Anthropic just cut the price of its top model by a fifth. It still costs double what Opus does. Cheap is relative this week. This is Lantern AI Daily Briefing for Tuesday, November 10th. Four stories today: Anthropic's Fable and Mythos 5.2, Google bolting two more systems onto its eighth TPU generation, Meta's Muse Ember going after the budget tier, and Microsoft letting an Oracle compute contract run out. Follow the show and the next briefing lands in your feed on its own.
Pragya Singha Roy, writing in Mint:
The two models share the same underlying architecture but have different safety restrictions, Anthropic says. Cybersecurity and biology protections are incorporated into Claude Fable 5.2, which the general public can use. Claude Mythos 5.2 is designed for trusted organisations engaged in cutting-edge cybersecurity and biomedical research.
Both shipped Monday. Fable 5.2 is on the Claude Platform, in Claude Code, and on AWS, Google Cloud and Microsoft Azure. Mythos stays behind Anthropic's Cyber and Life Sciences verification programs. Anthropic says output is up to thirty percent faster than the 5.1 generation, and list price drops from ten and fifty dollars per million input and output tokens to eight and forty. An optional fast mode runs up to two and a half times faster, at sixteen and eighty. The line I'd actually underline: Anthropic's own guidance still says start on Opus 5.5 for everyday coding. SH Consulting points out Opus is half Fable's price, at four and twenty. Fable is for the hard, open-ended jobs. A vendor telling you not to default to its flagship. I'll take it. The capability claims are Anthropic's, from prerelease testing: about a thousand Fable agents formalizing the proof of Fermat's Last Theorem in thirty hours, and two open problems in black-hole physics solved. It reports forty-nine point one percent on FrontierCode 2.0 at its extra-high effort setting, against forty-five point five for Opus 5.5. Anthropic also says it deliberately held back the research skills most likely to speed up AI development itself, like accelerator-kernel optimization, to roughly Opus 5.5's level. First outside-ish numbers are in. Cursor says Fable 5.2 scored fifty-six point nine percent on its own CursorBench at default effort, up from forty-nine point two for 5.1. And The Context's Hacker News roundup flags Artificial Analysis ranking it first of seventy-four models with a sixty-five, at twenty-six dollars fifty-four and forty-five minutes per index task. Multiply that by your task volume before you move a pipeline. Reuters frames the timing around the planned IPO. On Monday's show we noted Reuters' report that investors were questioning whether a roadshow could start this week. A flagship launch with a price cut is a handy slide for that deck.
Wylie Wong, writing in Data Center Knowledge:
To meet growing demand for AI inference, Google Cloud on Monday beefed up its Tensor Processing Unit family with two new chips: a more secure version of its TPU 8t accelerator and a higher-capacity version of its TPU 8i inference chip.
New systems, not new silicon. TPU 8tx uses the same chip as 8t, pairs it with AMD EPYC hosts, and keeps data and model weights protected even while they're being processed. Google claims up to two times lower latency than 8t with a near-zero performance penalty. TPU 8i Max packs two 8i chips into one package: double the compute, SRAM and HBM bandwidth, and three hundred eighty-four gigabytes of HBM, per Google Cloud's Mark Lohmeyer. The figure I'd circle is in Amin Vahdat's post: TPUs now have nearly a hundred customers, frontier labs and trading firms among them. Reuters notes TPU sales contributed meaningfully to a September quarter in which Google Cloud revenue more than doubled to thirty point seven one billion dollars. Lohmeyer also said select customers will be able to run 8tx in their own data centers. IDC's Dave McCarthy says a strong confidential-compute story gives Google an edge while Nvidia is still working out its own approach. My caveat: nothing we read gives availability dates or pricing, and the performance claims are Google's. The software matters more to me anyway. Google says PyTorch code and Hugging Face models now run natively on TPUs through TorchTPU, which it open-sourced last month. If that holds up in practice, the lock-in argument for Nvidia gets thinner. And 8i Max fits twice the silicon into the same rack footprint. When power and permits are the binding constraint, compute per watt is the spec that counts.
Meta, in its launch post for Muse Ember:
Muse Ember delivers more intelligence per dollar than Claude Haiku 5.5 or GPT-6 Luna at any reasoning setting: it matches Haiku's best score at about a third of the cost and beats every Luna setting.
This is the one to benchmark this week. Ten cents per million input tokens and fifty cents output on the Meta Model API, for prompts up to two hundred thousand tokens. A one-million-token context, up to five hundred twelve thousand tokens out, and text, images, audio and video in. It's built for agent grunt work: tool calls, routing, extraction, coding subtasks. Read the claim carefully. On the Artificial Analysis index, Ember and Haiku 5.5 both score thirty-six. The win is cost per task, and Meta says its Ember cost figures are estimates. On Meta's own table, Ember leads AutomationBench at thirty-seven point eight percent to Haiku's thirteen point nine, but trails on FrontierCode 2.0, twenty-seven point two to thirty-four point nine. Weights stay closed. That FrontierCode gap is the number I'd watch, because that test judges whether the code is mergeable. Among the partner results Meta chose to publish, You.com reports a p95 latency of two point two seconds on its coding tasks, with thirty-six to forty-eight percent fewer output tokens than leading open-weight models. Same score, a third of the bill. That's how small models win enterprise contracts. Just measure retries on your own data, because a cheap model that needs three attempts isn't cheap.
Anzar Mehraj and Aditya Soni of Reuters, relaying Bloomberg News:
Microsoft is winding down an arrangement to rent AI computing from Oracle as part of a larger reassessment of billions of dollars in cloud partnerships, Bloomberg News reported on Friday, citing people familiar with the matter.
This broke Friday, November 6th, and it's our first time on it. Per Bloomberg, Microsoft declined to renew and has been switching off the rented capacity in recent months, because the Oracle hardware ran an older generation of Nvidia chips that no longer met its needs. The deal was announced in 2023 to power the AI version of Bing, and its value was never disclosed. Oracle's response: a spokesperson told Reuters demand for its AI infrastructure remains strong, including for earlier-generation GPUs. Investing.com adds that Oracle said in September it was renewing contracts for four-year-old GPUs at a twenty percent premium. Microsoft didn't comment. The pattern is the story. The day before, Bloomberg reported Microsoft tightening oversight of data center projects run by neoclouds including IREN, Nebius and Nscale. IREN said it delayed some deliveries after Microsoft asked for network design changes, pushing its target for more than four billion dollars of annualized revenue capacity from year-end into January, possibly late February. This is the depreciation question nobody puts on the slide. The GPUs age out of a hyperscaler's specs long before the building does. Compute commitments keep stacking up, Anthropic's S-1 alone lists about eight hundred ninety-five billion dollars, and Microsoft just showed a buyer can walk once the hardware stops fitting. Two caveats. Bloomberg's sourcing is unnamed people, and this is one aging contract, not Microsoft retreating from AI. Vested puts its commitments to rent outside capacity at more than eighty billion dollars.
If the TPU and Oracle segments were your favorite part of today, try our sibling show, The Data Center Daily. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and the energy markets being reshaped by intelligence at scale. Search for it wherever you're listening.
What we're watching: independent evaluations of Fable 5.2 beyond Anthropic's own charts, and whether Google attaches dates and prices to the new TPUs. We're back tomorrow with the next round. That was Lantern AI Daily Briefing, one of the shows from Lantern Podcasts. See you Wednesday.