Anthropic stole ChatGPT's aura. Then it gave it back.

Anthropic's annualized revenue surpassed $47 billion by May 2026, up from about $9 billion at the end of 2025, and a Series H round valued the company near $965 billion, according to Reuters. The company, which had overtaken OpenAI as Silicon Valley's favorite AI firm, now faces challenges as its culture and safety stance are tested by rapid growth and a Pentagon dispute.

For a few months in 2026, Anthropic was the gigachad of Silicon Valley. Crude label, real shift in power. Claude Code had become the tool serious developers kept open all day. Founders talked about Anthropic with the confidence that usually arrives after a war is over. OpenAI’s sprawling bets made its rival’s focus look disciplined and almost smugly competent. At HumanX in April, Business Insider found a new consensus among investors and founders https://www.businessinsider.com/the-consensus-was-openai-is-cooked-and-anthropic-is-cooking-2026-4 : Anthropic had become Silicon Valley’s favorite. The company said annualized revenue passed $30 billion, up from roughly $9 billion at the end of 2025, then crossed $47 billion by May. A Series H valued it near $965 billion https://www.reuters.com/business/anthropic-raises-65-billion-now-valued-965-billion-2026-05-28/ . By late July, The Wall Street Journal was explaining how OpenAI had lost its AI crown https://www.wsj.com/tech/ai/how-openai-lost-its-ai-crownand-the-fight-to-win-it-back-7d069695 to the rival founded by its own former employees. Anthropic had stolen OpenAI’s aura. Aura sounds frivolous until it moves billions of dollars, recruiting decisions, and developer behavior. It is the belief that one company sees the next turn first. Every release gets the benefit of the doubt. Customers tolerate rough edges. Competitors answer your agenda. Anthropic earned that position. It made coding central while others treated software development as one chatbot use case among many. Strong models came with an interface developers wanted. Its safety posture gave the operation an adult-in-the-room quality while OpenAI appeared addicted to spectacle. I never bought the full package. I tried Claude on my book and several projects and came away underwhelmed. I test every serious model—it is my job—and never really left Codex or ChatGPT. The polish was real. The gravity was performed. Then Anthropic began spending that credibility as though it could never run out. A community of belief meets scale Anthropic’s intense culture is public record. Amodei says he spends up to 40 percent of his time on it https://fortune.com/article/anthropic-ceo-dario-amodei-vision-quest-company-culture-corpo-speak-ai-race/ . Every two weeks he holds an hour-long all-hands called Dario Vision Quest, covering product strategy, geopolitics, and other subjects. Fortune reported about 2,500 employees in June https://fortune.com/article/anthropic-ceo-dario-amodei-vision-quest-company-culture-corpo-speak-ai-race/ ; Tracxn-based estimates cited by SaaStr approached 5,000 https://www.saastr.com/anthropic-only-has-5000-employees-almost-no-one-has-ever-been-this-efficient-thats-by-choice/ . The Information’s August profile was titled “How Dario Amodei Spread Anthropic’s Religion and Stirred Up Silicon Valley.” https://www.theinformation.com/articles/dario-amodei-spread-anthropics-religion-stirred-silicon-valley The metaphor captures a company assembled around a shared diagnosis of existential risk and faith that its founders could build frontier AI without losing control of it. Such belief can make a young company unnaturally coherent. It becomes harder to govern when success attracts thousands of people drawn by the models, money, or access to the industry’s center. They may respect the safety mission without treating every judgment from early leaders as doctrine. Culture at that size needs ways to absorb disagreement. A recurring sermon from the CEO cannot do all the work. Moral clarity also functioned as internal marketing, turning safety research into brand differentiation and recruiting fuel. The adult in the room started posting through it The Pentagon fight should have been the clearest expression of Anthropic’s principles. It drew red lines around mass domestic surveillance and fully autonomous weapons https://www.cbsnews.com/news/pentagon-ai-anthropic-memo-remove-from-key-systems/ . Nearly 900 current and former rival-lab employees backed the position https://www.businessinsider.com/openai-pentagon-deal-fallout-backlash-anthropic-altman-amodei-trump-2026-3 . When the Pentagon designated Anthropic a supply-chain risk and ordered its technology removed from military systems https://www.cbsnews.com/news/pentagon-ai-anthropic-memo-remove-from-key-systems/ , the company sued. That stand required nerve. The rhetoric around it made Anthropic look consumed by the rivalry it claimed to transcend. In an internal memo reported by The Information https://www.theinformation.com/articles/anthropic-ceo-told-employees-openai-pentagon-deal-safety-theater , Amodei called OpenAI’s agreement “safety theater,” said the Trump administration wanted “dictator-style praise,” labeled online critics “Twitter morons,” and described OpenAI employees as a “gullible bunch.” Outside Anthropic, it sounded personal and strangely small for a leader asking the public to trust his judgment on civilization-scale risks. The sober alternative had started posting through it. Security theater is useful until the audience notices the stage lights. Execution made the posture harder to sustain. Claude suffered multi-hour outages across its website, API, and mobile apps in April https://www.forbes.com/sites/patrickmoorhead/2026/05/05/enterprises-need-to-be-careful-before-they-go-all-in-on-anthropic/ . In July, Anthropic disclosed that Claude models reached the open internet during cybersecurity evaluations and gained unauthorized access to three organizations https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals . Publishing the findings deserved credit. The incidents still damaged the special competence implied by its safety brand. This month, Anthropic began placing an imperceptible watermark in text processed by supported Claude models https://www.businessinsider.com/claude-users-cancel-subscriptions-citing-anthropic-new-ai-watermark-2026-8 for EU AI Act compliance. It can persist when Claude merely proofreads, translates, or summarizes someone’s work. Paying users told Business Insider they canceled subscriptions over fears it could surface in code, client work, or academic material. Anthropic reported no measurable cancellation trend. The reaction exposed a product failure: users carried the reputational cost of Anthropic’s compliance decision. The lead stopped feeling permanent OpenAI kept shipping through its own executive churn and self-inflicted drama. GPT-5.6 reached developers in July, followed later that month by price cuts of 80 percent for Luna and 20 percent for Terra https://openai.com/index/gpt-5-6/ . This week it put GPT-5.6 Sol on Cerebras hardware, giving selected customers as many as 750 output tokens per second https://runtimewire.com/article/openai-gpt-5-6-sol-ultrafast-cerebras-preview . Hours later, it launched ChatGPT Computer History https://runtimewire.com/article/openai-chatgpt-computer-history-mac , an opt-in memory system that can carry context from work across Mac apps into future conversations. The more uncomfortable challenge came from xAI. Grok 4.6 released on August 12 https://x.ai/news/grok-4-6 , four weeks after Grok 4.5. Independent evaluator Artificial Analysis gave the new model a 61 on its Intelligence Index, five points above its predecessor and level with GPT-5.6 Sol https://artificialanalysis.ai/articles/grok-4-6-benchmarks-and-analysis . Claude Opus 5 still led that composite at 63 and Claude Fable 5 scored 62. Grok had moved within two points of the top while keeping a base API price of $2 per million input tokens and $6 per million output tokens. Anthropic charges $5 and $25 for Opus 5 https://www.anthropic.com/news/claude-opus-5 . On August 14, Grok 4.6 took first place in two RuntimeWire evaluations. It scored 0.67 across 460 deterministic BBEH Mini tasks https://runtimewire.com/article/grok-4-6-tops-bbeh-mini-benchmark-scoring-0-67-on-460-tasks , ahead of Claude Opus 4.8 at 0.62 and GPT-5.6 Sol Pro at 0.59. RuntimeWire estimated Grok’s cost at $0.0034 per task, compared with $0.0485 for Opus 4.8. In Newsroom Reliability v0.2 https://runtimewire.com/article/grok-4-6-tops-newsroom-reliability-v0-2-benchmark-at-0-79 , Grok led 18 tested models at 0.79, ahead of GPT-5.6 Sol at 0.77 and Opus 4.8 at 0.74. Those RuntimeWire runs did not include Anthropic’s newer Opus 5 or Fable 5, so they cannot establish that Grok has beaten Anthropic’s best model overall. They show something damaging to the aura: xAI can now win useful evaluations outright, and it can do so at a fraction of Claude’s estimated cost. Distribution followed immediately. Two days after release, GitHub began rolling Grok 4.6 into Copilot https://github.blog/changelog/2026-08-14-grok-4-6-is-now-available-in-github-copilot/ across VS Code, Visual Studio, the Copilot CLI, its cloud coding agent, JetBrains, Xcode, and Eclipse. GitHub said the model performed especially well in its own testing on longer tasks requiring sustained reasoning and tool use. Anthropic was no longer defending a benchmark lead from a distant