The Token
Reckoning

On June 3, 2026, Uber CEO Dara Khosrowshahi appeared on the Invest Like the Best podcast and said: “We blew through our AI budget in a quarter, for the whole year. It is forcing us to adjust.” Patrick O’Shaughnessy posted the clip to X. It reached 510,274 views.^[1]

The arc had been visible inside Uber for months. In April 2026, The Information reported that CTO Praveen Neppalli Naga had told employees the company was “back to the drawing board” after a surge in agentic coding tool usage blew past internal expectations. Roughly 5,000 engineers had been encouraged to use Claude Code and Cursor. They were ranked on internal leaderboards by usage rate. By mid-March, the entire 2026 AI coding token budget was gone. Individual monthly bills ranged from $150 to $2,000 per engineer; a two-hour CTO demo cost $1,200 in tokens. On the Q1 2026 earnings call on May 6, CFO Balaji Krishnamurthy acknowledged it: “Candidly, when we set up budgets for 2026 in November, we underestimated the amount of impact the AI tools could have.”^[2]

COO Andrew Macdonald, appearing on the Rapid Response podcast on May 22, put the ROI question in language that defines 2026 enterprise AI: “When you hear companies talking about, hey, 25% of code commits over the last quarter were AI driven, it’s amazing. But then you sometimes go and you talk to your senior engineering leaders and you’re saying, OK, how many projects that were on the cutting room floor got moved above the line because of the productivity gains? And that link is not there yet. If you’re not actually able to draw a direct line to how much useful features and functionality you’re shipping to your users, that trade becomes harder to justify. AI is not free.”^[3]

Bloomberg confirmed on June 2 that Uber capped spending at $1,500 per month per employee per agentic coding tool, covering Claude Code and Cursor. Simon Willison calculated that at two tools per engineer, the cap amounts to $36,000 per person per year, roughly 11 percent of median software engineer compensation.^[4]

Uber is the most documented case, not the largest. On May 15, Salesforce CEO Marc Benioff appeared on the All-In Podcast: “I am going to probably use $300 million of Anthropic this year at Salesforce coding. We’re wasting tokens. A hot new company will sit between enterprises and Anthropic and OpenAI and make sure you only use their tokens when you actually need them.” His Q1 FY2027 earnings call put a volume figure on it: 28.6 trillion tokens in a single quarter, up 152 percent quarter over quarter.^[5]

“I am going to probably use $300 million of Anthropic this year at Salesforce coding. We’re wasting tokens.”

Marc Benioff, Salesforce CEO, All-In Podcast E273, May 15, 2026

Microsoft canceled roughly 100,000 Claude Code licenses. Mustafa Suleyman, Microsoft AI CEO, told Bloomberg and the Financial Times that Anthropic was “extremely expensive” and that Microsoft’s goal was to “reduce and ultimately eliminate that cost.”^[6] Blackstone President Jon Gray, May 28: “At our companies, their LLM spend was up 15-fold in Q1 of this year over last.” Sam Altman, June 4: AI costs had gone from “never came up” to “a huge issue.” His largest enterprise account burns 100 billion tokens per month, against 100,000 in 2019: a one-million-fold increase in 6.5 years.^[7,8]

The FinOps Foundation’s 2026 State of FinOps report surveyed 1,192 practitioners managing $83 billion in annual spend. 73 percent of enterprise agentic AI implementations went over budget. Those that did exceeded estimates by 2.4 times on average, burning $2.3 million in unanticipated costs per project. 98 percent of FinOps practitioners now manage AI spend, up from 63 percent in 2025 and 31 percent in 2024. Only 8 percent report to a CFO; the other 92 percent sit inside technology organizations, governing AI costs by uptime metrics, not margin.^[9]

The inflection was a February morning in 2025 when Anthropic shipped Claude Code. Before it, developers used AI assistants: one question, one answer, one context window, one invoice line. Finance teams could model it. After Claude Code, developers run AI agents. The agent receives a task, breaks it into steps, writes code, runs tests, reads failure output, rewrites, loops. The developer can go to lunch. The agent keeps running.

