Every week I go through what actually landed in my two inboxes — newsletters, daily briefs, funding notes, the occasional conference write-up that turns out to be sharper than the analysis around it. Here’s what stood out between 6 and 13 October, and why I think it matters if you are running agentic AI in production.
The shape of it: in a third of head-to-head tests, the model with the lower list price cost up to 28 times more to finish the same work. Google is giving its enterprise agent a workplace identity and an email address, and has set a November deadline for classifying which of your OAuth clients is an agent. A room full of NeoCloud operators, chipmakers and lenders spent a day converging on one unit of measurement: cost per outcome. Last week the story was that the model layer is commoditising. This week it became clear that the commodity is priced in a currency nobody is actually measuring.
The cheaper model billed 28 times more
Chamath Palihapitiya’s 11 October note carries the number I have not been able to put down. He cites a study by researchers at Stanford, UC Berkeley, Carnegie Mellon and Microsoft Research that ran eight reasoning models — models that work through a problem step by step before answering — across twelve kinds of task. The finding: in 32% of 336 head-to-head comparisons, the model with the lower list price cost up to 28 times more in total. At the study’s 1 May prices, Google’s Gemini 3 Flash listed 80% cheaper than OpenAI’s GPT-5.4, and yet running every test cost $705 on Gemini against $509 on GPT-5.4.
The mechanism is unglamorous and completely decisive. Reasoning steps — thinking tokens — are billed at the higher output rate, and on single-question tasks where the cheaper-listed model cost more overall, thinking tokens explained more than 95% of the gap. On one test set, Gemini 3 Flash produced more than 71 million thinking tokens against GPT-5.4’s 1.8 million. The buyer sees a price per token. The model decides how many tokens the job takes, inside an effort setting the buyer picks. Chamath sets this against Anthropic releasing Claude Haiku 5.5 on 7 October at $0.10 per million input tokens and $0.50 per million output — the same list price as OpenAI’s GPT-6 Luna for prompts under 100,000 tokens, and in the same range as frontier Chinese models at similar performance. Haiku 5.5 is Anthropic’s first small model with an adjustable effort setting, and Anthropic says its updated tokenizer uses slightly more tokens per task. The honest caveats are in the note: the study ran each model at a single reasoning setting, did not weigh cost against answer quality, and predates this week’s releases.
For two years the industry has quoted price per million tokens as though it were a price. It is not a price — it is a rate card for a quantity you do not control. The vendor sets the rate; the model sets the quantity; you get the invoice. And notice where the lever sits: the effort setting, the prompt, the stopping condition and the retry policy are all things you own, and they move the bill far more than the rate card does. I have watched teams spend six weeks negotiating a 15% discount on a model they then let reason for ninety seconds on a question that needed four. If you cannot produce cost per completed task for your top three workflows this week, you are not managing AI spend — you are receiving it.
Google gives its agent an email address — and you a November deadline
TechCrunch’s 9 October morning brief led with Google turning Gemini into an agent for business: it can plan, execute tasks and work across business apps and systems, delegate work to subagents, use multiple AI models, and gets its own workplace identity, complete with an email address. That last clause is the one to read twice.
Because the administrative half of it arrived in my inbox on 12 October, from Google Workspace directly. Google is rolling out Workspace Model Context Protocol (MCP) servers exposing Gmail, Drive, Calendar and other Workspace apps to external AI agents, and ahead of that it has introduced an “agentic” setting (is_agentic) on Google OAuth clients so that Google and Workspace administrators can tell agent traffic apart from ordinary app traffic. The timeline is tight and it is not negotiable. From 12 October developers can mark OAuth clients as agentic. From 11 November, Google will reject requests to Workspace MCP servers from any client not marked agentic — the error will say “the client isn’t marked as an AI agent”. From 12 November the setting becomes permanent: to change it you must create a new OAuth Client ID and update your app. Google’s own guidance is to use a separate OAuth client where an app handles both agent and non-agent traffic, because agentic clients get separate security controls applied.
Put the two halves together and something structural has happened quietly. An agent is no longer a feature of an application. It is an identity with a mailbox, a classification on a credential, and an irreversible flag with a date on it. I have argued for months that agent risk surfaces as over-broad identity and access rather than as model error; this is the week a hyperscaler agreed in writing and attached a deadline. Somebody in your organisation now has to open a console before 11 November and answer, client by client, “is this thing an agent?” — and after 12 November a wrong answer costs a migration rather than a toggle. If you do not already hold an agent inventory with owners, credentials and reachable systems, this deadline is going to find out. That register is not compliance paperwork; it is a load-bearing part of the harness, and this is the first month it has a due date.
