The Product Desk

Two levers, and only one depends on Microsoft's OpenAI deal

Microsoft EVP Jacob Andreou described Copilot to The Information as "this tapestry of all these different models where our job is to own quality and to own the outcome." In practice, Microsoft is pulling two separate cost levers.

The first is model substitution. OpenAI's GPT-5.6 Sol now handles some advanced Copilot tasks, and in-house MAI models handle some coding work. Microsoft can use OpenAI models without paying OpenAI. No Copilot competitor gets costs that low.

Any team can copy the second lever. Every AI-generated PowerPoint deck used to regenerate about 100,000 characters of generic code. Copilot now pastes in a pre-written script instead. Andreou's claim: "We can actually improve the quality while also reducing the cost a ton." The lesson is to stop paying frontier models for deterministic output, meaning output that is the same every time, such as templates, boilerplate code, schemas and formatting.

Who the cap is aimed at

Techpresso's figures show the target. Meta spent over $105 million on Claude Code in one 28-day stretch. That works out to about $1.37 billion a year, or roughly $1,750 to $3,500 per user per 28 days. The range depends on which headcount applied, and it is Clarity's derivation, not a reported figure. A $10,000 monthly ceiling sits far above that average. It targets the agentic power user whose agent loops can burn $100,000 a month, not the typical seat. Expect your customers' CFOs to ask for your tail cost per user, not your average.

Where the reporting complicates the story

Three accounts agree the money is shifting, not disappearing:

• Microsoft's customers are spending more on Claude through Copilot, and Applied AI says those pass-through payments roughly offset the internal cut.
• Augment reports that customer spend on Anthropic through Azure and Amazon Bedrock keeps growing.
• Microsoft kept its $5 billion Anthropic stake, and Anthropic's $30 billion Azure compute commitment is unchanged.

Andreou also says Claude is still the best fit for certain tasks. The baselines differ slightly: one account describes a peak above $1 billion a year in spending, another a plan to spend at least $1 billion a year. Treat the size of the cut as directional.

The in-sourcing risk applies to anyone who sells seats. Reuters reported that Anthropic's prospectus shows nearly a quarter of revenue coming from two unnamed customers. Once a giant account has a capable first-party tool, one internal mandate can erase your per-seat revenue there. Microsoft showed this when it steered engineers to GitHub Copilot CLI.

The pricing trap

Copilot costs $30 per seat per month, plus usage fees tied to model costs. If your AI fees pass token costs through to customers, every efficiency gain you ship is also a revenue cut. Applied AI reports that enterprise app vendors are already discounting AI add-ons, shortening contracts and writing performance guarantees into deals. Workday's AI monetization stumbled despite heavy promotion.

ServiceNow's new Flow shows an alternative unit. It charges $2 per conversation and $1 per automation run, with a $10,000 annual minimum. The prices are public and payable by card. A price per unit of work survives your own cost cuts. A price per token does not.

If your AI price is tied to tokens, every routing win your engineers ship comes out of your revenue.

What to replace first

A support lead triaging tickets wants a label back. OpenAI's Decisions API is built for that moment. It runs on GPT-6 Luna and returns only predicates, choices or scores, so it swaps out calls already in a product rather than routing among models. The candidates are moderation, intent detection, ticket triage, eligibility checks, fraud flags and LLM-as-judge grading.

AINews prices 1 billion input tokens of classification a month:

• $100 on the Decisions API
• $20 on Perplexity's pplx-decider-v1.1-27b
• $1,360 at Mistral Large 4's input rate, before a single output token

At max effort, GPT-6 Luna scores 38 on Artificial Analysis' index. That places it as a mid-tier model packaged for narrow outputs, and a yes/no call rarely needs frontier reasoning.

Narrow models break outside their lane. In Vals' testing, Jev matched GPT-6 Astra's 97.5% on claim verification at about 1/500th of the cost, then ranked last on LegalBench. Teams test on the easy middle of their traffic. Users send the hard domain cases, so those belong in the test set.

Per-token prices mislead

Mistral lists Large 4 at $1.36 input and $4.18 output per million tokens, yet Artificial Analysis measures $1.13 per task, more than four times the cost of open models with similar intelligence. Vals measured $13.78 per test because the model consumes so much context. Ling 3.1 Flash scores 41 on the same index, above Large 4's 38, at $0.30 input and $0.90 output.