Claude Code reached $2.5 billion in annualized run-rate within nine months of general availability, cited as the primary driver of Anthropic’s growth to $30 billion in annualized revenue by April 2026. At Uber, adoption among 5,000 engineers jumped from 32 percent to 84 percent in months. Engineers were ranked on internal leaderboards by usage rate. Nobody in the system had a structural incentive to slow spending when the bills started climbing.

Andrej Karpathy, who joined Anthropic on May 19, 2026, described the shift at Sequoia Ascent on April 30: “I started trusting the system more and more. I couldn’t remember the last time I corrected it. That realization sent me down the rabbit hole of infinite side projects.” He also separated productive agentic use from careless adoption: “Vibe coding is fine for prototypes and personal tools. Agentic engineering is what serious teams need. The agentic engineer does not blindly accept generated code. They design specs, supervise plans, inspect diffs, write tests, create evaluation loops, manage permissions, isolate worktrees, and preserve quality.”^[11]

The standard narrative about AI pricing is correct: the price of intelligence has fallen close to zero. It is also, for most enterprise buyers, beside the point. Commodity tokens fell from GPT-4’s launch price of $60 per million to $0.10 to $0.30 today, a 99 percent decline. But Claude Sonnet has been priced at $3 per million input tokens since March 2024: 27 consecutive months, through multiple model generations, while Anthropic’s revenue grew from $4 billion to $45 billion. Commodity tier deflated. Frontier tier held. Complex reasoning, agentic coding, multi-step analysis: all frontier tier, all flat.

The billing breaks down at the agentic level. One feature through Claude Code: read source files (18,000 tokens), write implementation (22,000), review for bugs (14,000), revise on test failure (28,000), check full codebase for compatibility (180,000 tokens of context, 32,000 response), write documentation (11,000). One feature: 305,000 tokens. At Opus rates, $9. 40 features per week across 20 engineers: $1,440,000 per month. From one team, on one task category, before any other use cases. That number was not in the budget. Fable 5 lists at $10 input and $50 output per million tokens. Its new tokenizer adds roughly 30 percent to token counts on identical inputs. The per-feature number climbs.

Tian Pan, formerly of Uber and Brex, named the mechanism in April 2026: “The unit price of intelligence is going down; the all-in cost of a successful resolution is going up. Token cost falls. Tokens consumed per outcome rise faster.” He called it “the token cost illusion.” His data: the honest cost-per-success divides total spend by count of successes. That number is two to four times the naive cost-per-token calculation, because failed and partial outcomes still cost money.^[10]

The FinOps Foundation’s report confirms the Jevons Paradox at work: enterprise blended token costs fell 67 percent year over year, from $18.40 to $6.07 per million tokens. Total AI spend accelerated anyway. The share of organizations spending more than $100,000 per month on AI doubled year over year. Cheaper tokens created headroom for more agentic deployments; each deployment consumed tokens at rates that more than offset per-unit savings. America Online ran the same experiment in reverse in 1996: flat-rate $19.95 per month unlimited access replaced hourly billing, usage exploded 500 percent, and the network collapsed. Enterprises knew the per-token price. The 73 percent who missed their forecasts did not miss because they ignored the rate sheet. They modeled chat-mode consumption and deployed agentic loops.

The market is splitting in two. Frontier labs are raising prices into their IPO windows. Open-source and efficient models keep falling, some to near zero. This week the gap between the cheapest capable model and the most expensive reached 70x.

Gemini Flash 3.5 was below $0.30 per million tokens twelve months ago. It is not there anymore. Claude Fable 5, launched June 9, carries a $10 per million input token price. The model it effectively replaces for production workloads, Claude Sonnet, was $3. That is a 230 percent increase for equivalent-tier access. Both labs are moving in the same direction at the same moment. The S-1 filings explain the incentive.

