The Product Desk

Three Converging Forces Just Killed Flat-Rate AI Pricing

In the span of 48 hours, the AI pricing landscape underwent a structural break that demands immediate attention from every PM shipping AI-powered features. The evidence is now overwhelming across twelve independent sources — per-seat pricing for AI features is not 'under pressure.' It has already been replaced.

74% of AI SaaS spend is now consumption-based, not seat-based. The market moved while most PMs were still debating it.

Start with the math that makes this irreversible. Agentic coding workflows consume ~1,000x more tokens than chat, with 30x run-to-run variance across identical tasks — and critically, spending more doesn't monotonically improve accuracy. This makes flat pricing impossible: your heaviest user might cost 30x your lightest, and you can't predict which session will be expensive. GitHub figured this out and is switching Copilot to token-based AI Credits on June 1 — $19/month in credits for Business, $39/month for Enterprise, with overage fees. Their CPO explicitly called this building a 'sustainable, reliable Copilot business,' which is a polite admission that flat-rate failed.

OpenAI's $8 ChatGPT Go: A Pricing Architecture Masterclass

OpenAI's internal projections show deliberate self-cannibalization at unprecedented scale: Plus subscribers drop from 45M to 9M (80% decline) while ChatGPT Go surges from 3.1M to 112M subscribers at $8/month (US) or $5/month (India). Run the subscription revenue math: 45M × $20 = ~$10.8B annualized. After the shift: (112M × ~$6.50 blended) + (9M × $20) = ~$10.9B. Subscription revenue is essentially flat. The entire ad stream from 112M users is pure upside.

This isn't cannibalization — it's a pricing arbitrage. OpenAI converts per-user subscription margin into subscription + ad margin while 2.5x-ing their total user base to 122M — comparable to half of Netflix's global base. Salesforce is signaling the same direction with 'Agentic Work Units' priced at roughly $0.60 per appointment. Markets are actively punishing seat-based B2B companies while rewarding consumption models like Cloudflare and MongoDB.

Where Sources Diverge: Cost Compression vs. Cost Expansion

There's a tension worth surfacing. On one hand, DeepSeek V4 priced 97% below GPT-5.5 and Chinese open-weight models are aggressively commoditizing inference. On the other, GPT-5.5 launched at 2x per-token cost over GPT-5.4 — though Ramp independently validated 40% token efficiency gains, making per-task costs roughly comparable. Meanwhile, NVIDIA B200 GPU spot prices surged 114% to $4.95/hour in six weeks. Inference is getting cheaper at the model layer but infrastructure costs are spiking. Your margin depends entirely on which layer you're exposed to.

The winning response is the same regardless: hybrid pricing (base subscription + consumption overage) with intelligent model routing that sends expensive queries to frontier models and commodity queries to cheaper alternatives. IBM's Bob and Cognition's Devin are already doing this. Your model abstraction layer isn't just good engineering — it's your margin protection.

The Exclusivity Is Over. The Leverage Is Yours.

A platform PM picked Azure in 2023 partly because OpenAI models lived there and nowhere else. This week that rationale stopped being true. Microsoft surrendered exclusive distribution rights in the OpenAI restructuring. OpenAI models are cleared for AWS Bedrock 'in coming weeks' (Andy Jassy's words). Microsoft retains non-exclusive model access through 2032 and revenue share through 2030, but the AGI clause that would have stripped those rights is gone. If privileged OpenAI access was part of the Azure decision, that rationale evaporated.

Model access is no longer a differentiator for cloud selection. Every cloud will offer every major model. Your moat must come from your application layer.

The tempting read is that multi-cloud optionality just got cheaper. The quieter signal cuts the other way. AWS enterprise customers are reportedly shrugging off OpenAI availability because they have already locked into Anthropic. The switching cost was never contractual. It was the prompt engineering, the custom tooling, the trained users, the optimized workflows. Teams that told themselves they would 'try multiple models later' are finding that later already happened.

