This week, frontier models got cheaper while the compute bill grew
Anthropic and Google cut top-tier prices, narrower models arrived, and an IPO, infrastructure earnings and NVIDIA's CEO exit showed who funds the hardware.
涵盖2026-11-08至2026-11-14
Two things happened at once between November 8 and 14. At the model layer, prices went down and products got narrower: the top tier got cheaper, and new models were built to do one job well. At the capital layer, the bill for AI compute got bigger and its leadership less settled. The week's IPO, earnings and boardroom news all concerned who pays for the hardware and on what terms.
The top tier got cheaper, and split in two
Anthropic's Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Fable 5.1, according to its developer documentation. It keeps a 1M-token context window. Claude Mythos 5.2 is the same underlying model, offered only to organizations vetted through Anthropic's verification programs. A day later, Google opened Gemini 4 Carbon to all paying developers at an introductory $2/$10 per million tokens. After November 30 the price rises to $4/$20. Both companies report benchmark gains from their own testing, and no independent evaluation has confirmed those figures.
For developers, the price cut matters less than Fable 5.2's breaking changes. A request that edits an earlier turn now fails with a 400 error. Thinking blocks are bound to the model that produced them, so a router that moves a conversation from Fable 5.2 to another model loses that reasoning. Forced tool choice is unsupported. If you build your own message history or route between models, test before you switch. Code that only appends to the history is unaffected.
Below the frontier, models built for one job
Google's Gemini Decisions API runs Gemini Decision 1, a model that doesn't write text. It scores yes/no, multiple-choice and rubric questions about an input of up to 256K tokens and returns probabilities. Google charges $0.10 per million input tokens and nothing for output. That makes Google the second major lab, after OpenAI's Decisions API in October, to sell a classifier-style endpoint for routing, policy checks and grading. Google's latency and calibration claims come from its own benchmarks.
Tether's QVAC team released BotPsy, a 0.8B model for on-device tool calling. Tether reports that it roughly doubles its Qwen3.5 base model on agent benchmarks, and it also reports a 31% unsafe rate under prompt injection. Two days later, llama.cpp and ggml reached v1.0.0 and adopted semantic versioning: a breaking change to the public C API now requires a major-version bump. At the hardware level, NVIDIA, Rebellions and SK Telecom plan racks that put NVIDIA GPUs on prefill and Rebellions NPUs on decode, linked by NVLink Fusion. The pattern across all three is to match each task to the cheapest component that can handle it.
Agents moved inside the perimeter
Anthropic's product news this week was about where agents run and under what rules. Self-hosted Claude Managed Agents, in private beta, lets customers run the harness, sandboxes, credentials and session history in their own cloud or data center. Claude Meetings, in beta for Team and Enterprise on Google Meet, adds Claude to calls as a visible participant that takes notes and builds documents. It saves transcripts but not audio. The company's revised Usage Policy also took effect on November 12. It adds requirements for Claude connected to physical hardware, including operator oversight and a safe state if the connection drops.
The compute bill, in public
Anthropic launched its IPO roadshow on November 10. It is offering 825 million Class A shares at $70 to $78 under the ticker ANT. The amended S-1 adds a $26.0 billion concurrent private placement bought by Google, NVIDIA, AMD, Broadcom, Micron, Microsoft and TSMC Partners. Those are the companies that sell Anthropic its chips and cloud capacity. Anthropic estimates net proceeds of about $84.6 billion. CNBC reports, citing sources, that pricing is expected November 18. The company has not confirmed that date.
The infrastructure earnings came in strong, and they showed how much the build-out is costing. CoreWeave reported Q3 revenue of $3.661 billion and a $132.5 billion backlog. It also posted a $709 million net loss, with $846 million in net interest expense for the quarter. Nebius grew revenue 571% to $980.8 million and spent $7.3 billion on property, equipment and intangible assets in the quarter. Cisco raised its expectation for AI infrastructure revenue from hyperscalers to $9 billion in fiscal 2027, and Applied Materials reported record quarterly revenue of $10.41 billion.
