Anthropic's TACO Day: Pentagon Access, Claude Code Pricing Model

TACO: Pentagon Regains Access to Anthropic

Just a month ago, Anthropic was the first American company to be classified by the Pentagon as a 'risk to national security.' The reason: The AI startup refused to release its models for mass surveillance and autonomous weapons. Now, the familiar TACO move ('Trump always chickens out'): Trump suddenly announces they've had 'very good conversations' and that Anthropic could be 'of great use.' What triggered the reversal? Mythos is the name of Anthropic's new model, which can allegedly identify and exploit software vulnerabilities on an unprecedented scale. Instead of a public release, there is a limited preview for selected corporations and governments under the name 'Project Glasswing.' While European governments reportedly reacted with alarm, the NSA is apparently already using the model. → Techpresso

Synthszr Take: Anthropic has flipped the script like a chess player intentionally sacrificing their queen. The refusal to supply AI for autonomous weapons wasn't a moral retreat but a calculated move to increase its value. With Mythos, the company now presents a model so powerful that even the hardliners in Washington can't ignore it. We know this pattern from the pharmaceutical industry: first, a drug is declared too dangerous for the open market, then governments get it under strict conditions (and at premium prices). Anthropic isn't just selling a product; it's selling exclusivity through artificial scarcity.

Anthropic Copies the TACO Move and Backpedals on Pricing

Anthropic updated the pricing page for Claude.com yesterday without any announcement, quietly removing Claude Code from the $20 Pro subscription. The feature, previously included in the Pro plan, is now slated to be exclusively available in the $100 or $200 Max tiers—at least according to the briefly visible pricing table that has since been removed. Amol Avasare, Anthropic's Head of Growth, tweeted after fierce reactions on Reddit, Hacker News, and Twitter that it was a 'small test with about 2% of new prosumer sign-ups.' Existing Pro and Max subscribers are not affected. Simon Willison, who documented the change, doubts this account, as the new pricing table was visible to everyone and had already been captured by the Internet Archive. The confusion is compounded by the fact that Claude Cowork—essentially a rebranded version of Claude Code—remains available in the $20 plan. → Techpresso

Synthszr Take: Anthropic is stumbling over the classic SaaS dilemma: how much is a category-defining feature really worth? Claude Code created the market for coding agents and already generates billions in annual revenue, but the jump from $20 to $100 a month is like trying to turn a busy public park into an exclusive country club. The stealth price change without prior communication is reminiscent of the early days of the airline industry, when carriers adjusted their fares multiple times a day, hoping no one would notice. OpenAI's Codex team immediately seized the opportunity for a trust offensive ('Transparency and trust are two principles we will not break'). What Anthropic is really testing here isn't price sensitivity, but how much goodwill they can afford to burn before their community defects to competitors.

Anthropic Mythos: Hackers Didn't Need Trump's Okay

A group of unknown users has gained access to Mythos, Anthropic's exclusive cybersecurity tool for enterprises. The model was actually only made available to select partners like Apple as part of 'Project Glasswing' to prevent misuse. According to Bloomberg, the breach was accomplished through a third-party vendor where a member of the group is employed. The Discord community, which specializes in unreleased AI models, guessed the model's online location based on Anthropic's previous naming conventions. Since the public announcement, the group has been using Mythos regularly and demonstrated access to Bloomberg via a live demo. Anthropic is investigating the incident but has so far found no evidence of any impact on its own systems. → Techpresso

Synthszr Take: Anthropic has created a classic chicken-and-egg problem: a security tool so powerful it becomes a threat itself. The Discord group operates like mushroom hunters who guess the best spots based on soil conditions and tree types, except here they're reading URL patterns instead of forest signs. Project Glasswing was supposed to be a controlled ecosystem, but software knows no walled gardens once people with access are involved. The irony: a tool designed to fend off cyber threats is compromised through social engineering and pattern recognition, not technical exploits. Anthropic is learning that in the digital age, exclusivity is a fiction.

