Anthropic dropped Opus 4.8 today, just 41 days after 4.7.

The fast turnaround may have something to do with the chilly reception Opus 4.7 got, which some users found disappointing.

Whatever the reason, Anthropic clearly wasn't hanging around.

The big benchmark wins are in Claude Code. Dynamic workflows let it tackle genuinely large-scale problems, and fast mode is now three times cheaper than before.

Yet, what people are REALLY noticing is that the model is 4x less likely than 4.7 to miss flaws in its own code and push them through without saying anything.

Which if you've ever had Claude confidently ship broken code at 2am, you'll understand why that matters.

Oh, and buried in the same announcement: Anthropic closed a $65 billion (!?!?!) Series H at a $965 billion valuation. That puts them ahead of OpenAI for the first time. In February they were valued at $380 billion.

Three months later, nearly a trillion. Wild times.

They also confirmed that Mythos, their most powerful model which has only been available to a handful of organisations, is coming to everyone "in the coming weeks."

That's probably the more important news, honestly.

Because Opus 4.8 is good.

But Mythos is the one the people are actually waiting for.

Speaking of valuations…

Claude Code and Codex are hogging all the headlines around AI coding, but Cognition just raised over $1 billion at a $26 billion valuation.

Eight months ago they were valued at $10.2 billion. Revenue run rate is now $492 million, up from $1 million ARR in September 2024. The growth is pretty insane by any measure.

Goldman Sachs, Mercedes-Benz and the US government are customers.

Devin has gone from a viral demo to actual enterprise software quietly in the background while everyone argued about Claude vs Codex.

Wu has called Cognition the "Switzerland" of this phase of AI, staying model-agnostic rather than betting on any single provider (a bit like Cursor). Given how quickly models are being switched, that's probably the right call.

The real question is whether that positioning holds as Claude Code and Codex get more aggressive.

For now though, the investors clearly aren't worried.

The labs aren't staying in their lane and everyone in AI coding knows it. The question isn't really whether they'll try to eat the space, it's who has enough of a moat to survive when they do.

Cursor probably does. They went from $100m ARR in January 2025 to $2 billion by February 2026 and developers aren't leaving. The IDE is sticky in a way the terminal isn't.

Most people run Claude Code for heavy lifting and keep Cursor open all day for everything else. That's not a dynamic that disappears overnight.

The vibe-coding tools are a harder case. Lovable and Bolt's pitch is that non-developers can build real products without writing code. That was genuinely new two years ago. Now the labs are making the exact same promise, with infinitely more resources. The window for that category is closing.

The deeper problem for everyone is dependency. Any tool reliant on frontier lab APIs is permanently exposed to those same labs using that access as leverage. Cursor saw it coming and built their own model. Most smaller players haven't, and probably can't afford to.

The labs won't eat everyone. But if your entire product sits on top of models that Anthropic and OpenAI are also shipping directly to users, you need a real answer to why someone keeps paying for the wrapper.

A lot of companies don't have one.