Then on November 12, NVIDIA said Jensen Huang had resigned as CEO and director, effective immediately. CFO Colette Kress is interim CEO, and Harvey C. Jones becomes the board's first chair. Four days earlier, the board had opened an independent investigation into reports that Huang was the target of a foreign intelligence effort. It said it had no reason to believe national security or NVIDIA secrets were compromised. TechCrunch reported that NVIDIA shares fell nearly 12% in early trading on the 12th, with CoreWeave and Nebius also down. NVIDIA still plans to report results on November 17.
What it adds up to
Per token, frontier intelligence got cheaper this week, and each company's most capable version stayed behind a verification gate. Developers have more ways to avoid paying frontier prices: scoring endpoints, sub-1B local models and a stable llama.cpp. Meanwhile, Anthropic's IPO structure, the suppliers' private placement and the neoclouds' interest bills all show compute being financed by the companies that sell it. Next week tests that arrangement twice: NVIDIA reports under an interim CEO, and Anthropic's pricing is reported for the following day.
为什么重要
- Fable 5.2's breaking changes affect anyone who edits message history, forces tool choice or routes conversations between models.
- Scoring endpoints and small local models offer cheaper options for routing, grading and simple tool calls than a frontier model.
- Anthropic's IPO and the neocloud results show that much of AI compute is financed by the companies that sell the chips and cloud capacity.
要点
- Claude Fable 5.2 costs $8/$40 per million tokens, down from $10/$50; Gemini 4 Carbon is $2/$10 until November 30, then $4/$20.
- Google's Decisions API charges $0.10 per million input tokens and nothing for output.
- Anthropic is offering 825 million shares at $70 to $78, plus a $26.0 billion private placement from its suppliers.
- CoreWeave reported $3.661 billion in revenue and a $709 million net loss; NVIDIA reports November 17 under interim CEO Colette Kress.
来源
- Anthropic一手来源Introducing Claude Fable 5.2 and Claude Mythos 5.2anthropic.com
- Anthropic一手来源What's new in Claude Fable 5.2platform.claude.com
- Google一手来源Gemini 4 Carbon is now availableblog.google
- Google一手来源Introducing the Gemini Decisions APIblog.google
- Tether QVAC (Hugging Face)一手来源BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Modelhuggingface.co
- ggml-org (Hugging Face)一手来源llama.cpp v1.0.0: A stronger foundation for the local AI ecosystemhuggingface.co
- NVIDIA一手来源NVIDIA, Rebellions and SK Telecom announce collaboration on sovereign AI infrastructure with NVLink Fusionnvidianews.nvidia.com
- Anthropic一手来源Operate Claude Managed Agents inside your security perimeterclaude.com
- Anthropic一手来源Introducing Claude Meetingsclaude.com
- Anthropic一手来源Updates to our Usage Policyanthropic.com
- Anthropic一手来源Anthropic launches roadshow for proposed initial public offeringanthropic.com
- U.S. SEC一手来源Anthropic PBC Form S-1/A (Amendment No. 2)sec.gov
- CNBCAnthropic sets price range of $70 to $78 in upcoming IPOcnbc.com
- NVIDIA一手来源NVIDIA Announces CEO Transitionnvidianews.nvidia.com
- NVIDIA一手来源Statement from NVIDIA Board of Directorsnvidianews.nvidia.com
- TechCrunchNvidia's Jensen Huang resigns; chipmaker names CFO Colette Kress as interim CEOtechcrunch.com
- CoreWeave / U.S. SEC一手来源CoreWeave Reports Strong Third Quarter 2026 Results (8-K Exhibit 99.1)sec.gov
- Nebius / U.S. SEC一手来源Nebius reports third quarter 2026 financial results (6-K Exhibit 99.1)sec.gov
- Cisco / U.S. SEC一手来源Cisco Reports First Quarter Earnings (8-K Exhibit 99.1)sec.gov
- Applied Materials / U.S. SEC一手来源Applied Materials Announces Fourth Quarter and Fiscal Year 2026 Results (8-K Exhibit 99.1)sec.gov
- weekly-recap
- api pricing
- model selection
- coding agents
- ai-infrastructure
- ipo
- earnings
- Anthropic
- NVIDIA
- CoreWeave
- Nebius
- Cisco
- Applied Materials
- Tether
- Rebellions
- SK Telecom
- Claude Fable 5.2
- Claude Mythos 5.2
- Gemini 4 Carbon
- Gemini Decision 1
- BotPsy-0.8B