The week of the point release, and a $500 billion bet on the compute underneath

Between 9 and 15 August, the notable model releases were point upgrades. Google shipped one to its cheap model. SpaceXAI shipped one to its flagship. OpenAI shipped no model at all, only a preview of a faster way to serve the one it already has. The largest number of the week came from none of them: it came from NVIDIA, and it was a financing announcement.

That is the throughline worth carrying out of the week. The labs competed on price, latency and how long a model can keep working, not on how much it knows. Meanwhile the capital structure underneath the compute got rearranged.

The releases were increments, and they competed on cost and speed

Gemini 3.7 Flash arrived on 13 August, three weeks after 3.6 Flash, which Google itself flags in the announcement. It carries an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, available through the end of the year. Google reports gains over its own predecessor on FrontierCode 1.1 Main (43.6% vs 34.4%), DeepSWE v1.1 (65.3% vs 49.0%), the GDP.pdf document benchmark (34.0% vs 22.0%), AutomationBench (30.4% vs 17.0%) and WebDev Arena Elo (1588 vs 1538). Every one of those is a Google-run comparison against a Google model; no independent evaluation of 3.7 Flash exists yet.

OpenAI did not release a model. On 13 August it previewed Ultrafast mode, a premium serving tier for GPT-5.6 Sol that its API changelog describes as delivering up to 14 times standard processing speed, available in limited preview to selected customers through a sign-up form. The headline was throughput, not capability — which is itself a statement about where the competition sits.

Grok 4.6 shipped on 12 August aimed at agents that run across many steps, holding Grok 4.5's list price of $2 and $6 per million input and output tokens. By SpaceXAI's own table it scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max and trailing Fable 5 Max at 62. SpaceXAI's announcement page returns a 403 to automated clients from our environment, so that release is carried here through our own coverage from Wednesday rather than a page we could open directly.

What that does to cost planning

For anyone budgeting an agent that runs continuously, the more useful observation is that the price you plan around is increasingly a temporary one. Gemini 3.7 Flash's rate is explicitly introductory and expires at the end of the year. Grok 4.6 advertises an unchanged headline price while the rate structure underneath it is not flat. Ultrafast is a premium tier with no public price at preview. That is three different shapes of "what does this actually cost" inside one week.

The practical consequence is dull but real: for agent workloads the sticker price is now among the least stable inputs in a cost estimate, worth re-checking on a schedule rather than once at model-selection time.

The week's biggest number was financial, not technical

On 10 August, NVIDIA announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent compute financing platforms, aimed at mobilising over $500 billion of third-party capital for AI infrastructure over time. NVIDIA's own subhead describes the goal plainly: turning its compute and full-stack infrastructure "into an investable asset class." Jensen Huang's quote in the release puts it more bluntly still: "In AI, compute is revenue."

Read the structure before the number. These are memorandums of understanding describing an aim, not closed funds. The release does not disclose committed capital, closing dates, or how the vehicles would be structured, and the $500 billion is capital to be mobilised over time rather than money raised. What was announced is an intention to make GPU capacity financeable by institutional investors the way toll roads and data centres already are.

It belongs next to the model releases because both are arguments about unit economics. If capability gains per release are narrowing while inference volume climbs, the scarce asset is the capacity to serve tokens — which is what six of the world's largest capital allocators were just invited to underwrite.

Open weights kept filling the small end

Meta released Muse Glimmer on 10 August, distilled to roughly 30B parameters and published under Apache 2.0 for local agentic use. NVIDIA followed on 11 August with Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts open model for high-volume agent tasks, which NVIDIA says delivers up to 4x faster output and 30% faster agentic task completion than others in its class — again, the company's own measurement. Alongside it NVIDIA released NeMo Switchyard, an open-source routing library for directing each request to a suitable model across a mix of open, proprietary and NVIDIA models.

Hugging Face's summer report, published 14 August, supplies the shape of that ecosystem with numbers instead of impressions. Public model repositories grew from 2.43 to 2.96 million between January and August, and datasets from 711,000 to 1 million. But roughly 85.6% of models have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of all downloads. The report also notes that the two organisations publishing the most new open models this year are the hardware vendors, AMD and NVIDIA, each with more than 200 new model repositories.

Also this week

• Gemini app passes 1 billion monthly users (11 August). Google called it the fastest-growing product in its history, reporting 150 million+ images generated daily and 100 million+ active users on iOS. Every figure is Google's own and was published without methodology.
• Anthropic explained Claude's text watermark (14 August), including where it does not work: short passages, factual writing with few equivalent word choices, proofreading of human text, and code, where the company says the watermark is not applied at all. The explainer is useful mostly because it states its own limits.

The throughline

This was a week in which the interesting engineering was in serving and financing rather than in training, and in which every performance figure published came from the company selling the thing measured. Neither of those is a complaint. But if you are deciding what to run in September, the week gave you far more information about what inference will cost and who will pay to build the capacity than about what the models can newly do.