The NeoCloud room settles on one unit: cost per outcome
IgniteGTM’s 11 October write-up of NeoCloud Summit 2026, held at Hotel Kabuki in San Francisco on 8 October, is the best-reported thing in my inbox this week. Ignite CEO Bill Barry opened by describing NeoClouds as the industry’s new “center of gravity”, and named the mismatch everyone in the room was living with: power and data centres take years to build, while agentic applications change in weeks.
Two things in the report cut against the usual conference consensus. First, Steven Hou, Head of Research at Silicon Data, pushed back on the hype directly: “This idea that everyone is deploying a swarm of AI agents. That’s not happening yet.” He argued that capital efficiency, utilisation and tenant quality now decide what gets financed. Second, AMD’s Nicola Tan gave the sharpest architectural test I have read all year: “The biggest risk is not choosing the wrong technology, but building something that cannot adapt” — and her test for whether you actually have an open stack, “if swapping one vendor forces you to replace everything, you don’t have a layer, you have a bundle.” On the tokenomics panel, WEKA Chief AI Officer Val Bercovici said what the pricing study above proves: “We need to get to cost per task, cost per outcome.” Hosted.ai’s Nilesh Shah gave inference buyers the same instruction — “Look at the cost of the outcome.” Elsewhere: Remanie.ai CEO Tiffany Wilson on capacity promises, “Commit to the capacity that you can deliver, not commit to capacity that you hope to deliver”; and Acasia’s Matt Renner to the finance panel, “You can’t finance the GPUs if you can’t secure power.” One number for how much this subject matters right now: across the 1,000-plus events of SF Tech Week, community builders put the average show rate at about 15%. NeoCloud Summit saw 93%.
“You don’t have a layer, you have a bundle” is the cleanest statement of the harness argument I have seen from someone selling silicon, and I intend to steal it. It is also a test you can run this afternoon: pick your largest agentic workflow and write down what breaks if you change the model vendor. If the honest answer is prompts, tool schemas, evaluation sets, guardrails, logging and the integration layer, you have a bundle, and your negotiating position is theatre. The other line worth sitting with is Hou’s. The people financing this build-out do not believe the swarm has arrived — and they are closer to the invoices than the keynote stages are. That is a reason to make your second and third agents cost a fraction of your first, which only happens if the layer underneath them is yours.
The most useful thing in my inbox was a sales coach explaining how to hire an agent
Matt Verlaque’s SaaS Academy deep dive on 10 October opens with a sentence I wish I had written: “I’d bet most of the AI agents running inside businesses right now would’ve been fired in their first week if they were people.” Somebody typed a vague request into a chat box, never explained what the job was or where anything lives, the output came back mediocre, and everyone blamed the model. He makes the point that if you did that to a new hire, you would know exactly whose fault it was.
His five steps are a harness specification in HR language, which is why they travel so well. Write the job description — the system prompt — with an outcome carrying a number and a deadline, a name and a role, and the rules one line each: how it reports, what it does when finished, where it gets information, and when it must stop and ask. His test is to read it back as if you were the new hire: if you would have to ask a question, the agent will guess instead. Onboard it with a “company brain”, one place holding a copy of your data, and define your terms, because if you do not spell out what “active customer” means the agent invents a definition. Train it with a skill file, written by the camcorder method: record yourself doing the task, narrate every decision and why, transcribe, then cut everything but the process. Coach it every couple of weeks with evals: pull recent work plus the skill file and data, write down what you wish it had produced, and ask your best model what to change to close the gap. And make it work in public — an audit log recording every decision and why, a report file per run, and a notification to a human on every run, so nothing happens out of sight.
I have read a great many enterprise AI governance frameworks this year and none is as actionable as this, which arrived in a newsletter aimed at founders. It works because it refuses the category error at the heart of most failed deployments: people buy a model and expect a colleague. A model is a capability. A colleague is a capability plus a job description, a map of where the truth lives, written-down craft, a feedback loop, and a record of what they did. Four of those five are things you write, not things you buy — and all four survive a change of vendor. The one line I would add from our own deployments: the audit log is not for the auditor, it is for the next eval. When an agent gets something wrong, a log of the decision and the reason lets you trace it to the exact line of the skill file that caused it. Without it you are tuning a prompt by vibes.
