Two of today's stories are about prices with dates attached to them, and two are about what you get to see. Google's cheapest new model is cheap until December 31 and then it is exactly twice as expensive, which makes every comparison table built this week quietly wrong for next year. A lab in Abu Dhabi released six models under Apache 2.0 and shipped the training data with them, which almost nobody at that scale does. McKinsey found that a third of organizations have already killed a software purchase because a coding agent could build the thing. And a single Anthropic contract doubled one datacenter company's backlog to $103 billion, which tells you where the money behind your API bill actually goes.
No sponsored or affiliate links in this digest — the links below are sources only.
Story I
Gemini 3.8 Flash is on introductory pricing for 114 more days, and the standard rate is exactly double
Google shipped Gemini 3.8 Flash on September 2, in a crowded week that also brought Claude Fable 5.1 on the 1st and GPT-6 Astra on the 3rd. The launch is not the interesting part. The interesting part is a date on the official pricing page: standard-tier input is $0.75 per million tokens through December 31, 2026, and $1.50 starting January 1, 2027. Output is $3.75 through the same date, then $7.50.
That is not a price increase so much as a discount with a stated end, which is arguably more honest than the alternative — but it has the same effect on your invoice. Counted from today, the cheap rate has 114 days left. Anything you benchmark this week against a flat-rate competitor is being scored on a number that expires before most of us finish a migration. The comparison that actually holds is the one you run twice: once at the introductory rate to see what you save this quarter, and once at the standard rate to see what you signed up for.
For builders
Take your August invoice for anything running on 3.8 Flash and double it. That is January, with zero change in your usage. Do that arithmetic before you write the model name into a config that ships to production, because the migration you would rather not do in December is much easier to plan in September. Two habits worth keeping past this one model: grep your provider's pricing page for the word "through" before you quote a number in a comparison, and pin the model version explicitly rather than an alias, so a rate change and a behavior change never arrive on the same day.
Story II
IFM published six models with the training data attached — the part of "open source" almost everyone skips
The Institute of Foundation Models, launched by MBZUAI in May 2025 with labs in Abu Dhabi, Silicon Valley and Paris, released the K2 Horizon family on September 3. Six models: 0.9B for very constrained targets like watches and glasses, 3.7B and 7B for phones and on-device work, a dense 32B for local hosting and on-prem servers, and two sparse models — 36B with 4B active, and 375B with 23B active. All Apache 2.0, on Hugging Face, running under vLLM and SGLang, with API access through Compass, Cerebras, AWS and Nebius.
What separates this from the usual release is that the weights come with the code, the training data, and the methodology. IFM founder Eric Xing put it plainly: open source is more than open weights, and every model ships with its training data, recipe and evaluations. That is the line most labs decline to cross, which is why "open" has quietly come to mean "you may run it, but you may not know what is in it." On performance, the lab claims state of the art at the 0.9B, 3.7B and 7B scales across reasoning, math, coding and agentic tasks — that is a self-reported claim on a release five days old, and it deserves independent evaluation before anyone repeats it as settled.
For builders
If you have ever had to answer a procurement or compliance question about what a model was trained on, this is the first release at this size where the answer can be a file rather than a shrug. Concretely: the 36B-A4B sparse model is the one to try first on a single box, because you pay inference cost on 4B active parameters while getting the quality of a much larger model — that ratio, not the headline parameter count, is what determines whether it fits your GPU. Pull it, run it against your own eval set rather than the benchmark table, and if the numbers hold you have a fallback that no vendor can reprice on you in January.
Story III
A third of organizations have already killed a software purchase because a coding agent could build it
McKinsey's State of AI 2026 survey reports that 32% of organizations decided against buying at least one software product or feature because they could build it internally with agentic coding tools. Read the unit carefully before you extrapolate: it is one product or feature, at least once — not a wholesale replacement of the SaaS stack. But the distribution is where it gets pointed. Among the roughly 6% McKinsey classifies as AI high performers, nearly half skipped a purchase, against 31% of everyone else. By sector: technology 41%, healthcare 39%, professional services and energy 38%. And among organizations above $1 billion in revenue, the share scaling agents in one or more functions went from 27% to 40% year over year.
The build-versus-buy line has moved before and moved back, so the honest framing is not that SaaS is over. It is that the build side of the ledger got cheaper to start and did not get cheaper to keep. A feature an agent writes in an afternoon still needs an on-call rotation, a security review, and someone who understands it in eighteen months. That cost does not show up in the survey, and it will not show up in your first sprint either.
For builders
Next time you are in a build-or-buy conversation, insist on pricing the run cost, not the build cost — because the build cost is the number that just collapsed and the run cost is the number that did not. A workable rule: write down the annual vendor price, then write down your own estimate of hours per year for maintenance, incident response, dependency upgrades and the security review, and multiply by a loaded hourly rate. If your internal number is not less than half the vendor number, buy it, because your estimate is optimistic. And if you sell software, note which sectors are furthest along on that list — technology at 41% is your own customers deciding they can write your product.
Story IV
One Anthropic contract doubled a datacenter company's backlog to $103 billion — a useful reminder of what sits under your API bill
Nscale is closing a pre-IPO round of up to $3.5 billion — roughly $2 billion from Nvidia, plus convertible notes led by Third Point with a conversion cap at a $30 billion valuation — ahead of a New York listing, with Goldman Sachs and JPMorgan advising. The number being pitched is a contracted revenue backlog that went from $51 billion to $103 billion in about a month. Nearly all of that jump is one agreement: a six-year, $45 billion deal with Anthropic signed around August 26 for 460 megawatts at a West Virginia campus, roughly 194,000 Nvidia Vera Rubin GPUs, with the facility projected online by the end of 2027.
Treat the headline figure with care, and the company's own framing says as much: $103 billion is total contracted revenue across the life of those agreements, not annual revenue, and backlog numbers in this sector are illustrative rather than forecasts. Two things are still worth taking from it. Capacity for the next two years is being committed now, by contract, before it exists. And the compute behind a model you call over HTTPS increasingly traces back to a single named site with a single delivery date on it.
For builders
This is the story that argues for portability rather than any particular action today. If a meaningful part of your product depends on one provider's inference, the practical hedge is not switching — it is making switching cheap: keep prompts and tool definitions in your own repo rather than a vendor console, put an adapter between your code and the SDK, and keep one alternative model wired up and passing your eval suite even if it never serves a request. The point of the second model is not that it is better. It is that you find out what breaks on a Tuesday of your choosing rather than theirs.
— The Vibe Gate news desk. We read the firehose so you can keep building.