Sam Altman Plays the Anti-Elite Card Again

OpenAI CEO Sam Altman criticized Anthropic's cybersecurity model Mythos as 'fear-based marketing' during a podcast appearance. Anthropic had released the model in early April only to select enterprise customers, arguing it was too powerful for the public and could be misused by cybercriminals. Altman mockingly compared the strategy to selling bomb shelters: 'It's clearly incredible marketing to say, 'We've built a bomb, we're going to drop it on your head. We'll sell you a bunker for $100 million.'' He warned that such tactics served to keep AI in the hands of a small elite. Critics note that Altman himself regularly invokes apocalyptic AI scenarios when it serves his agenda. → AI Secret

Synthszr Take: Altman accuses Anthropic of what the entire AI industry has perfected: the monetization of existential dread. The business model works like medieval indulgences—first paint a picture of hell, then sell salvation. Anthropic's 'too dangerous for the public' narrative is just a variation of the classic pharma playbook: exclusivity through manufactured scarcity. The real joke is that both companies are playing the same game, just in different costumes. While Anthropic plays the concerned scientist, OpenAI stages itself as the democratic liberator—and cashes in just the same. The AI industry has understood: nothing sells better than the apocalypse, except maybe salvation from it.

Google Now Codes 75 Percent with AI

Google now generates three-quarters of its new code using artificial intelligence, the company announced on Wednesday. The lines of code created by AI are subsequently reviewed by human engineers. This rate has steadily increased in recent years: in October 2024, it was still at 25 percent, and by fall, it had already reached 50 percent. CEO Sundar Pichai speaks of 'truly agentic workflows,' where engineers are increasingly taking on more autonomous tasks. A particularly complex code migration was completed six times faster through the collaboration of AI agents and engineers than it was a year ago with engineers alone. Google developers primarily use the in-house Gemini models for code generation, while some DeepMind employees have also recently been allowed to use Anthropic Claude Code. AI usage is already being factored into performance reviews for some employees. → techmeme

Synthszr Take: Google is proving that code generation is the most lucrative business in the AI value chain. While others are still philosophizing about AGI, Google has built a factory for synthetic code that feeds itself: AI writes code that trains new AI systems, which in turn write better code. This is reminiscent of the invention of the machine tool, which enabled the creation of more precise machine tools. The 75 percent mark signifies a phase transition: beyond this threshold, human programming becomes quality control, not production. Google's real innovation is allowing Anthropic Claude at DeepMind (despite internal tensions). They treat AI models like compilers: interchangeable, as long as the output is correct.

Google's TPU Split Shows: The AI Industry is Becoming a Specialization Industry

At Cloud Next 2026, Google split its eighth TPU generation into two separate chips: TPU 8t for training with 121 exaflops per pod and 2.8x better price-performance than its predecessor, Ironwood, and TPU 8i for inference with 80% better performance per dollar. The inference variant uses 288 GB of High-Bandwidth Memory plus 384 MB of on-chip SRAM to keep agentic workloads directly on the chip. Broadcom designed the training chip, MediaTek the inference chip, while Intel, Marvell, and TSMC complete the supply chain. Thomas Kurian, CEO of Google Cloud, calls the split a 'natural evolution'—two years in development before agents became mainstream. In parallel, Google is launching Workspace Intelligence as a context layer across all Workspace apps, which learns user behavior and generates authentic-sounding documents. → Techpresso

Synthszr Take: With the TPU split, Google is doing what biology has demonstrated for millions of years: specialization beats generalism when the environment becomes complex enough. Training needs throughput, inference needs latency—two fundamentally different optimization problems that Nvidia tries to solve with a single chip. The division of chip designs between Broadcom (training) and MediaTek (inference) is reminiscent of the division of labor in an ant colony: each does what it does best. MediaTek's stock price didn't explode to a record high for no reason—the Taiwanese chipmaker is becoming the silent winner of the inference wave while everyone is staring at Nvidia. The real punchline is hidden in the million-TPU deal: 3.5 gigawatts of capacity starting in 2027 means the demand for specialized inference hardware is just beginning. Anyone who still believes one chip can do it all hasn't understood the signs of the times.

