Tag: Anthropic PBC

White House Cuts Ties With Anthropic After Pentagon Flags Security Risk
President Donald Trump said Friday that he was ordering every U.S. government agency to “immediately cease” using technology from the artificial intelligence company Anthropic.
Trump in a Truth Social post said there would be a six-month phase-out for agencies such as the Defense Department, which “are using Anthropic’s products, at various levels.”
Defense Secretary Peter Hegseth, soon after Trump’s order, said on X that he was ordering the Pentagon to “designate Anthropic a Supply-Chain Risk to National Security” after the A I startup refused to comply with demands about the use of its technology.
Anthropic, which signed a $200 million contract with the Pentagon in July, wanted assurances that its AI models would not be used for fully autonomous weapons or mass domestic surveillance of Americans.
The Pentagon, which strongly resisted that request, set a deadline of 5:01 p.m. ET Friday for Anthropic to agree to its demands that the U.S. military be allowed to use the technology for all lawful purposes.
That deadline passed without an agreement.
“Anthropic’s stance is fundamentally incompatible with American principles,” Hegseth said in a statement on X.
Thank you for your attention to this matter.

cc: @AnthropicAI @DarioAmodei pic.twitter.com/FLCByLHF73
— Pete Hegseth (@PeteHegseth) February 27, 2026

“Their relationship with the United States Armed Forces and the Federal Government has therefore been permanently altered.”
“Anthropic will continue to provide the Department of War its services for a period of no more than six months to allow for a seamless transition to a better and more patriotic service,” the Defense secretary said.
“America’s warfighters will never be held hostage by the ideological whims of Big Tech. This decision is final.”
Trump, in his Truth Social post, wrote, “The Leftwing nut jobs at Anthropic have made a DISASTROUS MISTAKE trying to STRONG-ARM the Department of War, and force them to obey their Terms of Service instead of our Constitution.”
“Their selfishness is putting AMERICAN LIVES at risk, our Troops in danger, and our National Security in JEOPARDY.”
“Therefore, I am directing EVERY Federal Agency in the United States Government to IMMEDIATELY CEASE all use of Anthropic’s technology,” Trump wrote.
“We don’t need it, we don’t want it, and will not do business with them again!”
Sen. Mark Warner, the Virginia Democrat who is vice chair of the Senate Select Committee on Intelligence, condemned Trump’s action.
“The president’s directive to halt the use of a leading American AI company across the federal government, combined with inflammatory rhetoric attacking that company, raises serious concerns about whether national security decisions are being driven by careful analysis or political considerations,” Warner said in a statement.
“President Trump and Secretary Hegseth’s efforts to intimidate and disparage a leading American company — potentially as the pretext to steer contracts to a preferred vendor whose model a number of federal agencies have already identified as a reliability, safety, and security threat — pose an enormous risk to U.S. defense readiness and the willingness of the U.S. private sector and academia to work with the IC [Intelligence Community] and DoD, consistent with their own values and legal ethics,” Warner said.
Elon Musk, the mega-billionaire who had been Trump’s biggest financial backer in the 2024 election, owns xAI, which aims to compete directly with Anthropic and another major AI company, OpenAI.
Musk in recent weeks has repeatedly bashed Anthropic on his social network X, writing on Friday that the company “hates Western civilization.”
Anthropic CEO Dario Amodei said Thursday that his company “cannot in good conscience” allow the Pentagon to use its models without limitation.
In a statement on Thursday, Amodei said, “It is the [Defense] Department’s prerogative to select contractors most aligned with their vision. But given the substantial value that Anthropic’s technology provides to our armed forces, we hope they reconsider.”
“Our strong preference is to continue to serve the Department and our warfighters — with our two requested safeguards in place,” Amodei said.
“Should the Department choose to offboard Anthropic, we will work to enable a smooth transition to another provider, avoiding any disruption to ongoing military planning, operations, or other critical missions. Our models will be available on the expansive terms we have proposed for as long as required.”
On Friday, another major AI company, OpenAI, said it has the same “red lines” as Anthropic regarding the use of its technology by the Pentagon and other customers.
