Anthropic vs. OpenAI: the bill behind the two big AI labs

Every few months somebody re-asks a question that sounds eternal but deserves a fresh look in 2026: Anthropic or OpenAI, Claude or ChatGPT? A new video by Christoph Magnussen from September 21, 2026 walks through both AI labs, from products to pricing to the philosophies behind them. I used the video as a starting point and took its central claims to primary sources, and that verification turned out to be necessary, because one pairing in the transcript does not survive the check. What emerges is a picture more interesting than any versus title: on the surface the two offers look alike, but in founding philosophy and in the infrastructure bill behind them they diverge sharply. That bill is the part that matters for anyone running agents, because it explains why usage limits jump and what a month of operation actually costs. What is verified, what remains judgment, and where the video diverges from the reporting, I lay out below.

The core difference starts in the origin stories. Anthropic was built by seven co-founders who left OpenAI because they held a significantly different view on how models should be trained. OpenAI leaned for years on Reinforcement Learning through Human Feedback, RLHF: many humans rate an unrefined model's outputs and nudge it toward their preferences. Anthropic bet on Constitutional AI from the start: the model evaluates itself against a constitution of principles, improves through that loop and even helps train other models. Magnussen notes in the video that both methods are now standard on every side, but the constitution remains the philosophy at Anthropic: the model should improve itself, not merely be polished by people. The second difference is infrastructure. In the video's telling, OpenAI bet early on owning compute, the logic being more compute, more intelligence. Anthropic has to rent capacity at market prices, and that is where the money question begins. The talent market stays in motion too, well-known researchers rotate between Google, OpenAI and Anthropic, and that churn keeps both labs moving.

The money block of the video contains one pairing that I have to correct before anyone quotes it onward. The transcript names 1.25 billion per month for Google and 1.2 billion per month for SpaceX, both framed as sums Anthropic pays. The reporting draws a sharper picture. Per TechCrunch, Anthropic pays $1.25 billion a month to SpaceX, which merged with xAI in February 2026, renting capacity in the Colossus 1 data center near Memphis, the deal runs through May 2029. Google, per CNBC and TechCrunch, pays $920 million a month to SpaceX, spread over 32 months, for roughly half the capacity Anthropic books. Anthropic's Google Cloud partnership is a separate matter: Anthropic itself confirmed expanded use of Google Cloud TPUs in its own announcement. Yahoo Finance adds up the compute deals of Anthropic and Google to roughly $26 billion a year. The lesson from this correction matters more than any single figure: data centers have become rentals in this market, and the payment flows cross companies that stage themselves as competitors on stage.

On the product level Magnussen concludes that the two offers converge. OpenAI offers ChatGPT as the consumer entry point, business variants on top, and Codex as the coding line that has arrived in business use. Anthropic counters with Claude, long seen as the more agency-like, better-reading writer, had the Claude Code moment that broke coding agents open and turned it into Claude Cowork as a work surface. With Claude Design, Anthropic now moves straight into tooling otherwise found at agencies. Price lists sit in the same band, roughly 20 to 200 euros or dollars depending on tier, in companies via API or per seat. The visible difference is self-image: Anthropic thinks hard in language and design, its models behave in workflows more like a colleague with a character, which some people love and others find unsettling. OpenAI iterates faster on breadth and integration, updates arriving, in Magnussen's telling, multiple times a day. His anti-FOMO advice: both labs ship fast, a missed feature usually arrives weeks later, so do not switch brands over a single feature.

A point often missing from public comparisons concerns companies as customers. Both labs now sell individual seats, at OpenAI the premium variant arrived later than at Anthropic per the video, and via the API usage can be distributed per person. Anyone wanting Claude Code needs at least a higher tier, below that the coding tools stay locked. For companies, data protection agreements are the decisive building block, and the video advises checking those contracts case by case before work data enters either product. The big incumbents get their profile at this spot too: Google and Microsoft carry a different setting and a different legacy of responsibility. Their models are weaker in tool use, but fast, cheap and strong in multimodality, sound, image and text together. Per Magnussen, nobody wants Copilot to address them as a buddy, while some people find Claude's human-like behavior charming and others find it off-putting. These are valid selection criteria, they just deserve a deliberate decision against your own operation, not against headlines.

At the edge of the video sits a judgment with reach beyond the product comparison: are AI companies a new corporate form or a bubble? Magnussen points to AI-native startups that infuse AI into everything from day one and reach different orders of magnitude in revenue per employee. His take: some of it may be bubble, much of it is real. The margin comparison is telling. Classic SaaS vendors reached margins of 80 to 90 percent, AI company margins are lower because models are expensive, while efficiency gains keep driving model costs down, so the math improves every quarter. The open question of the video, which corporate form and team organization becomes the new standard, cannot be answered today. For agent operators the takeaway is one sentence: the cost side moves faster than the price lists, and efficiency per task is the number to watch, not list price per seat.

Honesty requires the other side of the ledger. This video is an informed mood picture, not a benchmark: model quality is compared by experience and feel, not by measurement, and the infrastructure figures are partially garbled in the transcript, as my check showed. Anyone who needs hard structures is better served by primary sources: TechCrunch and CNBC carry the deal structures with terms and cancellation clauses, and Anthropic itself describes its Google Cloud use. One question stays open regardless: how the two philosophies, Constitutional AI on one side and compute scaling on the other, pay off across the next model generations will be decided in practice, by teams running real workloads and publishing real numbers. A comparison video can frame that question well, and this one does, but the answer gets written by operators. Until then: useful for sorting, wrong to over-quote.

From an agent-operator view the decisive finding is not the product comparison but the infrastructure bill behind it. $1.25 billion a month from Anthropic to SpaceX, $920 million a month from Google to SpaceX, roughly $26 billion a year combined: that order of magnitude explains why usage limits, prices and model availability jump the way they do. TechCrunch reports that Anthropic raised its usage limits the very day the SpaceX deal was announced: that is the direct line from rental contract to user experience. I sign Magnussen's advice, run both labs as a professional and watch your own usage, because in my operation decisions happen per task, not per brand, coding rails and writing rails choose differently. What I do not adopt is the FOMO logic the video itself rejects. The right question is not who ships the better feature this week but which cost logic fits your operation: build infrastructure, rent capacity, or calibrate usage so the rent stays affordable. That question does not get cheaper by postponing it.

About the author: Marcel runs the HUNTER cyberdeck on a used Google Pixel 6a and documents daily agent operations, including a weekly token balance. This post is based on Christoph Magnussen's video plus TechCrunch, CNBC, Yahoo Finance and Anthropic's own announcement, retrieved September 21, 2026. Last fact-check: September 21, 2026.