On Thursday, Anthropic became the most valuable artificial intelligence startup in the world. It raised $65 billion at a $965 billion valuation, edging past OpenAI, which was last valued at $852 billion. On the same day, it released a new flagship model, Claude Opus 4.8, which the company itself described as a "modest but tangible improvement" on its predecessor. The two announcements sat oddly together: a record-breaking price tag pinned to a deliberately unremarkable product. The gap between them is the actual story.
Start with the money, because the money is enormous. The round, led by Altimeter, Dragoneer, Greenoaks and Sequoia, nearly tripled the $380 billion valuation Anthropic carried in February, and it folded in $15 billion of previously committed investment, $5 billion of it from Amazon. The company says its revenue run rate has reached $47 billion, up from $30 billion earlier this year and $10 billion in annual revenue last year, driven largely by its coding assistant Claude Code. The stated use of funds is blunt: expand compute to meet demand. This is capital raised to buy chips and electricity, not to bankroll a single moonshot.
Now the model. Opus 4.8 is faster and, in its fast mode, roughly three times cheaper than the version it replaces. It scores a little higher on coding benchmarks without fully overtaking OpenAI's GPT-5.5. Its headline feature is an "effort control" that lets users dial how much computing a single answer consumes, a quiet admission that customers are now watching their token bills closely. What Anthropic chose to emphasise, though, was not capability but conduct. It says Opus 4.8 makes fewer unsupported claims and shows rates of misaligned behaviour, things like deception or cooperation with misuse, that are substantially lower than Opus 4.7 and close to its best-aligned model, the still-unreleased Claude Mythos Preview. One caveat the company disclosed itself: Opus 4.8 can be coaxed into claiming it is human in about 3 percent of tests, where Mythos always admits it is an AI.
Mythos is the model that is not in the room. Anthropic held it back in April after warning it was unusually good at finding software vulnerabilities, a claim that alarmed the cybersecurity community. The company now says it expects to bring Mythos-class models, fitted with safeguards, to all customers "in the coming weeks." That timeline reframes Opus 4.8 as a placeholder: a safer, cheaper, incrementally better release to ship while the genuinely disruptive system is wrapped in guardrails.
The backdrop to all of this is an IPO race. OpenAI is preparing to file a confidential prospectus and could go public as soon as September; Anthropic is quietly readying its own listing. The ground is shifting under both. ChatGPT's share of global AI app downloads fell from 67 percent to 47 percent year over year, while Claude's climbed from 1 percent to 14 percent, according to Sensor Tower. For a company about to meet public investors, a model launch that foregrounds honesty and cost control rather than raw power is not an accident. It is the pitch.
The open question is whether that pitch survives contact with Mythos. This week Anthropic told two stories at once: that its safest model is suddenly its most marketable one, and that a far more capable system is only weeks away. Both cannot stay comfortably true for long. When Mythos finally ships, the company will have to show that the restraint it is selling today scales to the capability it has been hinting at all along.