When the gap between the cheapest capable model and the most expensive one is 70x to 100x, picking one model for everything stops making sense. A team spending $10 per million tokens on document classification is paying 70 times the necessary rate. The routing layer is not an optimization. It is the only rational response to a market that has split in two.

The June 2026 Price Sheet

Eight models, one market split in two. Sort by price or capability, filter by where each is heading — frontier tier climbing, open tier in freefall. Blended rate, 30% input / 70% output.

Sources: official provider pricing pages, June 2026. Fable 5 offline since June 12 (export controls).

Nobody explained this when enterprises adopted AI coding tools: Claude Code, Codex CLI, Gemini CLI. These tools are not interfaces. They are harnesses. An interface can be swapped without losing much. A harness accumulates organizational knowledge, encodes institutional decisions, and binds enterprise teams to output patterns that cannot be ported elsewhere.

Each lab has its own harness lock-in mechanism. The most visible is Anthropic’s CLAUDE.md: a project-level configuration file encoding architecture conventions, naming patterns, testing frameworks, and months of accumulated institutional knowledge. Teams running Claude Code end up with CLAUDE.md files tuned to Claude’s XML-delimited prompt structures; pointing them at GPT-4 or Gemini produces degraded results requiring weeks of re-tuning. OpenAI’s Codex CLI builds equivalent lock through codex-1’s specialized fine-tuning, embedding coding conventions no other provider can replicate. Google’s Gemini CLI accumulates lock through tight Cloud Run and Firebase integration, so every Gemini-specific function-calling schema and system prompt convention becomes another migration cost. The pattern is the same across all three labs. The harness files differ.

Tian Pan documented the accounting dimension in April 2026. His opening scenario: “The first time the CFO asks what the assistant costs per month, the engineering team produces a number. The second time, a different team produces a different number. The third time, finance produces a third number, and somebody opens a spreadsheet that re-derives the bill from spans because nobody trusts any of the previous answers.” The structural driver: a single user request can trigger an agent owned by one team, which calls a retrieval tool from a second team, which calls a model from one provider, which streams results through a re-ranking tool from a third team, which calls a different model from another provider. “One click; five owners; two invoices that arrive a month apart.”^[10,14]

“Switching harnesses is not like switching text editors. It is closer to switching version control systems mid-project, and nobody told the procurement team that was the kind of thing they were buying.”

The other half of the trap is the behavioral fingerprint. Months of production use build institutional knowledge: which prompts reliably produce well-structured code, how to frame a debugging task for useful output on the first try. That knowledge is real and valuable. It is also completely non-transferable, locked to a specific provider’s output patterns and unreachable by any migration script.

Every frontier lab has made design choices that deepen this lock. OpenAI’s Codex CLI runs on codex-1, fine-tuned through reinforcement learning on real-world coding tasks; that specialization is non-public and non-replicable regardless of benchmark scores. Google launched Gemini CLI in June 2025 as open-source, accumulated 105,000 GitHub stars and 6,000 community-contributed pull requests, and integrated natively with Cloud Run and Firebase. Anthropic built CLAUDE.md as a first-class product feature that makes Claude Code more productive over time. These are not accidents. Each decision that deepens behavioral differentiation is also a decision that raises the cost of leaving.

The math behind frontier price stability is the same for all three labs. Sonnet has held at $3 since March 2024 because the cost of switching exceeds the potential savings. In that period, Anthropic grew from $4 billion to approximately $45 billion in annualized revenue. No enterprise customer has publicly announced switching away from a frontier lab’s primary tool at scale. A company spending $5 million monthly on Sonnet, saving $1 million monthly by migrating at equivalent capability, generates $12 million in annual savings. The migration cost, roughly a quarter of original development time plus engineering labor, runs $10 to $25 million. Tian Pan: the eval harness re-validation alone “can eat a quarter of your roadmap when you actually try to migrate.” The exit costs more than staying.^[14]

The Model Context Protocol, introduced by Anthropic in late 2024 as an open standard, is the industry’s most serious attempt to reduce coupling between harnesses and specific model providers. MCP standardizes how AI agents communicate with external tools: file systems, databases, APIs, code execution environments. The gap between the promise and current reality is instructive. MCP addresses the tool communication layer, not the output behavior layer. A harness tuned for one lab’s specific response patterns still needs weeks of re-tuning when pointed at a different model, even if both speak MCP for tool calls. The structural coupling lives below MCP’s abstraction level.