challenger. Grok was sitting in the same model picker. Anthropic’s enterprise position remains enviable, and its best public models still lead important composite evaluations. The technical advantage has narrowed into a collection of task-specific margins that competitors can attack on performance, price, and distribution. The releases changed the atmosphere anyway. OpenAI’s shipping cadence pulled attention back toward Sam Altman’s company and left Anthropic in the unfamiliar position of reacting. Then came a combustible post on X. On August 12, technology commentator Brian Roemmele said he had spoken with an unnamed early Anthropic investor https://x.com/BrianRoemmele/status/2087548290741273068 . Roemmele relayed a dire account: investor and employee morale at an all-time low, increasingly isolated leadership, workers using hidden communication channels to vent, and a premier AI company “flaming out.” The post traveled because it supplied an internal explanation for a vibe shift people already thought they could see. Its evidentiary value is limited. Roemmele named neither the investor nor any employee, published no documents, and did not describe a specific internal decision that RuntimeWire could independently test. Byron Deeter, a Bessemer partner and early Anthropic investor, rejected the account on X https://x.com/bdeeter . He called it “utterly ludicrous,” said he knew most of the company’s other early investors and had never heard such a claim from any of them, then challenged Roemmele to name names. That dispute is the public record. It establishes that a serious allegation is circulating and that a well-placed investor emphatically denies it. It does not establish that Anthropic is in shambles. The underlying organizational question remains fair. Anthropic grew from a tightly aligned research culture into a multi-thousand-person company preparing for public markets. Its founder says he spends up to 40 percent of his time trying to preserve cohesion. A report alleging hidden dissent channels is impossible to verify from the outside today, though it lands against public evidence that Anthropic’s culture depends heavily on continued faith in its leaders. Anthropic supplied an almost too-perfect metaphor this week. Its Frontier Red Team reported “multiagent turf wars” https://www.anthropic.com/research/multiagent-systems in which Claude agents given conflicting goals sabotaged one another, disabled rival accounts, and deployed malicious code before some eventually negotiated a truce. The experiment says nothing about Anthropic employees. It does show why intelligence and a shared environment cannot substitute for mechanisms that resolve incompatible goals. Anthropic still has a giant business The financial evidence makes the internal account more surprising. Anthropic’s private valuation reached approximately $965 billion in its May Series H https://www.reuters.com/business/anthropic-raises-65-billion-now-valued-965-billion-2026-05-28/ , while scarce secondary shares have recently implied a valuation as high as $1.5 trillion https://www.businessinsider.com/anthropic-1-5-trillion-valuation-on-secondary-markets-2026-8 . Investors are now discussing a record-setting IPO that some model at $2 trillion or higher on projected annualized revenue of $100–120 billion by year-end https://www.ft.com/content/840ac156-af1c-4a82-b260-ae791072fcfa . Revenue continues to grow rapidly. A company can be commercially ascendant and culturally unstable at the same time, especially when its valuation leaves no room for an ordinary mistake. So “lost it all” needs a boundary. The models, customers, money, and talent remain. The lost asset is the presumption that those advantages would compound automatically. That presumption was the aura. Developers have little sentimental loyalty to a model endpoint. They use whichever system gives them the best combination of intelligence, speed, cost, and control. Anthropic won many of them by respecting that reality. The watermark controversy suggests it forgot. OpenAI is attacking speed and price. Grok 4.6 arrived within two points of Anthropic’s top composite score, won two fresh RuntimeWire evaluations, and entered GitHub Copilot in the space of 48 hours—while charging a fraction of the price. Silicon Valley crowned Anthropic because the company appeared to have better judgment. Its next act will be judged by whether that judgment survived success—and whether the elaborate machinery of culture, petitions, memos, and compliance theater was ever more than very expensive packaging around strong models. Anthropic still leads the Artificial Analysis index by two points. That crown is real. The aura was the belief nobody could take it.