Configuration moves results too. Mistral says many reported failures come from not setting reasoning_effort to "high". The spec sheets cannot agree on Large 4's context window: Vals and Artificial Analysis list 512K tokens, OpenRouter lists 1M. Every vendor bake-off has had this argument before, and the answer is still to test the configuration that will ship.

Open weights give leverage on conflicting timelines

Bloomberg reports AI startups are increasingly swapping big-brand models for cheaper ones, and that Reflection AI's first open-weight model adds pressure on OpenAI and Anthropic to cut prices.

The dates do not line up. Bloomberg has the model unveiled Monday. The Information and TLDR IT say Beam, at roughly 500 billion parameters, is in early access with weights due later in October. Reflection itself says Beam trails the leading Chinese models, a rare bit of modesty in a launch week.

Mistral's date is contested as well. Techpresso reports downloadable Large 4 weights on October 27, and Hugging Face's CEO says it isn't open-weight until the weights actually ship. For a PM, the practical value is a credible fallback for customers whose procurement teams resist Chinese-origin models, plus leverage at renewal.

Images reset what users expect

Google's Nano Banana 2.1 costs $0.034 per image against $0.134 for the previous Pro model, and Google says it beats that model. At 1 million images a month, the bill drops from $134,000 to $34,000. It is rolling out across Search and Ads. Once users see that quality everywhere, a premium image feature has to beat it to justify the price.

The grader is the weak point

An audit of Zapier's AutomationBench found 206 verifier bugs. Fixing them changed 27.9% of grades across 1,235 Kimi K3 runs. A model-selection pipeline built on an LLM judge needs that judge checked against human labels before it approves a cheaper model.

Price the decision with the cheapest model that passes your own eval, and treat list price as a guess until it does. On AutomationBench, 27.9% of grades moved after the verifier fixes.

Three ways the watermark touches your product

It changes your output

textGrain works by subtly adjusting word choices, with no visible symbols. Turning it on is effectively a model change. The most sensitive surfaces are code generation (eligible Codex output is in scope), legal or regulated copy, and brand-voice writing. If you enable it without running evals, you are shipping an untested model change into your most carefully worded outputs.

Your own pipeline may erase it

OpenAI says editing, translation and short passages make the mark harder to detect. Many products rewrite, template, translate or localize model output before users see it. Each of those steps can weaken the signal you would be relying on for compliance. Your post-processing is part of the compliance question, not separate from it.

You can't build a detector on it

A report co-authored with Penn and Yale researchers says detection requires a detector holding a secret key. Access starts with approved researchers and expert organizations. Techpresso's figures show how fragile the mark is: on 400-token passages, swapping 25% of words for synonyms cut detection to 17%. Short text is weakest, and short text is where reviews, comments and short answers live. OpenAI says plainly that a missing mark doesn't prove a person wrote the text.

Whether you have to turn it on

Both reports describe the same split. OpenAI covered its own apps and left the API choice to the companies building on it. Whether you must enable it depends on how counsel classifies your product under the AI Act's transparency obligations. Simplifying AI points to Article 50 as the place to start. You can no longer assume OpenAI handles this for your feature. Once EU users see watermarked ChatGPT output and unwatermarked output from your OpenAI-powered feature, buyers will start asking why.

Who should care most

• Teams building on the OpenAI API with EU users. The on/off decision, the legal sign-off and the quality eval are all on your roadmap now.
• Products that judge authorship, such as edtech, hiring, marketplaces and publishing. Never promise AI detection based on this mark. If authorship matters to your business, apply for detector access as an expert organization, and use any result only as one input to human review.
• Competitors to OpenAI. OpenAI is positioning itself as the compliance-first default. Marketing "watermark-free" output is a risky, short-lived angle while regulation is still setting the minimum standard.

Both reports summarize OpenAI's announcement and related research. Read OpenAI's own documentation before writing specifics into a PRD, and ask your account team which "select models" and "eligible" outputs are actually covered.

If your product calls OpenAI for EU users, the watermark decision is yours, and so is any regression it causes.