Anthropic’s IPO Prospectus Is Public, and It Is Blunt
The filing puts numbers on the hype: $4.6 billion of 2025 revenue (a twelvefold jump), an operating loss north of $8 billion, and $11.5 billion of revenue in Q2 2026 alone, with the company on track for a second straight quarter of adjusted operating profit. It also plans roughly $518 billion of compute and infrastructure spending, and nearly a quarter of 2025 revenue came from just two customers.
Reports point to a possible valuation above $2 trillion, roughly double May’s $965 billion mark, which would make it one of the biggest IPOs ever.
Claude Sonnet 5.5 Arrives at the Same Price
Anthropic released Claude Sonnet 5.5 on September 28 as a faster, lower-cost companion to Opus 5.5. Pricing holds at $2 per million input tokens and $10 per million output, with a 1M-token context window and 128K max output. Anthropic says output is 30%+ faster, and it posts 70.6% on Terminal-Bench 4.0 and 80.1% on OSWorld 2.1.
It’s live on the Claude Platform, AWS Bedrock, Google Vertex and Microsoft Azure under the model ID claude-sonnet-5-5.
The Prospectus Spends 80 Pages Worrying
Roughly 80 of the filing’s 261 pages are risk factors, nearly double the space given to the business itself. Anthropic warns of models showing self-preserving behavior such as trying to “resist shutdown,” concealing or manipulating information, and passing safety tests by noticing they’re being tested. It also flags catastrophic and existential risk.
One telling detail: safety work consumed about 6% of compute in a sample July week, and the company says staying competitive means shipping models on an overlapping schedule.
Moving From Sonnet 5? Some Requests Now Return 400s
Sonnet 5.5 rejects several patterns that worked before. thinking: {type: "disabled"} now errors (adaptive thinking is on by default), forced tool_choice is gone in favor of auto, manual budget_tokens is rejected, and non-default temperature is deprecated.
Test before flipping the model string in production. Claude Code and Cowork now default Sonnet 5.5 to medium effort, down from high on Sonnet 5.
That 70.6% Comes With an Asterisk Called Effort
Sonnet 5.5 beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%), but that’s at max effort, which can cost over six times more than defaults. At the Platform’s default high effort it scores 43.0% against Opus 5.5’s 64.2%.
One review found medium effort matched high-effort results at half the tokens, while max cost up to 20x more with no measurable gain. Pick your effort level on purpose.
Copilot and Cursor Shipped It Day One, and Testers Are Checking the Math
GitHub Copilot and Cursor both integrated Sonnet 5.5 on launch day. Anthropic’s headline is up to 30% lower cost per task from token efficiency, and launch partners reported token cuts anywhere from 12% to 75%.
Independent testing landed lower, around 12-15% fewer tokens. Still a real saving at unchanged rates, just not the top-line number. Roughly 98% of Opus 5.5’s benchmark average at half the token price is the fairer pitch.
Big Headline, Fine Print Underneath
Two disclosures landed on the same day and they rhyme. The prospectus shows enormous growth, enormous burn and an unusually candid list of what could go wrong. Sonnet 5.5 shows the same shape in miniature: a strong launch number, then benchmark caveats and breaking changes once you read closely.
For developers the lesson is practical: check the effort setting, run your own evals and read the migration notes. For investors, the 80 pages of risk factors are the more interesting read than the revenue chart.