Koko Daily · Wednesday, September 23, 2026

Koko: It's Wednesday, September twenty-third, I'm Koko, with Max and Sam — let's blitz it.

Max: Claude Opus five point five dropped, then ninety minutes later GPT-6 Sol and Luna landed — both slashing prices forty to fifty percent.

Sam: Meanwhile AI agents got caught actively recommending malware — Gemini and ChatGPT both pointed users to the same poisoned GitHub repo.

Koko: Okta's rallying AWS, Google, and Salesforce into a Blueprint Alliance to build kill switches for rogue agents.

Max: Accenture and Anthropic are putting two billion dollars combined into an independent AI evaluation venture.

Sam: PwC's leadership is publicly pushing a two-in-the-box model — pair an engineer with a business architect on every engagement.

Koko: McKinsey data says enterprises are blowing past budget on coding agents with little to show for it.

Max: Nvidia launched a DSX Ready certification program to standardize AI data center builds.

Sam: And Qualcomm shipped a new AI-focused Android chip right as a memory shortage squeezes device margins. Let's dig in.

Koko: Okay, this is the story of the day. Anthropic ships Opus five point five, and ninety minutes — ninety minutes — later OpenAI drops GPT-6 Sol and Luna.

Max: That's not a coincidence, that's a stakeout. Somebody at OpenAI had a finger on the trigger waiting for Anthropic's announcement to clear.

Sam: [chuckles] It's like two restaurants opening across the street and both slashing the menu the same night.

Koko: And the cuts are real — forty to fifty percent off API pricing, both sides. This isn't a promo, it's a repricing of the entire category.

Max: Right, and notice what they're NOT bragging about. It's not benchmarks anymore. It's tokens per dollar. Opus is forty percent cheaper and thirty percent faster — that's an efficiency pitch, not a genius pitch.

Sam: Which tells you something about where the ceiling actually is. If capability gains were still the story, you'd lead with capability.

Koko: So is this good news for buyers, or a warning sign that margins are getting squeezed somewhere upstream?

Max: Both. OpenAI's burning something like two hundred seventy-eight billion dollars through twenty-thirty by some projections. Prices don't fall fifty percent because costs fell fifty percent — they fall because the war demands it.

Sam: Somebody's eating that margin. Could be Nvidia's discounts trickling down, could be investors just footing the bill a while longer.</br>

Koko: Don't literally say that br tag, Sam.

Sam: [laughs] Noted. Point stands — this is a subsidized war, not an efficient one, yet.

Max: For enterprises, though, it's a genuine gift. Multi-model failover just got a lot cheaper to run for real, not just in a slide deck.

Koko: This next one is genuinely unsettling. Gemini and ChatGPT — independently — pointed users to the exact same malicious GitHub repo.

Max: Fourteen million downloads on that repo. That's not a fringe result, that's a top search suggestion from two different labs' flagship assistants.

Sam: How does that even happen? Did they both just... trust the same fake five-star reviews?

Koko: Basically, yeah — the repo gamed the signals both models use to judge trustworthiness. Stars, forks, activity. It looked legitimate to a model reading surface signals.

Max: Which is the uncomfortable part. These agents aren't being hacked exactly — they're doing exactly what they're built to do, and the thing they're built to do just got weaponized.

Sam: So it's less a bug and more... a new SEO. Malware-as-a-service optimizing for AI recommendation instead of Google ranking.

Koko: That's a great and horrifying way to put it, Sam.[chuckles]

Max: And it lands the same week Meta patched a zero-day letting attackers straight-up hijack its Muse agent. The pattern's not subtle anymore.

Sam: Which is exactly why Okta's move matters — get AWS, Google, Salesforce, all pointed at shared agent governance before every vendor builds its own incompatible kill switch.

Koko: Do you actually think a Blueprint Alliance moves fast enough to matter, or is this just four companies posing for a press release?

Max: Ask me in three months. Standards bodies move slow, attackers don't wait for consensus.

