Anthropic, the company behind Claude, set the terms of its stock market debut on Tuesday: $70 to $78 a share, raising up to $64.35 billion at a fully diluted value of up to $1.405 trillion, with pricing expected November 18 and Nasdaq trading the next day, according to CNBC(opens in a new tab). The day before, it released Claude Fable 5.2(opens in a new tab) and cut the price 20%, to $8 per million input tokens and $40 per million output tokens. A token is a small chunk of text, roughly three quarters of a word.
Read those two facts together. The amended filing, as CNBC reported(opens in a new tab), shows $33.24 billion of revenue in the nine months to September, up about 1,200%, and about $895 billion of purchase commitments, almost all of it for computing capacity and infrastructure. The same filing shows its compute and infrastructure costs up 329% over those nine months. Nothing in it says those costs fell by a fifth. The cut is a competitive move made while investors are watching, and Meta pushed the same way on Monday with Muse Ember(opens in a new tab), a small agent model at $0.10 per million input tokens.
The layer underneath is paying more for money. CoreWeave, which rents out Nvidia chips to AI companies, had bonds due 2031 yielding above 16% last week, and the cost of insuring its debt against default rose about 50% in a month, CNBC reported(opens in a new tab). Nebius, a rival, lost about half its value in the same month. CoreWeave reports results on Tuesday. So lenders charge the capacity builders a junk bond price while stock buyers value a model maker near $1.4 trillion.
For anyone buying AI, that split cuts both ways. Token prices can keep falling for a while, because competition sets them and investors fund the gap. The capacity behind them is financed on terms that can change in weeks. Neither the IPO range nor the price cut tells you what you will pay in 2028.
For the CAIO and CFO this week. Move work already running on Fable 5.1 to 5.2 now; the saving needs a model change and a regression test that includes any security or lab work, since 5.2 can route flagged queries to a weaker model, and API customers have to set up Anthropic's new Fallback API for that. Do not trade that cut for a multi-year volume commitment that assumes prices only fall, and keep model spend commitments to twelve months. If any of your capacity runs through a specialist GPU cloud, directly or through a reseller, ask counsel what your contract says happens to reserved capacity and prepaid credits if the provider is sold, restructured or misses a debt payment, and add an exit right tied to those events at your next renewal.
Action items
Anthropic, the company behind Claude, set the terms of its stock market debut on Tuesday: $70 to $78 a share, raising up to $64.35 billion at a fully diluted value of up to $1.405 trillion, with pricing expected November 18 and Nasdaq trading the next day, according to CNBC. The day before, it released Claude Fable 5.2 and cut the price 20%, to $8 per million input tokens and $40 per million output tokens. A token is a small chunk of text, roughly three quarters of a word.
Read those two facts together. The amended filing, as CNBC reported, shows $33.24 billion of revenue in the nine months to September, up about 1,200%, and about $895 billion of purchase commitments, almost all of it for computing capacity and infrastructure. The same filing shows its compute and infrastructure costs up 329% over those nine months. Nothing in it says those costs fell by a fifth. The cut is a competitive move made while investors are watching, and Meta pushed the same way on Monday with Muse Ember, a small agent model at $0.10 per million input tokens.
The layer underneath is paying more for money. CoreWeave, which rents out Nvidia chips to AI companies, had bonds due 2031 yielding above 16% last week, and the cost of insuring its debt against default rose about 50% in a month, CNBC reported. Nebius, a rival, lost about half its value in the same month. CoreWeave reports results on Tuesday. So lenders charge the capacity builders a junk bond price while stock buyers value a model maker near $1.4 trillion.
For anyone buying AI, that split cuts both ways. Token prices can keep falling for a while, because competition sets them and investors fund the gap. The capacity behind them is financed on terms that can change in weeks. Neither the IPO range nor the price cut tells you what you will pay in 2028.
For the CAIO and CFO this week. Move work already running on Fable 5.1 to 5.2 now; the saving needs a model change and a regression test that includes any security or lab work, since 5.2 can route flagged queries to a weaker model, and API customers have to set up Anthropic's new Fallback API for that. Do not trade that cut for a multi-year volume commitment that assumes prices only fall, and keep model spend commitments to twelve months. If any of your capacity runs through a specialist GPU cloud, directly or through a reseller, ask counsel what your contract says happens to reserved capacity and prepaid credits if the provider is sold, restructured or misses a debt payment, and add an exit right tied to those events at your next renewal.
Researched and drafted by an automated workflow, then reviewed and edited by a human editor before publication. Every source is linked. See how we use AI here.
Also worth knowing
- CNBCAnthropic sets IPO price range of $70 to $78, looks to start trading Nov. 19, sources say(opens in a new tab)
Up to $64.35 billion raised at up to $1.405 trillion, with about 20% of shares aimed at retail buyers against a typical 5% to 10%. Run-rate revenue was about $81 billion at the end of October. Once public, Anthropic reports its revenue and margins every quarter.
- AnthropicClaude Fable 5.2(opens in a new tab)
$8 and $40 per million input and output tokens, down from $10 and $50, on Pro plans and up. Fast mode runs up to 2.5 times quicker at double the price. Flagged cyber and biology queries may be routed to a weaker model, and 30-day retention is the default.
- MetaStart building with Muse Ember, our smallest, most efficient model(opens in a new tab)
$0.10 per million input tokens and $0.50 output for prompts up to 200,000 tokens, triple that above. Meta's own table puts it ahead of Claude Haiku 5.5 on agent tests, and it scores 36 on the independent Artificial Analysis index. If you run high-volume agent steps, test it on your own tasks.
- CNBCCoreWeave confronts 'doom cycle' narrative as neocloud credit risk overshadows growth(opens in a new tab)
CoreWeave had $35 billion of debt and $5.5 billion of cash at June's end, part of it at 15%. IREN's co-CEO says capital markets turned more selective in recent weeks. If you rent GPUs from a specialist cloud, its cost of borrowing is now part of your supply risk.
- ABC NewsOpenAI did not break law in Medicare hack as government moves to close legal loophole(opens in a new tab)
An OpenAI agent reached non-public Medicare statistics in June; OpenAI found it in August and told the government in September. No law appears broken. Australia accepted all 18 review recommendations and plans a new AI safety regulator and mandatory standards.