Daily Episode

Google Releases Gemini 4 Neon with Revolutionary Caching Discount

Google Releases Gemini 4 Neon with Revolutionary Caching Discount
0:000:00

Episode Summary

TOP NEWS HEADLINES Let's start with the big one: Anthropic officially rings the Nasdaq bell today under the ticker A-N-T, following yesterday's news that the IPO priced at 83 dollars a share. And ...

Full Transcript

TOP NEWS HEADLINES

Let's start with the big one: Anthropic officially rings the Nasdaq bell today under the ticker A-N-T, following yesterday's news that the IPO priced at 83 dollars a share.

And in a delicious footnote, SoftBank CEO Masayoshi Son publicly called passing on Anthropic "one of my mistakes" — ouch, especially since SoftBank is already $64.6 billion deep into OpenAI instead.

But Joanna flagged something that should temper the champagne: Anthropic is sitting on $70.7 billion in compute liabilities due within just twelve months, and over $900 billion stretching across 19 years.

That means revenue has to outrun a 45 to 55 percent annual collapse in frontier token prices just to keep the lights on.

Speaking of Google's token pricing, Gemini 4 Neon just dropped — Google's fourth Gemini 4 model in under two months — and Joanna spotted something developers are buzzing about: a 95 percent discount on cached input tokens, just 15 cents per million versus 3 dollars for fresh input.

We'll dig into why that's reshaping how agents get built in today's deep dive.

On the security front, Joanna surfaced a nasty one: a critical, CVSS 9.3 DNS rebinding vulnerability in Jupyter's MCP server that lets attackers hijack backends through nothing more than a browser trick.

And unconfirmed reports she's tracking describe a paper called "BlockJack," which documents a theoretical cross-tenant cache attack letting one user in a shared inference environment potentially peek at another user's cached computations — worth watching, not yet verified.

Meanwhile, Microsoft had a rough week: an amateur attacker slipped a credential-stealer directly into Microsoft's official agent-framework toolkit — the one with 41,000 weekly downloads — using what appears to be a stolen code-signing certificate.

The GAO also dropped a sobering report: researchers got ChatGPT, Claude, Gemini, and Grok to each produce workable attack plans for either explosives or grid sabotage, all in under 12 hours, by bypassing safety guardrails with common jailbreak techniques.

And Joanna flagged an integrity problem brewing in open-weight AI: an audit of 40 model reports found 22 of them copied benchmark scores straight from previous papers instead of rerunning the tests themselves — only one in 40 fully disclosed its training data, token count, and compute.

Comparative trust in this industry is looking shakier by the day.

Last one — Joanna also spotted reporting that Apple is quietly building its own native Mac AI agent for a late 2027 release, while simultaneously rolling out macOS warnings in version 27.3 that discourage users from granting full disk access to third-party agents.

DEEP DIVE ANALYSIS: Google's Gemini 4 Neon and the Caching Gambit Let's spend our deep dive today on Gemini 4 Neon, because this is Google's fourth Gemini 4 release in under two months, and buried in the pricing sheet is a strategic bet that could reshape how the entire industry builds AI agents. **Technical Deep Dive** Gemini 4 Neon is officially a "mid-tier" model, sitting below Google's flagship Carbon in the family lineup.

Neon actually edges past Carbon on the Artificial Analysis Intelligence Index, scoring 61 and trailing only Claude Fable 5.2 and GPT-6.2 Astra on the entire industry leaderboard.

In Google's own internal testing, Neon beat Anthropic's Claude Opus 5.5 and OpenAI's GPT-6.2 Sol on 16 of 19 benchmarks, with its biggest advantages showing up in long-document analysis, office automation, and cybersecurity tasks.

That's a mid-tier model outperforming two frontier competitors' flagship offerings.

One Google researcher, Minkai Xu, posted cryptically the night before launch that "many people are surprised by how fast RSI moves" — RSI meaning recursive self-improvement, where AI systems start meaningfully accelerating their own development cycle.

