Two Series A rounds in two days for the same idea: send each call to the cheapest model that can handle it
Sapiom, a platform for shipping and running AI agents, raised a $35M Series A announced 5 August, led by Dragonfly with Accel, Menlo Ventures, Okta Ventures and — notably — Anthropic among the backers, six months after a $15M seed. Its core product is a router that matches each model call to the cheapest model that can do the job instead of defaulting to the most expensive one. Sapiom says it cut one customer's monthly model bill from about $1.2M — most of it on Anthropic — to roughly $100K, close to a tenfold drop. A day later, the Berkeley-founded lab Naïve raised $28.5M, led by Nexus Venture Partners with Y Combinator, to build infrastructure for “autonomous companies”; two of its four stated research pillars are routing calls to cost-effective models and delegating work to subagents.
The layer attracting money this week is not a new model — it is the software that decides which model runs each request and keeps the bill down. Anthropic backing a startup built to shrink Anthropic bills is a fair signal of where the value is landing. For a regulated European buyer, that routing decision is also a control point: which model, from which provider, under which jurisdiction, handles each call. The cost question and the sovereignty question turn out to be the same question — where does this request go.
Ant Group open-sources a 124-billion-parameter model you can run yourself
Ant Group's open-weight lab, inclusionAI, released Ling-3.0-Flash on 5 August under the MIT licence: a 124-billion-parameter mixture-of-experts model that activates just 5.1 billion parameters per token, with a 256K context and a hybrid reasoning mode. The weights are on Hugging Face and ModelScope in two builds — BF16 at 255GB and an FP8 version at 128GB — and self-deploy through SGLang. A smaller Ling-3.0-Tiny shipped alongside it.
It is not a frontier model: it scores 38 on the Artificial Analysis Intelligence Index, against roughly 60 for the current leaders. But that is well above the median for open-weight models near its size, and the FP8 build roughly halves the memory footprint to about 128GB — small enough to serve on a single high-memory node rather than a large cluster. That is the practical point for a European team: an MIT-licensed model you download, run on your own infrastructure with no per-token vendor bill, and no prompt or output leaving your own network — at a size an ordinary server rack can actually host.
OpenAI's math results ship with proofs a machine can check
OpenAI said on 1 August that an unreleased internal version of its next model, Astra, produced ten results across mathematics and theoretical computer science, each a problem open for at least a decade, published as a roughly 249-page manuscript with machine-checkable Lean 4 certificates for every result on GitHub. The headline is the first explicit construction of a non-sofic group — a question that had stood since Mikhail Gromov introduced soficity in 1999. OpenAI put the total compute for all ten problems combined at about $2,000, not per problem. Manchester mathematician Thomas Bloom, who curates the Erdős-problems catalogue, called it “big news.”
The durable point for anyone who has to trust an AI output is not the maths — it is that every result comes with a Lean 4 proof a machine can verify without taking OpenAI's word for it. Both OpenAI and DeepMind now use Lean 4 as their verification backend, so AI-generated maths is converging on one checkable scoreboard where labs can compare results without trusting each other's claims. Verifiable-by-machine is the property enterprises will want wherever “the model said so” is not a good enough answer.
Quick Hits
- Defence-manufacturing money keeps flowing. — US defence-tech startup Hadrian raised about $1.37B at a roughly $7.9B valuation on 6 August, with WCM, Valor Equity Partners, Baillie Gifford and 137 Ventures leading and a16z, Founders Fund and Lux participating — a sign automation capital is moving into physical supply chains, not just software.
- Google is switching off its Imagen 4 image API. — The imagen-4.0 generate endpoints — standard, ultra and fast — shut down on 17 August and will return hard errors afterwards; the recommended replacement is the gemini-3.1-flash-image model, so any team still calling those endpoints has days, not weeks, to migrate.
- The EU AI Act's high-risk rules are not live yet. — The enforcement powers that took effect on 2 August cover general-purpose models and the Act's prohibited-practice and transparency rules; the separate obligations for “high-risk” AI systems did not start then — they were postponed to 2 December 2027 under the EU's “AI Omnibus” amendment, so any compliance plan built around an August 2026 high-risk deadline needs re-dating.
- FLUX 3 Video reaches general availability. — Black Forest Labs took FLUX 3 Video to GA on 4 August — clips up to 20 seconds with native audio and dialogue, 720p natively or 1080p upscaled, at about $0.17 per second of output on OpenRouter, or roughly $3.40 for a full 20-second clip before re-rolls.