OpenAI published 700 maths manuscripts. The interesting number is 42%.
Also from Chamath’s 11 October note: on 6 October OpenAI revealed over 700 mathematical manuscripts produced by an unreleased internal model, and says that about 42% of the headline results have proofs formalised in Lean, a tool that mechanically verifies mathematical reasoning. Computer scientist Scott Aaronson observed that cryptography is conspicuously missing from the release, and wrote that his sources say AI companies have begun discreetly investigating whether their latest models can break important cryptographic protocols.
The debate that followed is a study in proportionate response. Ethereum Foundation researcher Justin Drake urged a controlled move into addresses that keep public keys hidden until use; his worst case was a break within months, and he warned against rushing because panic causes more problems than it solves. Vitalik Buterin likewise recommended keeping funds in addresses not yet used for a transaction. Coinbase’s head of cryptography, Yehuda Lindell, countered that there is currently no evidence of anything major happening and no need to rush, recommending instead that blockchains add hybrid schemes alongside their post-quantum migrations.
Strip out the cryptocurrency and what is left is the most important enterprise lesson of the week. 42% formalised means 58% not formalised. OpenAI shipped the distinction in the announcement rather than burying it, and that is the right instinct — but notice what made the distinction possible at all. Lean is an evaluation harness. It is a mechanical, reproducible, vendor-independent check that sits outside the model and says yes or no, and its existence is the only reason anyone can quote a percentage rather than an impression. Every enterprise deploying agents needs its own Lean: a mechanical check, owned by you, that separates the output you can act on from the output that still needs a human. Most organisations have the opposite — a model they trust and a review step they hope somebody is doing. And Lindell’s response is the governance template: add a layer that does not depend on the threat being real, and do not rebuild your estate on a rumour.
Three frontier releases, two quiet deprecations, one leaderboard measuring lies
The churn underneath all of this got louder. TechCrunch reported on 12 October that Meta released Muse Wonder, its first frontier model; where Muse Spark aimed to be a lightweight model performing at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and better results on complex tasks such as coding. In the same brief, Microsoft president Brad Smith called for urgent AI regulation and said the UK is best placed to drive it, with the observation that milk and washing machines currently have more oversight than AI companies.
Meanwhile two deprecation notices landed in my own mailbox on 10 October: Gemini 3.5 Flash and Gemini 3.7 Flash have both been deprecated and auto-redirected, headed “No action required”. And TechCrunch reported on 9 October that Arena, the company behind the LMArena leaderboard, has raised $200 million led by Lightspeed and Khosla, nearly doubling its valuation to $3.1 billion in ten months — and is now measuring AI models on alignment issues such as lying. The same brief carried the report that OpenAI’s revenue is reportedly $20 billion less than previously projected, against earlier reporting of roughly $70 billion annualised.
“No action required” is doing a lot of work in those deprecation emails. An auto-redirect is a silent change to the component that decides how many thinking tokens your workflow spends and how often it gets the answer right — which, per the first story in this edition, is the component that sets your bill. If a model can be swapped underneath you with a cheerful notice and no action on your part, then your evaluation suite is the only thing standing between that swap and your customers. The genuinely encouraging item is the least glamorous: a leaderboard company raising $200 million to measure whether models lie. Evaluation is becoming an industry, and that is the clearest signal yet that the durable value is moving to the measurement layer rather than the model.
Quick hits
- $1.8 billion to build a virtual cell. Chamath reports that on 7 October Biohub — the nonprofit founded by Mark Zuckerberg and Priscilla Chan — expanded its effort to build an AI simulation of a living cell that predicts responses to drugs and genetic changes. On top of $500 million committed in April, it announced over $800 million more, including more than $500 million from the Department of Energy over five years and a combined $300 million from Meta, Google DeepMind and Isomorphic Labs; the NIH is contributing datasets built with more than $500 million of past federal spending. Commercial partners get one year of exclusive access to the data they develop, and the first dataset is expected in about a year.
- The AI build-out meets private credit. Chamath’s 9 October deep dive asks whether the AI build-out is creating a private credit bubble — how private credit works, how it grew, and what the build-out changes. Read alongside the NeoCloud finance panel’s “you can’t finance the GPUs if you can’t secure power”, it is the clearest view this week of where the capital actually comes from.