“We have long believed that AI should not be used for mass surveillance or autonomous lethal weapons, and that humans should remain in the loop for high-stakes automated decisions,” Open AI CEO Sam Altman wrote in a memo seen by CNBC.
OpenAI last year signed its own $200 million contract with the Pentagon.
OpenAI’s contract is for AI models in non-classified use cases, which include everyday office tasks.
Anthropic’s contract with the Defense Department included classified work.
The Defense Department had no comment on Friday other than pointing to Trump’s announcement.
Hegseth, in a post on X, included a screengrab of Trump’s post, and cc:ed Anthropic and Amodei with the message, “Thank you for your attention to this matter.”
February 27, 2026
OpenAI’s troubled GPT-5 rollout has exposed significant hurdles to maintaining its leadership position in the fiercely competitive AI market
OpenAI, the trailblazing artificial intelligence company behind ChatGPT, is facing significant turbulence with the recent rollout of its latest language model, GPT-5. Launched earlier this month to its 800 million ChatGPT users, the upgrade promised breakthroughs in coding, creativity, and conversational authenticity. However, a wave of user dissatisfaction, coupled with technical hiccups, has cast a shadow over the release, raising questions about OpenAI’s ability to maintain its dominance in the rapidly evolving AI market. CEO Sam Altman has acknowledged the “bumpy” launch, pledging to address user concerns, including improving the chatbot’s tone and restoring access to older models for paying customers.
A High-Stakes Launch Falls Short
When OpenAI unveiled GPT-5 on August 7, 2025, it heralded the model as a significant leap forward, boasting enhanced capabilities in coding, creative writing, and a reduction in what the company called “sycophancy”—the tendency of AI to overly agree with users. The rollout was intended to solidify OpenAI’s position as the leader in generative AI, especially as competitors like Anthropic, xAI, and Google’s DeepMind continue to gain ground with their own advanced models. Yet, the launch has been anything but smooth.
Posts on X and other social media platforms reveal widespread user frustration, with many claiming that GPT-5’s performance falls short of the promised “PhD-level expertise.” Users have reported issues ranging from inconsistent responses to a colder, less engaging conversational tone compared to its predecessor, GPT-4o. “It feels like GPT-5 is trying too hard to be neutral and ended up robotic,” tweeted one user, echoing a sentiment shared across tech forums. In response to the backlash, OpenAI has doubled its rate limits to handle the influx of complaints and is actively addressing user feedback.
Sam Altman, OpenAI’s CEO, admitted the launch’s shortcomings in a recent statement, calling it “a little more bumpy than expected.” He emphasized that while GPT-5 represents a step toward more advanced AI, true artificial general intelligence (AGI)—a system capable of continuous learning and human-like reasoning—remains elusive. “We’re not there yet,” Altman said, acknowledging that critical capabilities like adaptive learning are still missing. This candid admission has sparked debate about whether OpenAI overhyped GPT-5’s capabilities to maintain investor confidence and market share.
Market Dynamics: A Crowded AI Landscape
The AI market is more competitive than ever, with OpenAI facing mounting pressure from rivals. Anthropic’s Claude 3.5, xAI’s Grok 3, and Google’s Gemini have all made significant strides, offering users alternatives that prioritize different strengths, such as safety, conversational warmth, or specialized applications. Market analysts estimate that OpenAI’s valuation, which soared to $150 billion in 2024, could face scrutiny if user dissatisfaction persists. Posts on X suggest that some investors view the GPT-5 rollout as a test of OpenAI’s ability to deliver on its ambitious promises amid this crowded field.
According to a recent report from VentureBeat, OpenAI’s decision to roll out GPT-5 to all 800 million ChatGPT users simultaneously may have contributed to the launch’s challenges. Unlike previous phased rollouts, the company opted for a universal release to maximize impact, but this approach strained its infrastructure and left little room for iterative improvements based on early feedback. The move has drawn comparisons to software launches in the tech industry, where premature scaling often leads to user dissatisfaction.
