The Lyceum: AI Daily — Mar 14, 2026

Anthropic just made it dramatically cheaper to feed an AI an entire codebase in one gulp. Elon Musk publicly admitted xAI was built wrong — while preparing it for a reported $1.25 trillion public offering. And state legislatures, tired of waiting for Congress, are quietly building a patchwork of AI laws that will shape what companies can actually ship. Friday cleared the table; Monday's Nvidia keynote sets it again.

If you've been building anything that needs an AI to read a lot — a full codebase, a year of contracts, an entire research corpus — this is the pricing news you've been waiting for.

Anthropic flipped the switch on general availability for the 1-million-token context window across Claude Opus 4.6 and Sonnet 4.6, and killed the long-context pricing premium entirely. A token is roughly three-quarters of a word, so a million tokens is about 750,000 words — the complete Harry Potter series. Previously, sending that much text cost meaningfully more per word. Now a 900,000-token request costs the same per token as a 9,000-token one. Rates hold at $5/$25 per million tokens for Opus 4.6 and $3/$15 for Sonnet 4.6 — unchanged from base pricing.

For Max, Team, and Enterprise users running Claude Code, Opus 4.6 sessions now automatically access the full million-token window — fewer conversation compactions, more session history preserved intact. Developers are already testing whole repos and multi-month logs in single requests.

Google's Gemini 2.5 Pro matches the window size but still charges a premium above 200K tokens, putting Anthropic at a cost advantage in the segment that matters most: complex, long-running agentic workflows. The media limit bump — from 100 to 600 images or PDF pages per request — quietly unblocks document-heavy pipelines too.

The caveat worth watching: a Princeton NLP benchmark showed most models degrade past 32K tokens on summarization tasks. Users stress-testing million-token sessions report that models start losing fidelity on constraints set hundreds of thousands of tokens earlier. Accepting a million tokens doesn't mean the model retains every detail perfectly — practitioners advise deliberate chunking, explicit state handoffs, and treating the model as a reasoning engine on top of tools, not an infinite memory store.

The enterprise pricing war just escalated. Whether the capability matches the price tag at scale is the question the next two weeks of developer adoption will answer.

When a CEO publicly says his company "was not built right the first time," that isn't messaging. That's a confession.

Of xAI's original 11 co-founders, only two remain. "xAI was not built right first time around, so is being rebuilt from the foundations up," Musk posted on X. This week, co-founders Zihang Dai and Guodong Zhang departed after Musk complained that xAI's coding tools couldn't compete with Anthropic's Claude Code or OpenAI's Codex. The Financial Times reports that virtually everyone who appeared on the Grok-4 launch livestream has now left or been fired.

The attempted fix: Andrew Milich and Jason Ginsberg are joining from Cursor, the AI coding tool company — an acknowledgment that xAI needs to learn from people who've already won this market. Coding tools matter because they're where the money is; xAI's lag here isn't a perception issue, it's a business problem.

The backdrop makes this explosive. Tesla shareholders are suing Musk for breach of fiduciary duty over xAI's founding. Reports say a combined SpaceX–xAI entity is targeting a reported $1.25 trillion public offering — and Musk is rebuilding the AI division mere months before the listing window opens. If xAI doesn't ship a competitive coding product before summer, that valuation will face very hard questions.

Federal AI policy is stuck. State capitols aren't waiting.

Washington state this week gave final passage to a chatbot safety measure — the second such state-level measure to pass in 2026, following Oregon's approval last week. Utah passed so many AI bills in its short session that the Transparency Coalition devoted an entire separate post to all nine. Virginia passed three significant AI bills this week, with one more in play as its legislature hits scheduled adjournment today. Across multiple states, lawmakers passed measures limiting AI in schools, protecting people from deepfakes, and ensuring medical decisions are made by humans.

The R Street Institute's new "Terrible Ten" report calls out state laws that overreach or create disclosure regimes practically impossible for small teams. Healthcare providers face a particularly convoluted landscape — the Trump administration has limited federal guidance, while individual states legislate their own rules.

A patchwork of 50 state AI laws is the outcome nobody wants — including most advocacy groups that pushed for regulation. But if Congress doesn't act, that's exactly what companies shipping AI across state lines will get. Several measures are heading to governors' desks this month.

A practitioner conversation blowing up in developer forums highlights a fundamental weakness: today's AI agents are functionally blind to the last 24 hours. Even agents with web-browsing capabilities pull from indexed content, not live conversations happening on social media right now.

This isn't minor. For anyone building real-time market analysis, brand monitoring, or breaking-news tools, by the time an agent "sees" a customer reaction or competitor launch, the critical window has passed. The current workaround — manually feeding agents curated data or stitching together delayed APIs — points to a major infrastructure gap. Making it worse: third-party search APIs can hit quota limits or return authorization errors without warning, cutting off an agent's data sources mid-workflow.

If agents can't access truly live data, their value in fast-moving environments is fundamentally capped. Expect this gap to spawn startups focused on real-time data ingestion for AI.

Adobe's Photoshop beta now includes a "Rotate Object" feature that selects an element in a flat photo and AI-generates the hidden sides as you spin it. Early demos are split — sometimes convincing, sometimes uncanny, and some users report the tool doesn't work at all after updating.

If this stabilizes, it's photogrammetry without the multi-camera setup — enormous for e-commerce product shots and concept art. Every accept/reject action from users also becomes training signal, steadily improving Adobe's internal 3D priors. The tech is real; reliability is catching up to the demo.

A million-token window priced like a regular Tuesday. An AI lab whose CEO calls it a fixer-upper. A Photoshop button that hallucinates the back of your head. Somewhere, a fully blind developer is asking a local LLM to write the code that lets them bypass every visual bottleneck the industry never bothered to fix — and that might be the most important AI story nobody covered this week. Enjoy the weekend; GTC eats Monday. If someone you know builds things with AI, forward this — they'll want it before Jensen takes the stage.