled its newest flagship model, GPT‑5.6 Sol, and a workplace‑focused AI agent called ChatGPT Work, which can operate spreadsheets, calendars, email and other software on behalf of users.
- The release intensifies the rivalry with Anthropic, whose leading model Fable 5 shows comparable performance on standard benchmarks.
- Independent evaluator Vals AI rates Sol as “state of the art” for real‑world financial and legal tasks, even though it matches Anthropic’s Fable 5 in synthetic tests.
- The Trump administration initially sought to restrict the models over cybersecurity worries, signing an executive order that asked firms to give the government pre‑release oversight.
- After negotiations, both OpenAI and Anthropic were permitted to launch their technologies publicly, despite early fears that the models posed national‑security threats.
- Security experts largely deemed those fears overstated, noting that the dual‑use nature of the models—capable of both attacking and defending systems—requires careful guardrails rather than outright bans.
- OpenAI adopted a more open stance, providing Sol with fewer restrictions so businesses and individuals can use it for self‑defense, whereas Anthropic tightened guardrails on Fable 5 to block potentially harmful outputs.
- When Anthropic’s guardrails trigger, the model falls back to a less powerful system, Claude Opus 4.8, which still offers limited capability.
- Both Sol and Fable 5 are costly—Sol is sometimes more expensive than Fable 5, which itself runs about twice the price of Anthropic’s next‑tier model.
- Meta also released a new AI model on the same day, but it is markedly less powerful and far cheaper than the offerings from OpenAI and Anthropic.
- The release coincides with an ongoing copyright lawsuit filed by The New by against OpenAI and Microsoft, which both companies deny.
Overview of the Release
On Thursday OpenAI made public its most advanced artificial intelligence system to date, GPT‑5.6 Sol, together with a new productivity tool named ChatGPT Work. ChatGPT Work is an “AI agent” that can autonomously interact with common office applications—spreadsheets, online calendars, email services and similar software—to perform tasks on behalf of users. The announcement signals OpenAI’s push to move beyond pure language generation toward actionable assistants that can streamline everyday office work.
Escalating Competition with Anthropic
The launch comes amid an intensifying battle for AI supremacy with the rival start‑up Anthropic. Anthropic’s latest flagship, Fable 5, has been positioned as a direct competitor to Sol. Both companies are racing to deliver models that combine higher reasoning ability with broader multimodal competence, and each release is viewed as a strategic move to capture enterprise and research paper.
Benchmark Performance Comparison
According to Vals AI, a firm that monitors the latest AI technologies, Sol’s performance on standard benchmark tests roughly matches that of Anthropic’s Fable 5. While the numbers are close, Vals AI’s chief executive Rayan Krishnan emphasized that Sol’s strengths become more apparent in applied scenarios rather than synthetic tests alone.
Real‑World Applicability
Krishnan noted that in genuine‑world settings involving financial and legal work, Sol operates at a “state of the art” level. This suggests that the model’s improved reasoning, contextual understanding and tool‑use capabilities translate into tangible productivity gains for professionals who rely on accurate data extraction, contract analysis or financial modelling.
Government Intervention and Executive Order
The release followed a series of U.S. government actions aimed at curbing potential AI‑related cybersecurity risks. Although former President Trump had previously taken a laissez‑faire stance toward AI regulation, he signed an executive order last month requesting that tech firms voluntarily submit new models for government review before public release. The order reflected growing anxiety that powerful generative models could be weaponized for hacking, disinformation or other malicious activities.
Initial Restrictions and Subsequent Approval
In line with the order, OpenAI initially said it would share GPT‑5.6 Sol only with a small group of companies cleared by the Trump administration. Anthropic took a comparable approach, limiting access to its own models. After extended discussions, however, the administration reversed course and permitted both firms to release their technologies to the broader public, concluding that the anticipated threats were manageable with appropriate safeguards.
Security Experts’ Assessment
Many independent security specialists argued that the early national‑security concerns were exaggerated. They contended that while the models indeed possess the ability to locate software vulnerabilities—information that could be exploited by attackers—the same capability is invaluable for defensive cybersecurity operations. Consequently, experts advocated for nuanced guardrails rather than outright bans or severe access limitations.
Divergent Guardrail Strategies
OpenAI chose a relatively open path, releasing Sol with fewer restrictions so that enterprises and individual users can employ the model to protect their own systems. In contrast, Anthropic adopted a more defensive posture, installing robust guardrails on Fable 5 that block outputs related to cybersecurity, biology and other sensitive domains. When those guardrails fire, the model redirects the request to a weaker fallback system, Claude Opus 4.8, which retains limited functionality but avoids generating potentially hazardous content.
Anthropic’s Model Lineage and Limitations
Anthropic had previously introduced a more capable system called Claude Mythos in April, but elected not to share it publicly due to fears of enabling offensive cyber operations. Instead, the company released a toned‑down version, Fable 5, last month. Shortly thereafter, the Trump administration ordered Anthropic to suspend access for all foreign nationals, citing national‑security worries. To comply, Anthropic took the entire model offline, later re‑introducing it with the added guardrails described above.
Dual‑Use Nature of the Technology
Both Sol and Fable 5 excel at pinpointing weaknesses in the software that underpins the internet—a skill that can be turned either toward launching attacks or fortifying defenses. This dual‑use characteristic means that limiting access solely to prevent offensive misuse may also impede legitimate defensive efforts by businesses, governments and critical‑infrastructure operators. The challenge lies in crafting policies that allow beneficial use while mitigating risk.
Cost Considerations
Pricing remains a significant factor for adoption. Krishnan indicated that Sol is sometimes more expensive than Anthropic’s Fable 5, and that Fable 5 itself costs roughly twice as much as Anthropic’s next‑most‑powerful model. Consequently, organizations must weigh the performance gains against the higher operational expenses, especially when considering large‑scale deployment across multiple teams or geographies.
Meta’s Parallel Release
On the same day as OpenAI’s announcement, Meta unveiled its own new AI model. Meta’s offering is markedly less powerful than either Sol or Fable 5 and is correspondingly cheaper. While it may appeal to cost‑sensitive users or experimental projects, it does not directly compete with the high‑end capabilities that OpenAI and Anthropic are marketing to enterprise customers.
Ongoing Legal Challenges
The release occurs against the backdrop of a copyright lawsuit filed by The New York Times against OpenAI and Microsoft, alleging that the companies used Times content without permission to train their AI systems. Both OpenAI and Microsoft have denied the allegations, maintaining that their training practices comply with fair use and existing law. The outcome of this case could have broader implications for how AI developers source training data moving forward.