OpenAI's Financial Cracks Are Your Vendor Risk

OpenAI missed its Q1 2026 internal revenue target and previously missed ChatGPT user-growth goals. The company projects $25B cash burn in 2026 (up 213% from $8B in 2025) against a $30B revenue target. CFO Sarah Friar told leadership she wasn't sure revenue growth would support server spending commitments. Altman wants to accelerate the IPO. Friar says the company isn't ready in 2026.

Meanwhile Anthropic has nearly closed the revenue gap despite being founded five years later, is reporting booming sales, and trades higher on secondary markets ($1T vs OpenAI's $880B on Forge Global). Prediction markets give Anthropic a 64% chance of IPOing first. For a PM building on the OpenAI API, the concrete risks break into two components: pricing instability as they close the revenue-burn gap, and strategic distraction as an IPO-focused leadership team ships fewer API improvements. A documented fallback path is worth writing this quarter, not the quarter it is needed.

The Phone Is a Platform Bet — The Miss Is the Signal

The OpenAI partnership with MediaTek and Qualcomm on an AI-first smartphone (2028 mass production, Jony Ive design), where agents replace apps, is interesting but tactically distant. The nearer signal is the missed 1B weekly active user goal for ChatGPT. The ceiling on standalone AI chat products is lower than assumed. Embedded AI is where the retention is showing up, with Google gating natural language email search behind Pro/Ultra subscriptions. The forcing question for the roadmap review: does this feature solve a specific contextual workflow inside a product users already run, or is it a generic AI capability bolted onto the side? Features that answer the first ship; features that answer the second get cut.

Strip Away 'Workday' — This Is About Every Enterprise Category

a16z published a detailed investment thesis for an AI-native Workday killer, complete with wedge strategy, GTM playbook, and explicit founder call. But the patterns are universal: overlay AI features that don't change the underlying architecture, vanity AI metrics that mask low adoption, proprietary tooling creating consultant dependency, and multi-year contracts creating false security while users build workarounds. If your product has 12+ month implementations and a partner certification ecosystem, this is your wake-up call.

When your moat depends on complexity, and AI's core capability is reducing complexity, you have a structural problem no product roadmap can solve.

The Shadow AI Signal Is Your Most Urgent Indicator

HR practitioners are already building Claude MCPs to pull Workday data into other tools and routing approvals through Slack — treating the $10B-revenue HRIS as a read-only backend. This is the exact signal that preceded every major platform disruption: users routing around your product rather than requesting features through it. Workday's $400M 'AI ARR' is mostly Flex Credits procurement theater — CIOs buying AI line items to hit KPIs, not deploying production AI. 'Most organizations have no idea these capabilities exist, let alone how to activate them.'

The Implementation Speed Thesis Changes Everything

The most consequential claim: coding agents can compress enterprise implementations from 12-18 months to ~1 month by ingesting tenant configurations, auto-generating integrations, and replacing proprietary config languages with natural language. Tessera is already proving this with SAP ECC-to-S/4HANA migrations at F500 scale — projects that historically cost $700M and three years. Workday's 10,500+ certified consultants across Accenture, Deloitte, PwC, and KPMG represent a $10B+ services ecosystem whose existence depends on implementations remaining complex.

The Adjacent Budget Wedge

Rather than attacking locked multi-year contracts head-on, a16z proposes entering through adjacent budgets — ops, technology, transformation, innovation, consulting — that aren't controlled by the incumbentcontract. This is exactly how Workday itself displaced PeopleSoft. The tell: every enterprise CIO and CFO has 'AI investment' as a top-line KPI for 2026. That creates budget line items specifically for AI-native alternatives. a16z notes HCM is 'the last large enterprise software category without a serious AI-native challenger' — meaning ITSM, CRM, ERP already have them. Expect a well-funded AI-native HCM startup announcement within 6-12 months.