AI Weekly #3 — Aug 10–16, 2026

Four separate capability announcements this week, and not one of them involved a new base model.

Google shipped Gemini 3.7 Flash three weeks after 3.6 and said outright that it did not retrain from scratch. Z.ai shipped GLM-5.3 on the GLM-5.2 base and wrote that scaling post-training was the whole of the work. OpenAI's Ultrafast preview is GPT-5.6 Sol — the same weights, served roughly fourteen times faster on someone else's silicon. Alibaba's Qwen3.8-Max weights finally appeared on Hugging Face, which is a licensing event rather than a training one. The frontier moved four times in six days and the pretraining runs behind it were all finished weeks ago.

The pretrain has stopped being the news

This is worth sitting with, because it inverts how most people still reason about model releases. If capability now arrives through post-training, serving and pricing, then the interval between meaningful releases is set by how fast a lab can run reinforcement environments, not by how fast it can raise capital for a cluster. Gemini 3.7 Flash went from 49.0% to 65.3% on DeepSWE v1.1 and from 17.0% to 30.4% on AutomationBench in three weeks. Z.ai's ExploitBench score more than doubled, from 24.4% to 54.4%, on a base model it had already shipped.

The practical consequence is unpleasant if you build products: a benchmark you ran in July is describing a model that no longer exists, and the price you budgeted against has probably halved. Gemini 3.7 Flash is $0.75 per million input tokens until the end of December and $1.50 after, which means the cheap number is a promotional one and the planning number is the January one.

The second consequence is that Z.ai's decision reads differently. A lab that can produce a large cyber-capability jump without touching its base model has also discovered it can do so faster than it can evaluate the result — which is, in its own telling, why the open weights are staged behind the coding plan instead of shipping alongside it. That is the first time an open-weights lab outside the US has made that argument about its own model in public.

The mark moves to the output

Anthropic's watermarking post is the other side of the same shift. The change is not to the model but to the sampling procedure: token choices that used to draw on arbitrary randomness now draw on a cryptographic key, so the sequence carries a checkable signature and survives being pasted into a document. Files get C2PA credentials instead. It applies worldwide, not only in the EU, because Article 50 was the trigger and a single global behaviour is cheaper than a regional one.

Anthropic is careful about what the mark proves — that Claude took part, not who wrote or owns the result. That distinction will not survive contact with the first university that wires a detector into a plagiarism workflow. Worth knowing now, before someone runs it against your writing.

Money went to the layer above

Databricks took $5bn at a $190bn valuation and Lovable took $400m at $13.3bn, and both told the same story about what the money is for: infrastructure and cost of goods, not model training. SpaceX closed its $60bn all-stock acquisition of Anysphere on 14 August, converting a computing partnership into ownership of Cursor. None of that capital is chasing a bigger pretrain either.

By the numbers

The directory added 23 apps, 9 agent skills and 9 MCP servers this week, and now holds 390 apps, 141 skills and 167 MCP servers. Skills additions fell hard — 9 against 27 last week — and every one of the nine was first-party: OpenSearch, Brave, DuckDB, Figma, Notion and HashiCorp all published their own. The MCP servers were the same picture, with Pulumi, Snyk, New Relic, Miro, Algolia and Amplitude arriving as official vendor artefacts.

Which surfaced a problem worth naming. Of those nine skills, exactly one could be given a verifiable release date. Vendors are shipping skills as production artefacts and versioning them like scratch files — no release notes, no tags, no changelog — so "when did this ship" is usually unanswerable. That is why the picks below are mostly apps: the apps had launch dates and the skills did not. Software that agents load at runtime should be at least as datable as software humans install.

The apps side stayed on one theme all week. Paritok compresses the context an agent sends, Oqoqo benchmarks four harnesses against each other, HarnessRouter abstracts over them, and Spotify open-sourced the workspace it uses to run dozens of sessions at once. The tooling being built now is not for agents that write code — that is assumed — but for the problem of running a lot of them at once without losing track of the cost.