The Product Desk

The New Buying Motion Isn't Bottom-Up Anymore

A PM at a mid-market SaaS company noticed three of her top twenty target accounts changed ownership fields this week. All three sit under Blackstone portfolio companies. The deals stalled not because a competitor appeared but because the buyer committee was replaced overnight by someone with a portfolio-wide vendor consolidation mandate and a spreadsheet that does not care about her champion's product love.

Seven independent intelligence sources point at the same pattern this week. OpenAI closed a $10B raise from a 19-firm Wall Street consortium explicitly designed to push ChatGPT agents into every mid-market company those firms own. Five days later Anthropic closed its $1.5B JV with Blackstone, Goldman Sachs, and Hellman & Friedman under the same structure. Blackstone alone carries 250+ portfolio companies. The full consortium runs into the thousands.

For three years the labs sold direct through enterprise sales and API contracts. Now they sell once to a PE sponsor and deploy to hundreds. This is Accenture-style distribution at venture speed.

What This Actually Means For Your Product

The motion has two faces that compound against independent vendors:

1. Top-down mandate: When Blackstone owns a company and says 'implement Claude for operations efficiency,' the company implements Claude. The outside sales call arrives after the decision is made.
2. Consultant-bundled deployment: Anthropic is not selling an API anymore. It is selling a transformation program with Claude inside. OpenAI, Anthropic, and Salesforce have each built consulting arms for the same reason: enterprise buyers do not want a model, they want someone to run the workflow change.

The commercial implications diverge sharply by segment. PE-owned accounts move top-down the moment the sponsor writes the mandate into the operating plan. They do not run bottom-up experiments, and a third-party tool is not going to win on developer love. Independent companies still pick on time-to-value and usage depth. One motion does not serve both.

Where You Can Still Win

The PE JV deploys general-purpose AI across operations. It does not deploy domain-specific workflow tools that require proprietary data and context. The 2x2 that matters: on one axis, is the product a layer the foundation model provider will eventually ship as part of a deployment package, or a layer they will keep routing customers toward because it makes their consulting engagements faster? Build in the second cell. The products that survive are consultant-resellable and priced per-outcome. They slot into the JV's deployment playbook rather than competing with it.

The Timeline Is Quarters, Not Years

Goldman Sachs is already cutting Claude access for Hong Kong bankers over contract concerns while simultaneously being part of the $1.5B JV. The internal contradictions tell you the playbook is still forming. The window to position as complementary rather than competitive is the next 6-12 months, before the sponsor's standard vendor checklist and operating-plan template harden around a specific deployment pattern across hundreds of companies.

The Numbers That Should Change Your Next Pricing Meeting

A developer named Theo sat down with GitHub Copilot and ran 15 agentic messages. One of those messages consumed more than 60 million tokens. By the end he had burned $221 of inference tokens on a $40/month subscription. Uber, running the same playbook at scale, disclosed Claude Code costs of $500–$2,000 per engineer per month and burned its entire 2026 AI coding budget in four months. This is not misuse. This is the power-user behavior the product was designed to enable.

The distribution of cost per user in an AI product is not normal. It is a long tail where the top 5% of users generate roughly half the inference spend. A flat subscription is a bet that the average covers the tail. On current model costs, it does not.

Why Agentic Workflows Break Every Cost Model

The compute multiplier is the structural problem. A chatbot query generates one model call. An agent workflow generates 10–50x that in tool calls, retries, context assembly, and evaluation loops. The Atlantic documented a full reversal in data center sentiment from 'too much capacity' to infrastructure panic, pinned specifically on agents rather than foundation models.

Anthropic, reading the same tea leaves, doubled Claude Code enterprise token costs and is optimizing for high-margin enterprise over developer adoption. Meanwhile DeepClaude now offers "identical autonomous loops" at 17x lower cost by swapping Claude's backend for DeepSeek V4 Pro. The spread between what users pay and what inference costs is widening in both directions.

The Pattern That Works

Microsoft's shift to consumption-based pricing is not thought leadership. It is a pricing architecture change from the company generating more SaaS revenue than any other on Earth. The shape: seats as packaging for prepaid consumption, with overage billing per token, per agent action, or per outcome. Replit's 300% net revenue retention and positive gross margins are the existence proof. They own the full stack and sell to non-technical users who do not price-compare against API bills.

Here is the 2x2 for this sprint. One axis: is the product priced on model cost, or on user outcome? The other axis: is the model a dependency, or a substitutable component? Products in the 'outcome pricing + substitutable model' cell survive both the subsidy unwind and the Jupiter/DeepSeek resets. Products in the 'cost-plus + single-model dependency' cell get repriced by someone else's strategy.

A 13.7-Point Improvement Without Touching Model Weights

Mason Drxy demonstrated that changing only the prompts and middleware in a coding agent harness moved GPT-5.2-codex from 52.8% to 66.5% on Terminal-Bench 2.0. Anthony Maio stated the implication directly: lock-in comes from 'how repo state is fetched, ranked, and compressed into the prompt,' not from the shell or the model. A second researcher claims >20x cost reduction from tuning open models inside well-designed harnesses.

OpenAI has now formalized 'harness engineering' as a distinct discipline, with four named capabilities agents need from their environment: queryable runtime states, progressively-disclosed documentation (AGENTS.md), merge-first CI, and background cleanup agents. This is the first time a frontier lab has described the wrapper as more important than the model.

A thin wrapper around an API with basic retrieval has no moat and is one LangGraph upgrade from commoditization. Domain-specific context compression — knowing which pieces of a user's data matter for a query and how to rank them into a prompt — is defensible IP.

Context Overload Is the #1 Production Failure Mode

Teams shipping agents this quarter consistently report the same finding: the most common reason agents fail is too much context, not too little intelligence. A typical agent has a 200K token context window. The actual instructions it needs are ~400 tokens, buried under tool definitions, reference docs, and brand guides. The SKILL.md pattern — two clear routing lines per skill file — is what's working in production, letting the LLM route without embeddings or retrieval layers.

The Investment Framework

Separate three layers and invest accordingly:

1. Model layer (commodity): Swapping models is a weekend of work. No moat here — leadership rotates monthly across Anthropic, OpenAI, Google, and DeepSeek.
2. Context pipeline (defensible IP): How you fetch, rank, and compress user data into prompts. This is 6 months to rebuild and where engineering investment belongs.
3. Orchestration layer (watch): Sakana's 7B Fugu model hit SOTA on GPQA-Diamond by learning communication topologies via RL — signaling that hand-engineered orchestration DAGs may get displaced by learned policies within 18 months.

Multi-Model Is Now Mandatory

The Claude-vs-GPT character split is a segmentation signal: users pick Claude for drafting and GPT for structured output. Products assuming one model fits both lose half their segment. With MCP reaching 10,000+ public servers and OpenAI, Google, and Anthropic all speaking the same protocol, switching costs between providers trend toward zero. The harness that routes dynamically per task — not the model choice — is where retention lives.