If you set your AI compute budget in October, some of it is already out of date. In the last ten days a GPU cloud raised its list prices, Google confirmed that its newest models double in price on 1 December, and OpenAI put an end date on a promotion. Only Anthropic moved the other way. Here's what happened and what we'd do about it.

1. Lambda raises on-demand GPU prices from 16 November

Lambda's pricing page now lists the H100 SXM at $4.59 per GPU-hour, the B200 at $7.69 and the A100 80GB at $3.19. The page says these prices "take effect November 16, 2026, at 00:00 UTC". Until then, current prices are only visible to logged-in customers.

Lambda's page doesn't show the old prices. Pulse, which tracks pricing pages, compared captures from 2 and 9 November and records the H100 moving from $3.99 to $4.59, the B200 from $6.69 to $7.69 (+15%) and the A100 80GB from $2.79 to $3.19.

Why it matters: the usual assumption is that older GPUs get cheaper once a new generation ships. Here a neocloud raised the price of the older H100 by about the same proportion as the new B200. If you rent Lambda on demand, your bill goes up next week. For how Lambda now compares with seven other clouds, see our H100 price comparison.

2. Gemini 4 Carbon goes GA, and Gemini 4 pricing doubles on 1 December

Google made gemini-4-carbon generally available on 10 November. Its pricing page lists Carbon and Argon at $2 in and $10 out per million tokens "through November 30, 2026", then $4 in and $20 out "starting December 1, 2026". Batch and Flex tiers are half those rates.

Separately, Google deprecated Gemini 2.5 Pro and 2.5 Flash on 2 November. From 1 December, requests to them are routed to gemini-3.8-flash and gemini-3.5-flash-lite and billed at those models' rates. The changelog lists breaking API changes.

Why it matters: any cost model built on November's Gemini 4 prices will undercount December's spend by half. If you're still on 2.5, you get a forced migration that changes cost and behavior at once. Test the replacement models before the routing switch does it for you.

3. Anthropic cuts frontier pricing with Claude Fable 5.2

The other direction. Anthropic released Claude Fable 5.2 on 9 November. Its release notes say it costs "$8 USD per MTok of input and $40 USD per MTok of output", down from $10 and $50 for Fable 5.1. That's 20% cheaper on both. A fast mode offers "up to 2.5x higher output tokens per second" at $16 and $80. The release notes say the 5.2 models "require 30-day data retention".

Why it matters: the top tier got cheaper in the same week that Gemini 4's mid-tier price was confirmed to double. If you have zero-data-retention requirements, read the retention terms before switching.

4. OpenAI: GPT-6 Cyber launches, and GPT-5.6 Sol's promotion has an end date

OpenAI's 5 November changelog added gpt-6-cyber, a restricted-access model listed at $20 in and $100 out per million tokens on short context. More relevant to most budgets, the pricing page now says "GPT-5.6 Sol's promotional pricing is available until November 30, 2026." The page doesn't say what Sol will cost after that.

Why it matters: this is another price with an end date. If Sol is in your stack, plan for a change after 30 November and watch for the new rate.

5. Moonshot publishes Kimi K3.1 Pro weights at 2.8 trillion parameters

Moonshot's model card on Hugging Face lists 2.8T total parameters, 104B active and a one-million-token context, with weights quantized to MXFP4. TechNode reported the release on 5 November. It ships under a custom "Kimi K3.1 Pro License", not a standard open-source licence. TechNode also reports that Huawei has announced support for serving it on Ascend hardware.

Why it matters: this is a frontier-scale model you can download and run inside your own perimeter, which matters for sovereign and private deployments. But at this size it needs multiple nodes or a full rack, not one server, so "free weights" doesn't mean cheap to run. Have a lawyer read the licence before any commercial use.

6. Power is the bottleneck, and the big buyers are building their own

Three stories from the same week point the same way:

  • Google announced a $17 billion data center in Wyoming, powered by a 702 MW natural-gas fuel-cell project contracted with AEP. Google says it "will pay for 100% of the power we use".
  • Drax, Brookfield and Radiant announced a partnership to develop a data center of more than 1 GW at a UK power station. Phase 1 is "up to 100MW, operational from Q1 2028". It still depends on planning approval, a tenant and a final investment decision.
  • Anthropic is reportedly buying 1.7 GW of natural-gas generators for $3.4 billion, according to Fortune citing the Financial Times. Neither company had commented when Fortune published. Treat this one as unconfirmed.
When hyperscalers start building their own power stations, it tells you the grid can't supply power fast enough.

Why it matters: if power, not chips, sets the pace of new capacity, the GPU-hour price is unlikely to fall as quickly as chip supply grows. Lambda's increase this week fits that picture, though one provider's price change doesn't prove a trend. For buyers outside these mega-deals, our reading is that reserved capacity at today's prices looks more attractive than it did a quarter ago. That's an editorial judgment; check it against your own quotes.

7. Open-source serving got faster again

Both main open serving engines shipped big releases in early November. vLLM v0.33.0 reports a speculative-decoding change that took one model on GB300 "from 1373 to 3390 output tok/s in the PR's benchmark" and "39-48% lower mean TTFT at 100k input on 8xB200". Its builds are now CUDA 13 only. SGLang v0.5.23 reports "up to 44% faster single-user decode on GB300 and 75% on MI350X", plus TPU7x support.

Why it matters: these are the projects' own benchmarks on specific setups, not numbers to plan around. The direction is what counts: the throughput you get from a GPU you already rent keeps rising. That's the number that decides whether self-hosting pays, so re-measure after you upgrade. Plan the CUDA 13 driver upgrade before you need the new vLLM.

What we're watching next: Lambda's prices going live on 16 November. Whether any other GPU cloud follows. The post-promotion price of GPT-5.6 Sol. Gemini 4's price switch on 1 December.