Signal SentrySearchUpdated Nov 19, 05:30 UTCAM Drop
Drop #095 · Thu, Nov 19, 2026 · Morning

Neon lands, Airbnb's eval surprise, LiteLLM exploited

10-second version
  1. Google's Gemini 4 Neon lands at $3 in, $15 out per 1M tokens, with Google claiming lower cost per task than Carbon.
  2. Airbnb says a model that won its evals cost several times more per user in production; CISA flags an exploited LiteLLM flaw.
  3. Anthropic raised $81.3 billion in its IPO, CNBC says; two preprints find more thinking or more agents didn't buy accuracy.
Our read

What changed

Every signal →

Our reading of this Drop's sources: what to rethink, and one thing worth checking this week. Not a prediction.

  1. New: eval cost per task can miss how people actually use a model

    If you assumed
    that a model's cost per task in your evals predicts what it will cost in production
    Now test
    whether spend per user moved after your last model switch, and how much of each request is re-read cached context
    This week
    Put eval cost per task next to production spend per user for your current default model and any model you plan to adopt

    Based on:Airbnb: a coding model that won its evals cost several times more per user in productionGoogle ships Gemini 4 Neon at $3 in, $15 out per 1M tokens, pitched on cost per task2 primary sources

  2. Extra thinking and extra agents are not reliably buying accuracy

    If you assumed
    that a higher reasoning effort or a team of agents makes an agent's answers better
    Now test
    whether your agent actually scores higher at high effort, or with several agents, than one run given the same budget
    This week
    Rerun one key eval at low and high effort, and with one agent versus several, matched on samples or spend

    Based on:Preprint: more thinking made shopping agents more confident, not more often rightPreprint: multi-agent LLMs matched or trailed equal-budget baselines in 74% of testsAirbnb: a coding model that won its evals cost several times more per user in production3 primary sources

The Drop in 60

This Drop in 52 seconds

Tap to play. It never plays by itself, and the captions are in the picture.

Transcript

Gemini 4 Neon costs three dollars in, fifteen out per million tokens. Google claims 32 percent lower cost per task than Carbon. Airbnb says a model that won its evals cost several times more per user in production, driven by cached context. CISA flagged an exploited LiteLLM flaw that could let attackers run code remotely. Anthropic's IPO raised 81.3 billion dollars, the largest on record, CNBC says. Palo Alto Networks agreed to buy Applied Compute, which runs custom open-weight models, for 4.5 billion. Rerun evals on your own tasks, and track spend per user after every model switch.

Made with AI from this Drop's sourced items. The facts come from the sources linked on each item. Script: Claude · Voice: OpenAI gpt-4o-mini-tts · Visuals: motion graphics

452models priced on OpenRouter
+597★stars today on the top repo
4,154open roles at AI labs
9 ddays to the next AI-law deadline
Ranked by impact

Top 5 that matter

1
Models

Google ships Gemini 4 Neon at $3 in, $15 out per 1M tokens, pitched on cost per task

Google DeepMind says Neon has a 1M-token context, is rolling out in the Gemini API, Antigravity and the Gemini app, and costs 32% less per task than Carbon at its default level. Artificial Analysis scores it 61.4.

Why it matters: A mid-priced model with frontier-level vendor claims is a reason to rerun your own evals; Google's benchmark and cost figures are its own, so check them on your tasks first.

2
Agents

Airbnb: a coding model that won its evals cost several times more per user in production

Airbnb says the model beat its previous best by over three points at similar eval cost, but in production users sent about three times the requests at about three times the cost each. A later model cut that spend to about a seventh.

Why it matters: Eval cost per task can miss how people use a model; Airbnb found cached context, not output, drove the bill. Track spend per user after each model switch.

3
Security

CISA adds LiteLLM code-execution flaw to its actively exploited list

CISA added CVE-2026-40217 in BerriAI's LiteLLM to its list of actively exploited vulnerabilities. Its entry says remote attackers could run arbitrary code through bytecode rewriting at the /guardrails/test_custom_code URI.

Why it matters: If your team routes model calls through LiteLLM, find every instance, check whether that guardrail test endpoint is reachable, and update it.

Confirmed Way in: a remote request to LiteLLM's guardrail test endpoint that can run attacker code

Controls that would have helped:

4
BusinessFollowing since Nov 15, 2026in 8 Drops

Anthropic prices IPO at $83 a share and raises $81.3 billion, a record, CNBC says

Anthropic says it sold 980,000,000 shares and expects Nasdaq trading as ANT on November 19. CNBC reports $33.24 billion in revenue for the nine months to September 30 and run-rate revenue near $81 billion at end of October.

Why it matters: A key model vendor now reports publicly each quarter, giving buyers hard numbers on its revenue, losses and stated risks to weigh in vendor reviews.