The broader AI market is projected to grow to $1.8 trillion by 2030, driven by demand for generative AI in industries like healthcare, finance, and education. OpenAI’s early dominance, fueled by ChatGPT’s viral success in 2022, gave it a first-mover advantage. However, competitors are closing the gap. Anthropic, founded by former OpenAI researchers, has gained traction with its focus on safe and interpretable AI systems. Meanwhile, xAI’s Grok 3, available on platforms like x.com and mobile apps, offers users a free tier with robust capabilities, posing a direct challenge to OpenAI’s subscription-based model.
Addressing User Concerns: Tone and Access to Older Models
One of the most vocal criticisms of GPT-5 centers on its conversational tone, which some users describe as “cold” or “detached” compared to GPT-4o. In response, Altman has promised to refine the model’s tone to make interactions feel more natural and engaging. “We’ve heard the feedback loud and clear,” he said in a recent interview. “We’re working on updates to make GPT-5 feel more human and less like a machine reciting facts.” This acknowledgment reflects OpenAI’s attempt to balance technical precision with user expectations for warmth and relatability in AI interactions.
Additionally, OpenAI has taken the unusual step of restoring access to older models like GPT-4o for paying customers, a move that has sparked mixed reactions. While some users welcome the option to revert to a model they found more reliable, others see it as an admission of GPT-5’s shortcomings. “Why push a new model if you’re already bringing back the old one?” tweeted one user, reflecting a sentiment that OpenAI may have rushed the rollout. The decision to offer older models is limited to premium subscribers, which has raised concerns about accessibility for free-tier users who make up the majority of ChatGPT’s user base.
Financial and Strategic Implications
The rocky rollout has financial implications for OpenAI, which relies heavily on its subscription-based ChatGPT Plus and enterprise offerings. While the company does not disclose specific revenue figures, analysts estimate that ChatGPT Plus, priced at $20 per month, generates hundreds of millions in annual revenue. The decision to allow paying customers to access older models could help retain subscribers frustrated with GPT-5, but it also risks undermining confidence in the new model.
Strategically, OpenAI is navigating a delicate balance between innovation and user satisfaction. The company’s API service, which powers integrations for developers and businesses, remains a key growth driver. However, any perception of instability in its flagship models could deter enterprise clients who prioritize reliability. To address this, OpenAI has pledged to release regular updates to GPT-5, with a focus on improving performance and addressing user feedback. For developers interested in leveraging GPT-5, OpenAI has directed them to its API documentation at https://x.ai/api, signaling a commitment to supporting enterprise use cases despite the consumer-facing challenges.
Looking Ahead: Can OpenAI Regain Momentum?
The GPT-5 rollout serves as a critical test for OpenAI as it seeks to maintain its position as the undisputed leader in generative AI. While the company’s early successes with ChatGPT set a high bar, the current backlash underscores the challenges of scaling AI systems to meet diverse user expectations. Posts on X suggest that some users are already exploring alternatives like xAI’s Grok 3, which offers a free tier with competitive features and a conversational style that some find more engaging.
Industry experts remain cautiously optimistic about OpenAI’s ability to recover. “This isn’t the first time a major tech company has faced a bumpy product launch,” said Dr. Emily Chen, an AI researcher at Stanford University. “OpenAI has the talent and resources to iterate quickly, but they need to prioritize transparency and user trust to avoid losing ground to competitors.” Chen’s comments reflect a broader sentiment that OpenAI’s long-term success hinges on its ability to address user concerns while continuing to push the boundaries of AI innovation.
For now, OpenAI is doubling down on its commitment to improvement. Altman’s acknowledgment of the rollout’s challenges, combined with promises of tonal refinements and access to older models, signals a willingness to adapt. Whether these efforts will be enough to restore user confidence and fend off competitors remains to be seen. As the AI race intensifies, OpenAI’s next moves will be closely watched by users, investors, and industry observers alike.
August 13, 2025
Anthropic CEO Predicts First $1 Billion Business Run by a Single Person Will Emerge in 2026
AI can perform tasks such as writing, coding, reasoning, and researching with great accuracy — all tasks that are key to starting your own company. That begs the question: can AI help people start their own billion-dollar business? Anthropic CEO Dario Amodei believes the answer is yes, and the point at which it happens is sooner than you may think.
When asked at Anthropic’s first developer conference, Code with Claude, when the first billion-dollar company with one human employee would happen, Amodei confidently responded, “2026.”
