Albacel Reads Issue #6

Reads ·

Issue #6

Covers November 17 – November 18, 2026

Choosing what runs where: enterprises go multi-model and open-weight, two new workhorse models compete on cost per task, and the serving layer gets faster for agents.

This issue is about choosing what runs where. The Databricks AI Index is the clearest evidence yet that enterprises have gone multi-model: open-weight models now carry almost a third of the token traffic on its gateway, most companies now use more than one provider, and the model with the cheapest tokens can be the most expensive per task. Gemini 4 Neon and GPT-6.2 arrived within two days of each other to compete for that workhorse tier, while Cloudflare’s open Clairic and vLLM’s MonoKernels work on the cheaper, faster layers underneath. If you read two, start with the Databricks numbers and Google’s test of its agent anomaly detector on benchmarks it didn’t write.

The posts

vLLM

Keeping Agents Fast: High-Interactivity Decode with FlyDSL MonoKernels in vLLM

Key number 1.20–2.26× faster decode at concurrency 1 across six merged vLLM changes

At small batch sizes, much of a decode step goes on starting and draining many short GPU kernels rather than on useful work. MonoKernels, written in AMD’s FlyDSL, run several dependent stages of a decoder layer in one persistent launch: thread blocks move on to the next stage without a new launch, weights for later stages are prefetched, and results pass forward through tagged “mailboxes” as soon as they are ready. In vLLM the path is opt-in (--monokernel), runs per layer and per step, and is merged for five models on AMD Instinct GPUs; on MI325X one decoder layer goes from 32 kernels in 320 µs to 2 kernels in 115 µs. Across six merged changes, interactivity at concurrency 1 improves 1.20× to 2.26×, with accuracy within noise. The gains shrink as load grows: for DeepSeek-V4.1-Flash at concurrency 4, one TP8 server delivered 26% less throughput than two TP4 replicas.

Inference stackBuilding agents

Read at vLLM (opens in a new tab)

Google DeepMind

Introducing Gemini 4 Neon

Koray Kavukcuoglu

Key number 12.8% → 39.0% SWE-sweep score, mini-SWE harness versus Google Antigravity

Google’s new middle model in the Gemini 4 family, between Carbon (hardest, long-horizon work) and Hydrogen (fast everyday tasks). It is priced at $3 per million input tokens and $15 per million output, with a 1-million-token context. Google reports 61 on the Artificial Analysis Intelligence Index at its highest thinking level (#3 overall), first place on that firm’s Coding Agent Index (58.6), and 65.2% on Terminal-Bench 4.1. At the default level it claims a 32% lower cost per task than Carbon. The same model scores 12.8% on SWE-sweep in the mini-SWE harness and 39.0% in Google’s Antigravity harness. On safety, Google reports shortcuts on 2.45% of tasks in its coding cheating eval and, with real-time safeguards at the highest thinking level, no successful attacks on Gray Swan’s indirect prompt injection benchmark.

Choosing modelsAI ecosystem

Read at Google DeepMind (opens in a new tab)

Cloudflare: AI

An updated Clairic and Clairic-flash are now open-weight, from the new Workers AI Superintelligence team

Alex Reneau, Michelle Chen, Kevin Flansburg

Key number $0.10 per million input tokens for Clairic-flash; output tokens aren't charged

Cloudflare released a redesigned Clairic, its System One extraction model, as open weights at 27B parameters, along with a 9B Clairic-flash. Fields are now generated autoregressively in parallel slots, each stopping at its own end marker, instead of in a single forward pass, and dedicated heads handle estimates, durations, date calculations and arithmetic. Responses still carry confidence scores, character spans and probabilities, and the models now read PDFs as well as images, up to four pages or images per request. On Cloudflare’s own evals Clairic passed all 1,263 tracked customer assertions, against 1,033 for the preview, with median latency between 358 and 691 ms across its test sets. Pricing is per input token only: $0.24 per million for Clairic and $0.10 for Clairic-flash, with both at $0.10 until December 1.

Inference stackChoosing models Decision models

Read at Cloudflare: AI (opens in a new tab)

SaaStr

5 Interesting Learnings from the Databricks AI Index: Open-Weight Jumps to 31% of Enterprise Tokens, 72% of Companies Now Use Multiple Model Providers, and a 281x Gap Between Median and Power Users

Jason Lemkin

Key number 7.3% → 30.9% open-weight share of enterprise tokens, July to November

Lemkin works through the Databricks AI Index, built from the token traffic of more than 10,000 companies that route their AI through Databricks’ Unity Gateway. Open-weight models went from 7.3% of tokens in July to 30.9% in November, mostly for high-volume work such as classification and extraction, and GLM 5.2 alone accounts for about 40% of open-weight tokens. The share of companies relying on one provider fell from about 64% to 28%. On a coding benchmark built from real merged pull requests, GPT-6 Luna and Claude Opus 5.5 both scored 89.1%, but Luna cost $0.05 a task against $1.62; Gemini 3.8 Flash took 140 turns and $11.97 a task despite its low token price; and one model scored 61.7%, 48.1% or 38.4% depending on the harness. The median user spends $2.49 a week on tokens, and the 99th percentile $699.74.

