But don't worry! DeepSeek is raising prices. #
- Welcome to *. Cautious Optimism, a newsletter on tech, business, and power. Modestly upbeat
Thursday. Going public? Might as well snap up a startup for $6 billion, right? That’s what Anthropic is considering. After all, what’s a single-digit-billion-dollar purchase when you are targeting a possible $2,000 billion valuation? Decart, the reported target, is a company I’ve interviewed before. It does two things: improves GPU efficiency and builds world models. Both of which, you can imagine, Dario is interested in! Now, to work! — Alex
📈Trending Up:More wars for oil…digital letters of marque…major pickups…VCs hating on Anthropic…paying up for the harness…split custody of the Strait of Hormuz…no shit, Sherlock…gaming footage as critical AI training data…📉Trending Down:Wholesale inflation in the United States? …Microsoft’s app-spread…ethics…technology companies killing media? …Databricks’ IPO timing…prediction market accuracy…
DeepSeek raises prices: The latest AI model from Chinese AI lab DeepSeek is out. DeepSeek-V4-Pro is no longer in preview, and its new iteration brings with it a host of benchmark gains. At a cost. Recall that DeepSeek is busy raising billions of dollars, and recently told customers that it would raise prices. Yes indeed. The new V4-Pro model will cost $0.66/$1.98 per million input/output tokens during off-peak hours, and $1.32/$3.96 per million tokens during peak demand. During its preview period, V4-Pro cost $0.435/$0.87 per million input/output tokens. (V4-Flash now costs $0.44/$1.32 during peak periods, far more than its preceding $0.14/$0.28 rate.)
- As we argued a few days back, the era of super-cheap, performant models from China is closing. They cost more now, even if they still represent an attractive, low-cost option even with a multiplied price. - DeepSeek-V4-Pro was impressive before its upgrades and full release; with an intelligence score of 53 per AA, the preview version of V4-Pro was smarter than all but eight AI models.
Everyone is clowning on Google #
Here’s a run of recent news, setting aside updates from Anthropic and DeepSeek that we covered above: **SpaceXAI’s **fresh Grok 5.6 release raised the AI lab’s prominence to near frontier-lab status at a very attractive price point. An open-weight model from South Korea (Motif 3 by Motif Technologies) set an AA intelligence score of 47, besting
MiniMax’s M3.
Microsoft’s recently released coding model has
[already received a major upgrade](https://microsoft.ai/news/mai-code-1-1-flash-br-better-faster-at-a-quarter-of-the-cost/); Microsoft also made its first
*general*model, MAI-Thinking-1,
available to its Foundry customers, though precisely how strong that model is remains a little opaque.
Meta’s own fast-cadence update to its Muse 1.1 Spark model
will be released with open weights soon. The impressive model will help boost the American open-AI frontier forward;
Nvidia’s
latest Nemotron model family is also starting to roll out, bringing even more open-weight heft to these shores.
And then there’s Google. It’s precisely nowhere. Observe the dark green candles:
From market leader at the end of 2025 to getting mogged by a half-dozen labs in less than a year is an accomplishment. Worse, the similarly-performant Gemini 3.6 Flash costs $1.50/$7.50 per million input/output tokens while OpenAI’s GPT-5.6-Luna costs just $0.25/$1.20 for the same number of tokens. I was hesitant to pile on Google for its delayed release of Gemini 3.5 Pro. The AI treadmill is brutal, and every lab misses a step here and there. But every day that passes pushes Google further back in the AI stack ranking; every day that slips by without Gemini 3.5 Pro out raises the bar it must clear to be market-relevant.
It’s time to be a little rude. That OpenAI and Anthropic are beating Google is not a massive shock; they were Alphabet’s rivals in the American frontier AI trivium. But falling behind, variously, SpaceXAI, Moonshot, Alibaba, Meta, DeepSeek, and Z.ai, and nearly getting beaten by a South Korean lab that you hadn’t heard of until this morning is downright embarrassing.
This is not to say that Google is down/out in the AI game; it makes oodles of money via its cloud business, as we recently explored. No matter if its models are good or not, Alphabet shareholders are earning an AI dividend. But those gains could be far larger if the company could sling its own models as easily as it’s finding customers for its TPUs.
Microsoft’s quick improvements to its coding model are especially galling, as it’s the precise place where Google is supposedly working to close a gap to the market. If Google can’t catch up on coding, it likely cannot accelerate its AI development with more AI. Meanwhile, OpenAI and Anthropic are turning their data and model foundations into tools to accelerate building new LLMs, something that Microsoft is imitating.
Precisely what has gone wrong inside Alphabet is not clear. What Google does show us, however, is that no leading lab is inevitable. And having a mountain of compute is necessary, but not sufficient. For now, at least!
The money tornado keeps spinning #
**Sweden rising! **Shortly after vibe-coding (intonation: neutral) giant Lovable announced a $400 million raise at a $13.3 billion valuation on the back of a June-era revenue run rate of $500 million, another Swedish startup is raising at an eleven-figure valuation. Legora, a legal AI unicorn that competes with Harvey and Spellbook, is raising at a $10 billion valuation per the FT, “roughly double the $5.6 [billion] it achieved only four months ago.”
Why? Quick growth: Legora grew from $1 million ARR to $100 million in 18 months,it announced in April.Harvey grew from$100 million ARR in August 2025 to $300 million in June, with the company announcing in July that the second quarter saw its first with “over $100M ARR added.”Spellbook was “on track to hit $100 million USD in annual recurring revenue in 2026, after tripling its revenue over the past year” this March.- Venture investors expect faster portfolio company growth than ever. This trend concentrates capital into fewer startups. Sometimes, even in the same competitive space. I struggle to fully understand ~100x revenue multiples, but as an index fund guy at heart, I am not the target market here.
Bending Spoons earnings: Italian startup undertaker (conglomerate) Bending Spoons reported earnings today. Revenue of $704 million (+126%) beat analyst expectations, though the company’s full-year guidance may have come in light. Regardless, Spoons’ earnings release had several interesting nuggets:
Bending Spoons’ purchased assets are barely growing (in aggregate): Per BSP earnings, “organic revenue growth was 3% in Q2 2026, withTractive andWeTransfer making the largest contributions.” That’sthin. Today, Bending Spoons is growing almost purely on the back of inorganic methods (M&A).Bending Spoons is very proud of its internal tech team: From detailing its ‘Alt-Spooner’ AI agent for staffers that takes into account employee data context, runs on locally-hosted open-weight models, to claiming dozens of product updates atEventbrite andVimeo in a single quarter, Bending Spoons is far more than a PE chop-shop.- Shares of Bending Spoons are down 6% this morning.
AIObs is hot hot hot: Dynatrace is buying Arize. Or if you want that in corporate-speak, an AI-powered observability platform is buying an AI observability platform. Obs now requires AIObs, if you will.
The deal will cost the larger company nearly a billion dollars. Recall that
Cisco bought another AIObs startup,
Galileo,
[earlier this year](https://blogs.cisco.com/news/cisco-announces-the-intent-to-acquire-galileo), and
[.](https://clickhouse.com/blog/clickhouse-acquires-langfuse-open-source-llm-observability)
ClickHouse boughtLangfuse There are still a number of startups working in the space, presenting attractive acquisition targets: Braintrust, Fiddler AI, Arthur AI, and Evidently AI come to mind.
- Why are so many companies building AIobs tools? From a high level, probabilistic systems are inherently observability-primed. And with everyone trying to keep their agents on-task and their AI costs in check, it’s a fertile time to sell tools to help folks understand their AI usage and spend.