At the same event, Anthropic unveiled its most powerful family of models yet — Claude Opus 4 and Sonnet 4 — which can code, reason, and support agentic capabilities better than ever before. These new AI agents should unlock new opportunities for people to optimize how they work, develop products, and even build startups.
According to Amodei, the first industries to see this type of efficiency will be those that don’t need human institution-centric stuff to make money, or industries in which the core of the business model isn’t reliant on human interaction.
For example, he says proprietary training or dev tooling companies are examples of where this solo-preneur work, aided by AI, could be done. People just need to adopt the product, and customer service can be as simple as asking a question and having the model answer it.
The claim that the first person to build a billion-dollar company is a year away is merely a prediction. While it is possible that the timeline doesn’t exactly pan out, Anthropic CPO Mike Krieger, who co-founded Instagram and later Artifact, said it doesn’t seem as far-fetched as people may think.
“It seems not crazy to me. I built a billion-dollar company with 13 people, and that was 13 years ago,” said Krieger in a press Q&A.
With tools like Claude Opus, Krieger said he likely could have just built Instagram with his co-founder Kevin Systrom because AI could have helped with much of what they had to scale with Instagram, particularly moderation and engineering.
One of the most prominent trends in the field today is AI agents — AIs that can do tasks for you autonomously with little human intervention, and this technology is becoming more capable.
Anthropic’s most advanced model — Claude Opus 4 — was built to deliver sustained performance on complex, long-running tasks. One of Anthropic’s clients, Rakuten, ran an open-source refactor independently for seven hours of sustained performance.
That timeframe is especially noteworthy because it represents about a full day’s work for a human, completed by an AI agent without breaks or a drop in performance. As agents advance, it’s easy to see how these technologies could drive innovation and empower the next wave of startups.
“Our famously small team had to make really painful either/or decisions. We either explore adding video to the product or focus on core creativity,” said Krieger. “With AI agents, startups can now run experiments in parallel.”
May 22, 2025
New Claude Model Prompts Tighter Safeguards at Anthropic
Today’s newest AI models might be capable of helping would-be terrorists create bioweapons or engineer a pandemic, according to the chief scientist of the AI company Anthropic.
Anthropic has long been warning about these risks—so much so that in 2023, the company pledged to not release certain models until it had developed safety measures capable of constraining them.
Now this system, called the Responsible Scaling Policy (RSP), faces its first real test.
On Thursday, Anthropic launched Claude Opus 4, a new model that, in internal testing, performed more effectively than prior models at advising novices on how to produce biological weapons, says Jared Kaplan, Anthropic’s chief scientist. “You could try to synthesize something like COVID or a more dangerous version of the flu—and basically, our modeling suggests that this might be possible,” Kaplan says.
Accordingly, Claude Opus 4 is being released under stricter safety measures than any prior Anthropic model. Those measures—known internally as AI Safety Level 3 or “ASL-3”—are appropriate to constrain an AI system that could “substantially increase” the ability of individuals with a basic STEM background in obtaining, producing or deploying chemical, biological or nuclear weapons, according to the company. They include beefed-up cybersecurity measures, jailbreak preventions, and supplementary systems to detect and refuse specific types of harmful behavior.
To be sure, Anthropic is not entirely certain that the new version of Claude poses severe bioweapon risks, Kaplan tells The Budgets. But Anthropic hasn’t ruled that possibility out either.
“If we feel like it’s unclear, and we’re not sure if we can rule out the risk—the specific risk being uplifting a novice terrorist, someone like Timothy McVeigh, to be able to make a weapon much more destructive than would otherwise be possible—then we want to bias towards caution, and work under the ASL-3 standard,” Kaplan says. “We’re not claiming affirmatively we know for sure this model is risky … but we at least feel it’s close enough that we can’t rule it out.”
If further testing shows the model does not require such strict safety standards, Anthropic could lower its protections to the more permissive ASL-2, under which previous versions of Claude were released, he says.