Markets & safe useChoosing modelsAdvising What agents cost

Read at SaaStr (opens in a new tab)

LangChain

What We’ve Learned Building Proactive Agents

Sydney Runkle

Key number 70%+ of Managed Deep Agents runs now start from a schedule, not a prompt

LangChain describes two background agents it runs internally: a sales agent that watches leads and drafts outreach, and Robin, an engineering agent that drafts and reviews pull requests. They wake from three sources: real-world events, schedules set by developers, and schedules the agent sets itself, and over 70% of runs on its managed platform now start from schedules. Memory comes from seeded history, feedback on drafts, and outcomes after the agent acts. Keeping it useful means solving salience, expiry and duplication, often with a scheduled “dreaming” job that reads recent traces and updates memory. On UX, the post argues for surfacing only what needs the user, making each check-in quick to review so feedback flows back into memory, and posting in Slack with a one-line headline. Robin uses a decision model to judge when to join a conversation.

Building agents Context and memory Harness and multi-agent design

Read at LangChain (opens in a new tab)

Latent Space

[AINews] GPT-6.2 Astra, the #2 Frontier Model is better and 20%-600% cheaper, but has mixed reviews; 6.2 Sol likely main workhorse

Key number ~15% of the cost per task at which GPT-6.2 Sol ties September's Astra

Latent Space’s daily roundup on OpenAI’s GPT-6.2 Astra and Sol. In Artificial Analysis’ tests, Sol ties September’s GPT-6 Astra at about 15% of the cost per task, and the new Astra scores 62 on the Intelligence Index against 65 for Claude Fable 5.2. API prices per million tokens are $8 input and $40 output for Astra, and $2 and $10 for Sol, both with a 1M context. Independent results include 67.2% for Astra on Mercor’s long engineering benchmark (Fable 5.2: 54.6%) and, on Perplexity’s research-agent test, 0.783 at about $12.54 a task for Astra against 0.746 at about $2.65 for Sol. OpenAI rated Astra High for AI self-improvement, a first for one of its released models. Early testers were split: one called it “essentially unusable” in agent workflows, while others blamed their harnesses.

Choosing modelsAI ecosystem What agents cost

Read at Latent Space (opens in a new tab)

Anthropic Research

Updating our coordinated vulnerability disclosure policy timelines

Key number 90 → 30 days disclosure window for high- and critical-severity vulnerabilities

Anthropic is shortening the disclosure window in its coordinated vulnerability disclosure policy from 90 days to 30 for high- and critical-severity issues and to 45 for the rest, for reports made from January 1, 2027. It will also escalate to an outside coordinator after 15 days without a maintainer response instead of 30, and publish full technical details 30 days after a patch instead of 45. The reasoning is that disclosure norms assumed human-speed fixes and rare rediscovery; models now help write fixes, and Anthropic is seeing more independent rediscovery of the same bugs. Hard-to-patch cases such as critical infrastructure or specification changes get negotiated timelines, the seven-day window for actively exploited bugs is unchanged, and AI findings without human review are excluded.

AdvisingAI ecosystem

Read at Anthropic Research (opens in a new tab)

Google Developers: AI

How Agent Anomaly Detection scores on benchmarks we didn't write

Nakul Javali

Key number 87% of carried-out harmful instructions flagged for review

Google ran its Agent Anomaly Detection service, which reads agent traces after the fact and flags sessions that leave their intended boundaries, against five public agent-security benchmarks without tuning: AgentDojo, InjecAgent, Agent Security Bench, AgentDyn and AgentHarm. An ADK agent on gemini-2.5-flash ran 1,774 scored sessions, after at least 300 benign sessions per agent to learn normal behaviour, and each benchmark’s own code decided whether the harm happened. The detector flagged 87% of the harmful instructions the agent carried out, with false alarms on 2–4% of benign office sessions and 16% on AgentHarm, whose benign tasks use riskier tools. It caught only 4 of 18 sessions where poisoned long-term memory made the agent act, and with about 220 sessions of history it flagged 33% of attacks on AgentDyn, against 99% with 360. It reports rather than blocks, so Google pairs it with pre-execution controls and human approval for high-impact actions.

Building agentsAdvising Agents acting in the world

Read at Google Developers: AI (opens in a new tab)