Sam: Which means for now the actual defense is embarrassingly manual — don't trust an agent's link just because the agent trusts it.

Koko: Let's talk about a genuinely new move — Accenture and Anthropic committing a billion dollars each, five years, into an independent AI evaluation venture through Accenture's Faculty unit.

Max: Two billion dollars to build red-teaming and safeguard testing as an actual business line. Not a pilot, not a lab — a revenue category.

Sam: This is smart positioning. Everyone's selling AI deployment. Nobody's selling AI assurance at scale yet.

Koko: Right, and it lands the same week PwC's Paul Griggs is out there publicly pushing this two-in-the-box model — pairing an engineer with a business architect on engagements.

Max: Which is basically an admission that the old billing model — bodies on a project, hours on an invoice — doesn't survive an AI-delivered world.

Sam: If the engineer plus architect pair can do what used to take a team of eight, what happens to the eight?

Koko: That's the question nobody on stage wants to answer directly.[pause]

Max: PwC Switzerland already gave you the preview — quietly halving bonuses, blaming AI cost pressure. Margin compression doesn't wait for the strategy to be finished.

Sam: So Accenture's evaluation venture is interesting on two levels. It's a new fee line, sure. But it's also a hedge — if you're the firm that certifies everyone else's AI is safe, you're indispensable no matter how the labor math shakes out.

Koko: Selling shovels and now selling insurance on the shovels.[laughs]

Max: And PE firms evaluating consulting targets should be pricing that distinction into every deal right now — who's building the assurance layer versus who's just riding headcount.

Koko: Nvidia's new DSX Ready program certifies data centers to a standard build spec — basically a seal of approval for AI infrastructure.

Max: Qualcomm shipped an AI-focused Android chip right as a memory shortage is jacking up smartphone bill-of-materials costs — margin squeeze incoming for device makers.

Sam: IREN's latest earnings call showed nineteen billion dollars raised against just four billion in contracted revenue — a capital structure heavy on debt-like instruments worth watching.

Koko: And Alibaba shares popped again on fresh AI chip and data center plans — infrastructure disclosure is officially a stock-moving event now.

Koko: Okay, quick recap. Anthropic and OpenAI both slashed prices forty to fifty percent within ninety minutes of each other — the frontier race is now a margin war.

Max: AI agents got caught recommending malware to real users, fourteen million downloads deep, and Okta's rallying the industry toward shared kill switches.

Sam: And Accenture and Anthropic are putting two billion dollars behind AI evaluation as its own business, while PwC restructures delivery teams around AI leverage.

Koko: Now, if you're sitting in professional services or private equity today, here's the lens. Every consulting and SI relationship you have is about to be repriced — literally, because model costs just dropped, and structurally, because two-in-the-box delivery models mean fewer billable bodies per engagement.

Max: So the action item is concrete: before your next contract renewal, ask your services vendor to show you their cost savings from these price cuts and demand it show up in your rate, not just their margin.

Sam: And if you're doing diligence on a professional-services target, price in the assurance layer now — firms building AI evaluation and governance capability, like Accenture just did, hold pricing power differently than firms just reselling headcount with an AI logo on it.

Koko: Here's a question worth sitting with: when delivery labor keeps compressing, does professional-services equity re-rate on revenue per partner, or on something entirely new — evaluation contracts, safety certifications, assurance revenue?

Max: Watch for this within the month — at least one more Big Four or major SI firm announcing its own AI evaluation or assurance practice, copying Accenture's playbook.

Sam: And watch this quarter — whether Okta's Blueprint Alliance actually ships interoperable agent kill switches, or whether AWS, Google, and Salesforce quietly go their own competing ways instead.

Koko: Longer horizon — within a year, expect professional-services fee structures to visibly shift toward fixed-bid and outcome-based pricing as the default, not the exception, as this price war and delivery restructuring both compound.

Max: Lots to chew on.

Koko: That's the show. For the full brief, every story, and insights built for your exact role, go to koko knows dot A I. [excited] See you tomorrow.