Whether that's literally what's happening inside Google's training pipeline or just an flex of confidence, something has clearly changed internally.

Google went from being mocked for shipping slowly through most of 2026 to pushing out Argon, Carbon, Hydrogen, and now Neon in rapid succession.

That's a four-model cadence that no competitor is currently matching. **Financial Analysis** Here's where it gets interesting for anyone running an AI budget.

Neon's API pricing is 3 dollars per million input tokens and 15 dollars per million output tokens — a quarter below both Claude Opus 5.5 and Google's own full-price Carbon.

But the real financial story, which Joanna flagged from her monitoring, is the caching discount: cached input tokens cost just 15 cents per million, a 95 percent reduction off the standard rate.

Think about what that does to unit economics for any application that maintains long-running context — a coding agent working through a large codebase, a customer service bot with persistent conversation history, a research assistant chewing through a document library.

If your architecture is smart about reusing cached context instead of re-sending it fresh every turn, your effective cost per interaction can drop by an order of magnitude.

That's not a marginal pricing tweak; that's Google trying to make stateful, context-heavy agent design the economically rational default, rather than the stateless, pay-per-call pattern that's dominated API usage until now. **Market Disruption** This pricing structure is a direct shot at how Anthropic and OpenAI have built their businesses.

Remember, Anthropic just went public carrying $70.7 billion in near-term compute liabilities, and its entire IPO thesis depends on revenue growth outpacing the 45 to 55 percent annual decline in frontier token prices that Joanna's reporting highlighted.

If Google can simultaneously undercut on list price and create a cheaper-still caching tier, it tightens the vice on Anthropic's margins precisely when the company can least afford it.

It also puts pressure on how developers choose their default model.

The ski-rental problem referenced in TLDR's engineering coverage today — deciding when to switch models mid-session — becomes a lot less painful when Google's caching economics make staying on Neon for long sessions the cheap choice by default.

Meanwhile, OpenAI is leaning into custom silicon with its Jalapeño chip, reportedly cutting circuit area by 10 to 13 percent using its own models to optimize chip design — a hardware-side bet rather than a pricing-side one.

Three different strategies, three different bets on where the margin war actually gets won. **Cultural & Social Impact** There's a broader cultural shift tucked inside this story too.

As AI interactions get cheaper and more persistent, the nature of how people and businesses interact with these systems changes.

Cached, long-running context means your AI assistant genuinely "remembers" your project, your preferences, your prior conversations, without you paying a premium to maintain that memory.

That's part of why Cambridge Dictionary just named "human" its word of the year — not because the word is new, but because compounds like "human-made" and "human-written" are suddenly meaningful distinctions worth paying for.

As AI becomes cheaper, more persistent, and more deeply embedded in daily workflows — the Rundown's own Nate Grahek described using Claude to navigate a hardware store's inventory and place an order entirely on its own — the premium shifts toward proof of human involvement rather than proof of AI capability.

Cheap, persistent AI is becoming invisible infrastructure, and invisible infrastructure changes what people value enough to call out explicitly. **Executive Action Plan** So what should business leaders actually do with this?

First, if you're running any agent-based workflow with long or repeated context — document review, customer support, coding assistance — audit whether your current provider's caching discount matches or beats Neon's 95 percent reduction, and if not, start benchmarking Neon directly against your current frontier model spend.

Second, don't just chase the headline price; test actual task completion quality against your own benchmarks, because today's audit finding that over half of open-weight model reports recycle old benchmark numbers should make every CTO skeptical of marketing claims, Google's included.

Third, start designing your agent architecture around persistent context rather than stateless calls now, because the economics are clearly moving that direction across the industry, and retrofitting a stateless system later will cost you more engineering time than building for caching from the start.

Never Miss an Episode

Subscribe on your favorite podcast platform to get daily AI news and weekly strategic analysis.