- $3.7 billion in one week, 87% of it megadeals. IgniteGTM reports Bay Area startups raised $3.7 billion in October’s first full week, $3.2 billion of it in megadeals dominated by AI infrastructure and applied AI. The standout: TypeSafe AI closed an $870 million Series A for Jev, a model that makes fast structured decisions rather than generating conversational text — less than a month after a $40 million seed, having picked up more than 25% of the Fortune 500 as customers. Oxide Computer raised $445 million for on-premises cloud-grade servers.
- Safety culture under strain. TechCrunch reported on 9 October that three fired OpenAI safety researchers dispute allegations of mishandling sensitive information, warning in an open letter that their dismissals are creating a chilling effect on the company’s AI safety culture.
- From our own desk: Veehive Mind reported a deliberately quiet week — nothing published between 5 and 11 October, with one campaign sitting in the approval queue waiting on a human decision. I am going to defend that. An agent that holds and says it is holding is behaving correctly; an agent that publishes to avoid an empty report is the failure mode nobody writes a policy for.
The theme of the week
For nine weeks the argument has been the same: rent the model and the compute, own the context, the harness, the evaluations, the controls and the record of truth. This week sharpened it into something more specific. The thing you own is not a defensive asset. It is the only place the actual numbers are decided.
Look at what each story is really about. The pricing study says the vendor sets the rate and your harness sets the quantity, so your bill is mostly your own design. Google’s agentic OAuth flag makes your agent inventory an administrative fact with a date attached. The NeoCloud room says the unit of account is cost per outcome, computable only against your own tasks. Verlaque’s five steps say four of the five things separating a model from a colleague are written by you. Lean says a percentage is only possible when a mechanical check sits outside the model. And two silent Gemini deprecations say the model can change underneath you while you sleep.
Six independent sources, from a venture newsletter to a Google admin email, arrived at the same structural claim in one week. The model is the part you rent, the part that churns, and the part whose price tells you almost nothing. The harness is the part that decides what you spend, what you can prove and what survives the next deprecation notice. Rent the model. Own the harness. This week, own the meter.
What this means if you’re deploying agentic AI
- Replace price-per-token with cost-per-completed-task. Instrument your top three agentic workflows to report total cost per successful outcome, broken out by input, output and thinking tokens. Until you have that number, a cheaper rate card is not evidence of a cheaper system.
- Treat the effort setting as a budget control. Thinking tokens explained more than 95% of the cost gap in the study. Set explicit effort levels, step limits and stopping conditions per workflow, and review them the way you review cloud spend.
- Classify your OAuth clients before 11 November. Identify every client used by an AI agent, mark it agentic, and split any client carrying both agent and non-agent traffic. After 12 November the flag is permanent and changing it means a new Client ID and an app update. While you are in there, record owner, purpose, credentials, reachable systems and expiry for each one — Google has now issued agents an email address, and your access review will ask who that mailbox belongs to.
- Apply the bundle test to your largest workflow. Write down what breaks if you change model vendor tomorrow. If the list is long and lives in one provider’s API, you have a bundle, not a layer — and that list is your migration plan.
- Write the job description, the skill file and the references folder. Outcome with a number and a deadline, where the truth lives, your definitions of key terms, the process captured by narrating it once, and real examples of good and bad work. Read it back as a new hire: every unanswered question is a place the agent will guess.
- Build your own Lean. For each agent output class, define a mechanical check that runs outside the model and returns a verdict. Report the share of outputs that pass it. That share, not a benchmark, is your production number.
- Make every run leave a trace. An audit log of decisions with reasons, a report file per run, and a notification to a named human. The log is what lets you trace a bad output back to the line that caused it at the next eval.
Frequently asked questions
Why can a cheaper AI model end up costing more to run?
