Anthropic Lost $8B Operating in 2025. You Feel That as Usage Caps.

September 29, 2026 — Reuters reviewed Anthropic’s IPO prospectus. 2025 revenue: nearly $4.6 billion. Operating loss: more than $8 billion (follow-on copy: $8.06 billion, up from $2.98 billion in 2024). Net loss: about $42 billion, of which ~$34 billion was a non-cash charge from marking up financing that can turn into stock. Compute last year: $7.33 billion of $12.65 billion in operating expenses. Future cloud/compute/infra obligations: $518 billion. Cash at year-end 2025: $20.28 billion. Anthropic declined to comment. Listing talk: after November midterms, target more than $2 trillion — more than double the $965 billion May mark.

Same 48 hours, OpenAI told ChatGPT Pro buyers the reopened $200 seat nets to half the old API dollars, while a $500 Pro Max leak sat on top of DevDay.

That is not two stories. It is one bill. The labs ran a loss-leader frontier. The losses are now public. The consumer meter is catching up. We already argued, in point of no return, that dependence plus a later price hike is how you get stuck. The hike is this week.

TL;DR

Do not cite $42 billion as “they spent $42 billion on models.” Cite $8 billion of operating hole on $4.6 billion of sales, plus a mark-to-market on paper that got more valuable because the private valuation went up.

Read the losses like a token bill, not a ticker

explainx.ai is not a markets desk. An IPO date does not change your SDK call until the meter does — same rule as the September 13 valuation-talks post. The prospectus is useful because it shows why the meter is moving.

Revenue $4.6B vs opex $12.65B means 2025 Claude was not a cash machine. Compute $7.33B — threefold vs 2024, more than half of opex — is the line that maps to your Codex and Claude Code sessions. $518B of future obligations is how you get $200 that suddenly prints half, Astra 4x cuts, and a Claude Code net ~17% cut in the same season.

Customer concentration: nearly a quarter of 2025 revenue from two customers, many large clients not on long-term contracts (Reuters). If those two sneeze, you are the residual demand they will reprice.

FT/TechCrunch 2026 print ($11.5B in one quarter, adjusted profit) is why bankers still want >$2T. It does not refund 2025. It tells you the company is trying to grow into the compute stack while public-market buyers will ask when the consumer subsidy ends.

Earlier rumor math on our Sep 13 IPO-talks post put year-end 2025 run rate ~$9B. The prospectus 2025 revenue is $4.6B. Prefer the prospectus. Run-rate chatter for mid-2026 is a different year.

What “AI bubble” means if you actually pay for models

Reuters: AI and chip stocks sold off; SpaceX’s IPO popped then faded; Anthropic’s sale tests whether the AI trade survives scrutiny. Fine. That is their bubble.

Your bubble was unlimited-feeling frontier at $20 / $100 / $200. Token economics were always: push agents, burn tokens, win share, fix the P&L later. Point of no return called the later part: deskilled teams cannot leave when the bill jumps.

The pinch, this week, is not 10× overnight. It is worse because it is quieter:

1. Same $200 sticker, half the API-dollar print — Tibo. You did not get a refund. You got a smaller tank.
2. A $500 rung hanging over DevDay — still unconfirmed. If it ships, EZRA’s joke becomes product: old Pro lives at $500. Speed (Work/Codex) as the SKU, not “more chats.”
3. Last-gen Unlimited* — 5.6 Sol/Luna/Terra starred on Pro while GPT-6 stays Expanded. The buffet is yesterday’s model.
4. Claude side already cut the promo meter — 17%. Anthropic does not need an IPO day to squeeze Code.

That is a bubble deflating at the checkout, not a prediction about NVDA. Off-balance-sheet compute and vendor financing were the delay. The prospectus is the invoice.

OpenAI: $200 halved, $500 leaked — do not mix the two

Confirmed: new Pro sign-ups reopen; usage calculation changes; Tibo: half the dollar in API spend vs old $200; no 5-hour cap coming back. Full decode.

Not confirmed on the pricing page: Pro Max $500. Treat it as the next shelf until openai.com lists it. DevDay can ship, delay, or kill the string.

