OpenAI Says GPT-5.4's Uncontrollable Reasoning Is Working as Designed

01Anthropic Said No to the Pentagon. The Pentagon Put It on a List.

Dario Amodei spent two years building Anthropic's reputation as the AI company that would draw lines. Last week, he drew one in public. In an interview reported by TechCrunch, the Anthropic CEO called OpenAI's characterization of its military contracts "straight up lies." It was a sharp departure from the careful, policy-paper language Amodei typically uses. Days later, the Pentagon responded with bureaucratic force: it formally designated Anthropic a supply-chain risk.

The label, reported by the Wall Street Journal, places Anthropic in a category typically reserved for foreign adversaries and compromised vendors. Defense procurement officers must now flag any contract or subcontract that routes through Anthropic's models. Prime contractors using Claude in defense-adjacent products face new disclosure and justification requirements. The designation doesn't ban Anthropic from government work outright. It does something quieter: it makes choosing Anthropic a decision that requires paperwork, review, and risk tolerance that choosing a competitor does not.

Amodei's rhetorical escalation followed a pattern. When Anthropic first declined a Pentagon contract in late 2025, the company framed it as a principled policy choice. Amodei spoke in measured terms about responsible deployment and use-case boundaries. By March 2026, the framing had changed. "Straight up lies" isn't policy language. It's an accusation, directed at a rival that had secured the contract Anthropic walked away from.

That shift moved the conflict from a philosophical disagreement into a public fight. Pentagon officials don't typically intervene in vendor disputes. But a CEO accusing a defense partner of lying about the terms of military work creates a problem the department has to address. The supply-chain risk designation reads less like a security assessment and more like an institutional response to a company that made itself politically inconvenient.

For Anthropic, the commercial fallout extends beyond direct government contracts. Defense integrators, aerospace firms, and intelligence-community vendors all watch the supply-chain risk list. A company on that list becomes a liability in any proposal that touches federal money. Anthropic's enterprise sales team now carries a label that competitors will reference in every procurement review.

Amodei built Anthropic's brand on the premise that safety and commercial success aren't contradictory. The supply-chain designation tests that premise with a concrete cost. Anthropic still has its principles. It also has a new line item in every defense contractor's risk register.

02OpenAI Calls GPT-5.4's Uncontrollable Reasoning a Safety Feature

GPT-5.4 launched Thursday as OpenAI's self-described "most capable frontier model." It scored 83% on GDPval, the company's knowledge-work benchmark, and makes 33% fewer factual errors than GPT-5.2. The API version supports a 1-million-token context window and native computer use. That same week, OpenAI published research concluding reasoning models like GPT-5.4 Thinking cannot control their own chains of thought.

OpenAI says that's a good thing.

The company's open-source CoT-Control evaluation suite tested over 13,000 tasks across frontier reasoning models. It measured whether models could deliberately reshape or obscure their internal reasoning. GPT-5.4 Thinking successfully controlled just 0.3% of its chains of thought at 10,000-character reasoning lengths. No frontier model exceeded 15.4%. OpenAI's logic follows: a model that can't manipulate its reasoning trail can't hide deceptive intent from safety monitors. Low controllability means high monitorability. If the thinking is involuntary and visible, oversight teams can read it.

The GPT-5.4 Thinking System Card backs this up. On agentic misalignment detection, monitorability "approaches 100%," according to the card. OpenAI estimates 99.95% of outputs will not violate its harassment policy before other safety layers activate. The card also notes GPT-5.4 Thinking is the first general-purpose model to include mitigations for high-capability cybersecurity risks.

The same system card, though, complicates the picture. GPT-5.4 Thinking has lower aggregate CoT monitorability than its predecessor, GPT-5 Thinking. The model got harder to monitor overall, even as the company built its safety case on monitoring.

The structural tension runs deeper. OpenAI's framework depends on a chain: visible reasoning, involuntary transparency, effective oversight. "Involuntary" does the heaviest lifting, and CoT-Control's 0.3% figure is the evidence. But the same research shows controllability rises with model scale. GPT-5.4 scored 0.3% where GPT-5.2 scored 0.2%, a small gap moving in the wrong direction.

OpenAI built its safety architecture on a property it measured in current models, not one it engineered. The CoT-Control suite, now open-source, gives outside researchers the tools to verify whether that property holds as models grow.

03Nvidia Exits AI Lab Investing as OpenAI Pushes Into Enterprise Software

Two moves this week from opposite ends of the AI stack point in the same direction. Nvidia said it is done investing in AI model companies, and OpenAI released a ChatGPT plugin for Microsoft Excel. Neither announcement referenced the other. Together, they signal that the AI industry's brief era of vertical entanglement is ending.

Nvidia CEO Jensen Huang told reporters Wednesday that the company's stakes in OpenAI and Anthropic will "likely be the last" such investments. His reasoning was muddled. He cited potential conflicts of interest with Nvidia's customers, but the company has held these positions for years without apparent friction. A second justification, that AI labs no longer need outside capital, sits uneasily beside OpenAI's recent $40 billion fundraise. TechCrunch observed that Huang's explanation "raises more questions than it answers."

What is clearer than his rationale is the strategic direction. Nvidia has been pouring billions into hardware infrastructure: optical interconnects, networking, custom silicon. Holding equity in model companies offered optionality when the industry's structure was uncertain. Now that uncertainty has resolved. Nvidia's moat is hardware, and capital deployed elsewhere is capital not reinforcing it.

OpenAI is making the mirror-image bet. ChatGPT for Excel, announced this week, embeds GPT-5.4 directly into spreadsheet workflows for financial modeling and data analysis. The product targets regulated industries where users need AI inside their existing tools, not through a separate chat interface. It marks a shift from selling API access to selling finished software. OpenAI is no longer content to be infrastructure for other people's products.

The pattern: each layer of the AI stack is retreating to defensible ground. Hardware makers sell hardware; model companies ship applications. The middle position — investing in adjacent layers, providing generic APIs, building platforms that serve everyone — is compressing. Nvidia no longer wants returns from AI labs; it wants those labs to buy more GPUs. The same logic drives OpenAI's Excel play: owning the product beats powering someone else's.