5
Business

Palo Alto Networks to buy open-weight model startup Applied Compute for $4.5 billion

SecurityWeek says the all-stock deal brings a platform for adapting open-weight models to company data and running them, to be offered in public clouds and sovereign setups. Quarterly revenue rose 39% to $3.44 billion.

Why it matters: Security vendors are moving into hosting the models themselves; if you plan open-weight deployments, add your security suppliers to the shortlist and check lock-in.

Agents

Agents at work

See all Agents →
Agents

Preprint: more thinking made shopping agents more confident, not more often right

Why it matters: Don't treat an agent's stated confidence or a higher effort setting as grounds to let it buy unchecked; keep a person's confirmation step for purchases.

More detail: Preprint: more thinking made shopping agents more confident, not more often right

Researchers tested ten frontier models on three product-choice benchmarks. Raising thinking effort did not meaningfully improve picks but lifted stated confidence; most models were poorly calibrated, and many erred on the same tasks.

Agents

Preprint: multi-agent LLMs matched or trailed equal-budget baselines in 74% of tests

Why it matters: Before adding agents, compare against one model given the same samples or budget; extra agents reading identical evidence may add cost without accuracy.

More detail: Preprint: multi-agent LLMs matched or trailed equal-budget baselines in 74% of tests

A synthesis of nine studies and 516 configurations finds multi-agent accuracy averaged -0.042 against baselines matched on samples or budget. The author argues gains need split evidence, outside checks or varied models.

On the Frontier watch →
Security

Attacks and flaws

See all Security →
Security

Nadella says faster AI-found patches are destabilizing old systems after Outlook outage

Why it matters: Faster patching is now a risk of its own. Check that urgent patches still get a staged rollout and that one failure cannot take down every region.

More detail: Nadella says faster AI-found patches are destabilizing old systems after Outlook outage

Nadella told CNBC a latent bug hit by one traffic pattern caused last week's Outlook outage and Microsoft should have tested more. At DealBook he said much critical infrastructure firmware is unpatchable, TechCrunch reports.

Security

Preprint: a federated trust layer for cross-company agents lifted containment to 84%

Why it matters: Agents calling other companies' agents need identity checks outside the model. A judge-scored simulation is early evidence, not proof, but the design is open source to test.

More detail: Preprint: a federated trust layer for cross-company agents lifted containment to 84%

Researchers report ATI checks agent certificates, approved model, tool and prompt setups, and three policy rules before each request. In a simulated eight-org federation, containment rose from 43% to 84%, adding about 174 microseconds.

On the Frontier watch →
Business

Money and deals

See all Business →
Business

Intel signs first outside customer for its 18A chip process, unnamed, CNBC reports

Why it matters: Another advanced foundry option matters for AI chip supply; until the customer is named, treat it as a signal, not capacity you can plan on.

More detail: Intel signs first outside customer for its 18A chip process, unnamed, CNBC reports

CEO Lip-Bu Tan said the deal is signed but did not name the client; Intel's technology chief said it is for the 18A-P node. Tan also named Microsoft the first public customer for its EMIB packaging.

Jobs

Jobs pulse

See all Jobs →
Jobs

Nadella says Microsoft's 60,000 buyout offers were about reorganizing work around AI

Why it matters: If this model spreads, roles shift toward directing agents; decide how your teams will be organized around agent work and who gets retrained.

More detail: Nadella says Microsoft's 60,000 buyout offers were about reorganizing work around AI

TechCrunch reports Nadella said the October buyouts let staff choose before jobs shift to directing AI, citing an 11-developer Cowork team that builds a software factory rather than the product.

Repos

Climbing fastest

See all Repos →
ReposFollowing since Nov 17, 2026in 3 Drops

rust-anything ports apps to Rust piece by piece, checked against originals; 3564 stars

Why it matters: Checking every step against recorded behavior is a useful pattern for any large rewrite; try it on a small, well-tested tool before anything critical.

More detail: rust-anything ports apps to Rust piece by piece, checked against originals; 3564 stars

The MIT-licensed repo, created November 16, ports CLIs, HTTP servers and libraries from TypeScript, Python, C, Go, Ruby and Java to Rust, checking recorded behavior at each step. It gained 597 stars since our last check.

Today's Talk

What people are arguing about

See all Today's Talk →
Today's Talk

X posts debate a reported US plan to make firms disclose Chinese AI model use

Why it matters: Nothing official is published here. If you run Chinese open-weight models, know where they sit in your stack in case disclosure becomes a requirement.

More detail: X posts debate a reported US plan to make firms disclose Chinese AI model use

@unusual_whales posted, citing Reuters, that the White House and Treasury are discussing website disclosures for firms using Chinese AI models. @Jason called it a name-and-shame play; @bindureddy called it ridiculous.

Play the NewsAgent on a LeashAbout 2 minutes, built from this Drop's items.