How they fit: you cannot halve the popular tier and keep Astra Codex all week unless you invent a more expensive tier or you dump work onto cheaper models (Sol/Luna 50% API). Tibo said (b) is over time. Consumers heard “more work.” The usage page will settle the argument.

Compare labs without morality: Opus 5.5 vs Sonnet 5.5 is still a routing choice. Switching because a tweet said bubble is dumb. Switching because your weekly bar died on Tuesday is rational.

What this means for what you build or pay

If you are a $20 / $100 user: you were never the 2025 operating-loss customer. You are the habit. Expect more Expanded / fair-use stars and less “unlimited frontier.” Budget API for anything you cannot miss.

If you are a $200 Pro / Max / Claude Code heavy: you are the 2025 compute line. Screenshot remaining this week. Route grunt to cheaper SKUs. Keep one eval that is yours so a limit cut does not delete your job.

If you run a team that tokenmaxxes: the point of no return trap is live. A deskilled org facing half tank + $500 ultra cannot “just think harder.” Kill the token leaderboard before finance does it for you. True-cost math still starts with subscription vs API.

If you were waiting on Anthropic’s IPO to “make Claude safe/cheap”: public markets tighten consumer subsidies; they do not gift them. PBC language and pacing essays sit next to $518B of infra promises. Safety risk factors (shutdown resistance, concealment, blackmail-like behavior in tests — Reuters/FT via TechCrunch) are for investors. Your rate limit is the part that ships.

Do not treat adjusted 2026 profit as “the bubble is over, prices fall.” Profit can arrive because they stopped giving Astra away.

Claude vs ChatGPT this week (same cost structure, two meters)

On Claude, you already ate a promo-to-permanent haircut on Code limits. The prospectus says why: compute was already more than half of 2025 opex, and $518B more is contracted-looking spend, not a vibe. Opus 5.5 shipping into that P&L is sales, not generosity.

On ChatGPT, Tibo pre-briefed the cut so DevDay would not look like a surprise downgrade. Unlimited* 5.6 is the pressure valve. GPT-6 Expanded is the product they still have to ration. The $500 leak is how you recreate last month’s $200 for people who will pay speed. Until it is on the pricing page, do not wire a budget to it. Do wire a budget to half tank. Check openai.com before you treat $500 as live.

Neither lab “burst the bubble” for NVIDIA holders. Both are passing the 2025 hole to weekly remaining. If you still have the skill to route — cheap model, cache, two vendors, AI-off drills — you are in the group point of no return said would survive a reprice. If you do not, the IPO PDF is not your problem. Tuesday’s usage bar is.

Two customers, a quarter of 2025 revenue is the other consumer tell. Those accounts will negotiate. Everyone else gets the list meter. That is how a $2T story and a $20 Plus habit coexist: the enterprise keeps the lights on; the weekly bar keeps the crowd from lighting the cluster on fire.

What to do this week

1. Read Tibo and the Models grid as one OpenAI story: meter GPT-6, star 5.6, maybe sell speed at $500.
2. Prefer prospectus figures over Sep 13 rumor run-rates when you cite Anthropic 2025.
3. Keep two labs. Nvidia–Anthropic $180B contracted value is their entanglement, not a reason to single-home your agents.
4. Re-run the point of no return drill: AI-off one small ship. If you cannot, the pinch is not a pricing problem. It is a skill problem with a pricing problem.

Related reading

• AI is taking us to a point of no return
• ChatGPT Pro $200 reopen — half API dollars
• ChatGPT Pro Max $500 leak
• Anthropic $100B IPO talks (Sep 13 — valuation rumor, pre-loss table)
• Why AI companies want you using agents
• True cost of ChatGPT / Claude / Gemini subs
• Claude Code ~17% limit cut
• Astra usage 4x cut
• Nvidia as compute financier
• Primary: CNBC / Reuters prospectus · TechCrunch (FT 2026 run-rate)

$42B net vs $8B operating is the distinction most headlines will skip. $500 Pro Max is leak-only until OpenAI’s pricing page moves. Anthropic declined to comment on the Reuters read of the prospectus.