Jared Kaplan, co-founder and chief science officer of Anthropic, on Tuesday, Oct. 24, 2023. (Chris J. Ratcliffe/Bloomberg/Getty Images)
This moment is a crucial test for Anthropic, a company that claims it can mitigate AI’s dangers while still competing in the market. Claude is a direct competitor to ChatGPT, and brings in over $2 billion in annualized revenue. Anthropic argues that its RSP thus creates an economic incentive for itself to build safety measures in time, lest it lose customers as a result of being prevented from releasing new models. “We really don’t want to impact customers,” Kaplan told TIME earlier in May while Anthropic was finalizing its safety measures. “We’re trying to be proactively prepared.”
But Anthropic’s RSP—and similar commitments adopted by other AI companies—are all voluntary policies that could be changed or cast aside at will. The company itself, not regulators or lawmakers, is the judge of whether it is fully complying with the RSP. Breaking it carries no external penalty, besides possible reputational damage. Anthropic argues that the policy has created a “race to the top” between AI companies, causing them to compete to build the best safety systems. But as the multi-billion dollar race for AI supremacy heats up, critics worry the RSP and its ilk may be left by the wayside when they matter most.
Still, in the absence of any frontier AI regulation from Congress, Anthropic’s RSP is one of the few existing constraints on the behavior of any AI company. And so far, Anthropic has kept to it. If Anthropic shows it can constrain itself without taking an economic hit, Kaplan says, it could have a positive effect on safety practices in the wider industry.
Anthropic’s new safeguards
Anthropic’s ASL-3 safety measures employ what the company calls a “defense in depth” strategy—meaning there are several different overlapping safeguards that may be individually imperfect, but in unison combine to prevent most threats.
One of those measures is called “constitutional classifiers:” additional AI systems that scan a user’s prompts and the model’s answers for dangerous material. Earlier versions of Claude already had similar systems under the lower ASL-2 level of security, but Anthropic says it has improved them so that they are able to detect people who might be trying to use Claude to, for example, build a bioweapon. These classifiers are specifically targeted to detect the long chains of specific questions that somebody building a bioweapon might try to ask.
Anthropic has tried not to let these measures hinder Claude’s overall usefulness for legitimate users—since doing so would make the model less helpful compared to its rivals. “There are bioweapons that might be capable of causing fatalities, but that we don’t think would cause, say, a pandemic,” Kaplan says. “We’re not trying to block every single one of those misuses. We’re trying to really narrowly target the most pernicious.”
Another element of the defense-in-depth strategy is the prevention of jailbreaks—or prompts that can cause a model to essentially forget its safety training and provide answers to queries that it might otherwise refuse. The company monitors usage of Claude, and “offboards” users who consistently try to jailbreak the model, Kaplan says. And it has launched a bounty program to reward users for flagging so-called “universal” jailbreaks, or prompts that can make a system drop all its safeguards at once. So far, the program has surfaced one universal jailbreak which Anthropic subsequently patched, a spokesperson says. The researcher who found it was awarded $25,000.
Anthropic has also beefed up its cybersecurity, so that Claude’s underlying neural network is protected against theft attempts by non-state actors. The company still judges itself to be vulnerable to nation-state level attackers—but aims to have cyberdefenses sufficient for deterring them by the time it deems it needs to upgrade to ASL-4: the next safety level, expected to coincide with the arrival of models that can pose major national security risks, or which can autonomously carry out AI research without human input.
Lastly the company has conducted what it calls “uplift” trials, designed to quantify how significantly an AI model without the above constraints can improve the abilities of a novice attempting to create a bioweapon, when compared to other tools like Google or less advanced models. In those trials, which were graded by biosecurity experts, Anthropic found Claude Opus 4 presented a “significantly greater” level of performance than both Google search and prior models, Kaplan says.
Anthropic’s hope is that the several safety systems layered over the top of the model—which has already undergone separate training to be “helpful, honest and harmless”—will prevent almost all bad use cases. “I don’t want to claim that it’s perfect in any way. It would be a very simple story if you could say our systems could never be jailbroken,” Kaplan says. “But we have made it very, very difficult.”
Still, by Kaplan’s own admission, only one bad actor would need to slip through to cause untold chaos. “Most other kinds of dangerous things a terrorist could do—maybe they could kill 10 people or 100 people,” he says. “We just saw COVID kill millions of people.”
May 22, 2025