Wispr Ignites a Voice Revolution: No Keyboard, No Prompts

The startup Wispr Flow has raised $81 million to develop a 'Voice OS'—software that integrates voice input into any text editor on Mac, Windows, iPhone, and Android. Users simply speak, and the service transforms their words into structured, error-free text at 220 words per minute, instead of the usual 45 when typing. The unique feature: Flow works without special prompts or commands directly in any app, from VS Code to WhatsApp. Clay's sales team is already reporting 20 percent more customer calls per day because they can simply dictate notes and follow-ups instead of typing. Wispr is aimed at developers, content creators, support staff, and people with disabilities—essentially, anyone who wants more output with less mechanical work. → TAAFT - There's An AI For That

Synthszr Take: Wispr is delivering on what Siri has been promising since 2011: turning natural language into usable text without having to say, 'Hey Computer, write a letter.' The business model follows the principle of invisible infrastructure—like electricity from a socket, but for thoughts. The $81 million in funding shows that investors believe in a future where the keyboard becomes as obsolete as the rotary phone. The real test will come when Wispr has to compete against the built-in voice features from Apple, Google, and Microsoft, which are integrated into the operating systems. Wispr is betting that specialized excellence will beat the platform giants—just as Zoom once outshone the video features of the major tech corporations.

Sullivan & Cromwell Lawyers Let the AI Hallucinate

The renowned law firm Sullivan & Cromwell had to apologize in court for AI-generated false citations after its lawyers cited fabricated court decisions in a legal brief. Partner Arthur Robinson took responsibility and admitted that the firm had failed in its use of AI tools. The hallucinated citations concerned a case in which a former UBS trader is suing the Swiss bank. Sullivan & Cromwell represents UBS in this case. The firm promised to revise its internal processes for handling AI tools and to ensure in the future that all sources are manually verified. → Techpresso

Synthszr Take: Sullivan & Cromwell is the latest elite law firm to learn that AI hallucinations cost more than the research time saved. The phenomenon is reminiscent of early Wikipedia errors, except this time, highly paid partners are signing off on the fabricated sources. The industry faces a classic innovator's dilemma: AI tools promise efficiency, but every hallucination can mean a million-dollar lawsuit or reputational damage. Law firms will likely soon hire 'AI Compliance Officers' whose sole job is to verify machine-generated text. The irony: in the end, it takes more human labor to control the machines than the original research would have cost.

AI Search 2026: The New Two-Tier Society of Visibility

A new study by AirOps analyzes 15 million AI search queries and shows that the search landscape is splitting between classic Google search and AI assistants like ChatGPT, Gemini, and Perplexity. Companies like Ramp, Carta, and Webflow are reporting dramatic success through targeted AI optimization: Webflow recorded 6% more AI-driven sign-ups within a few days; Chime tripled its citations in AI answers from 24 to 68 priority questions. The study identifies four key strategies that successful companies use to remain visible in both worlds. It reveals that those who rely solely on traditional SEO are already losing relevant customer groups today. The data comes from real-world implementations at leading SaaS companies and shows measurable pipeline effects from AI search optimization. → AirOps

Synthszr Take: The splitting of search is reminiscent of the emergence of parallel worlds in urban development: while Google remains the main street of the internet, AI assistants are creating exclusive express lanes for informed users. Companies face a resource dilemma, like restaurant owners who suddenly have to serve both walk-in customers and delivery services. The tripling of Chime's citations shows that AI systems follow different relevance criteria than PageRank. Those who only optimize for Google today are building digital Potemkin villages tomorrow that will remain invisible in AI search. The bifurcation of the search ecosystem is becoming the new normal, where companies must invest in two currencies simultaneously: classic SEO authority and AI understandability.