Because the list price is a rate, not a bill. Chamath Palihapitiya's 11 October 2026 note cites a study by researchers at Stanford, UC Berkeley, Carnegie Mellon and Microsoft Research that ran eight reasoning models across twelve kinds of task. In 32% of 336 head-to-head comparisons, the model with the lower list price cost up to 28 times more in total. At the study's 1 May prices, Google's Gemini 3 Flash listed 80% cheaper than OpenAI's GPT-5.4, yet running every test cost $705 on Gemini against $509 on GPT-5.4. The cause is thinking tokens — the reasoning steps a model writes before answering, billed at the higher output rate. On single-question tasks where the cheaper-listed model cost more overall, thinking tokens explained more than 95% of the gap; on one test set Gemini 3 Flash produced more than 71 million thinking tokens against GPT-5.4's 1.8 million. The buyer sets the rate card and the effort setting; the model decides how many tokens the job actually takes. The study ran each model at a single reasoning setting, did not weigh cost against answer quality, and predates releases such as Anthropic's Claude Haiku 5.5, launched on 7 October 2026 at $0.10 per million input tokens and $0.50 per million output.
What is changing for AI agents that access Google Workspace?
Google is rolling out Google Workspace Model Context Protocol (MCP) servers that expose Gmail, Drive, Calendar and other Workspace apps to external AI agents, and has introduced an "agentic" setting (is_agentic) on Google OAuth clients so that Google and Workspace administrators can distinguish AI agent traffic from traditional app traffic. The timeline announced on 12 October 2026 has three dates. From 12 October 2026, developers can mark OAuth clients as agentic in the Google Cloud console. From 11 November 2026, Google will reject requests to Workspace MCP servers from any OAuth client not marked as agentic, with an error stating that the client isn't marked as an AI agent. From 12 November 2026 the setting becomes permanent, so changing it requires creating a new OAuth Client ID and updating the application. Google advises using a separate OAuth client where an application handles both AI agent and non-agent traffic, because agentic clients have separate security controls applied, such as agent access controls. Applications that are not AI agents and only call Google Workspace REST APIs are unaffected. Separately, TechCrunch reported on 9 October 2026 that Google is turning Gemini into an agent for businesses that can plan, execute tasks, work across business apps, delegate to subagents, use multiple models, and receives its own workplace identity complete with an email address.
What did NeoCloud Summit 2026 conclude about pricing agentic AI?
That the GPU-hour is the wrong unit. IgniteGTM's 11 October 2026 report on NeoCloud Summit 2026, held in San Francisco on 8 October, describes a day spent converging on cost per outcome. WEKA Chief AI Officer Val Bercovici said on the tokenomics panel, "We need to get to cost per task, cost per outcome," and Hosted.ai's Nilesh Shah told inference buyers to "look at the cost of the outcome". Ignite CEO Bill Barry framed the structural mismatch: power and data centres take years to build while agentic applications change in weeks. Silicon Data's Steven Hou argued that capital efficiency, utilisation and tenant quality now decide what gets financed, and pushed back on the hype directly: "This idea that everyone is deploying a swarm of AI agents. That's not happening yet." AMD's Nicola Tan offered the test for architectural lock-in: "The biggest risk is not choosing the wrong technology, but building something that cannot adapt," and "if swapping one vendor forces you to replace everything, you don't have a layer, you have a bundle." The finance panel added that capacity must be bankable, with Acasia's Matt Renner noting, "You can't finance the GPUs if you can't secure power."
From Veehive Labs
Nine weeks in, the compressed version holds and gets more concrete: rent the model and the compute; own the context, the harness, the evaluation and the record of truth. What this week added is the meter. The vendor publishes a rate; your harness decides the quantity, the stopping condition, the retry and the check that says whether the work was any good. That is where the money is spent and the proof is produced — and it is the only part of the stack a deprecation notice cannot quietly rewrite. Veehive Labs is a Dubai-based AI innovation lab and custom AI product development company, building model-independent enterprise AI harnesses, custom agents, RAG pipelines and sovereign, private-AI deployments for regulated and operationally complex organisations across the UAE, KSA, GCC and beyond.
If you’re moving from AI strategy to production, start with a focused AI Discovery Sprint — we map your use case, data, systems, agent inventory, permissions model, cost ceilings, risks and delivery architecture before the build begins.
- Read the deep dive: Own the harness, rent the model
- Read the last edition: Memory out-earns the models, consistency becomes the metric
- Explore Veehive Labs enterprise AI services · See our solutions
Build Your Organisation’s AI Capability
Start with a focused AI Discovery Sprint. We will map your use case, organisational knowledge, systems, MCP integrations, security requirements and delivery architecture.
Agentic AI This Week is written by Sathish Jeyakumar, Founder & CEO of Veehive. Bring your preferred model — we make it understand, speak and operate like your organisation.