The ROI analysis runs in both directions. METR, the AI safety evaluation nonprofit, published a randomized controlled trial on experienced software engineers using state-of-the-art AI coding tools on real tasks. Engineers with AI assistance completed tasks 19 percent more slowly than engineers without it. Not because the tools were bad. Because the overhead of integrating AI into workflows engineers already knew how to execute, context-switching, verification work, time deciding whether to trust a suggestion, offset the speed gains. High-volume, well-defined, autonomous tasks show the clearest productivity gains. Broad deployment into general workflows where engineers already have expertise often does not. Both findings are true simultaneously. The deployment context determines everything.^[15]

Anthropic’s planned IPO is the first major test of how public markets will value this structural position. The company is widely expected to file an S-1 by August 31, 2026. Goldman Sachs, Morgan Stanley, and JPMorgan are competing for the mandate. At the reported $965 billion pre-money valuation from Series H talks, this would be the largest U.S. technology listing since Aramco. A 21-times revenue multiple against $45 billion annualized. Salesforce peaked at roughly ten times revenue at its high. OpenAI’s reported valuation trajectory and Google’s cloud AI revenue growth suggest the same structural premium is being priced across the frontier lab category broadly. No publicly traded U.S. technology company has sustained a 20-times revenue multiple without eventually compressing, either through growth that catches the multiple or through multiple compression as growth slows.

The frontier labs run a business that standard valuation frameworks do not easily capture. All of them share the same structural economics: very high fixed costs to manufacture the product, margins that improve with scale, and rapid growth driven by consumption-based billing. Anthropic illustrates the pattern in the most disclosed detail available. Compute spending: approximately $1.25 billion per month on GPU time, memory bandwidth, cooling, power, and infrastructure management at scale. Gross margins: approximately 44 percent. The median large-cap software company earns 72 percent gross margins, but software companies do not pay $1.25 billion monthly to manufacture their product. The most revealing detail in Anthropic’s capital structure involves memory chips. Samsung, SK Hynix, and Micron, the three dominant global suppliers of High Bandwidth Memory required for large-model inference, all hold equity stakes in Anthropic simultaneously. That has not occurred before in the history of the semiconductor supply chain.

The hyperscaler channel adds structural complexity for all frontier labs. Amazon Web Services, Google Cloud, and Microsoft Azure each resell Anthropic’s API under their own branding: Amazon Bedrock, Vertex AI, Azure AI Foundry. This channel represents an estimated 38 percent of Anthropic’s revenue. The entanglement runs across the entire category: Microsoft is OpenAI’s largest customer and primary infrastructure partner; Google is simultaneously DeepMind’s parent and its principal distribution channel. Each frontier lab’s commercial fate is bound to businesses that are also, in meaningful ways, its competitors.

The renewal cliff is a calendar event. Multi-year enterprise AI contracts signed in 2024 are expiring now. Teams that signed in Q1 2024 are renewing in Q1 2026. Teams that signed in Q3 2024 are renewing in Q3 2026, precisely when Anthropic’s IPO roadshow will be underway. In 2024, the enterprise team had a pilot result and an executive mandate, small spending, sparse governance. In 2026, they have eighteen months of production data, an actual invoice history, and in many cases a routing layer that has already demonstrated 60 to 75 percent of their consumption could move to cheaper models. Bessemer Venture Partners coined the term “renewal cliff” in their February 9, 2026 AI Pricing and Monetization Playbook to describe exactly this moment.^[12]

On June 9, 2026, Anthropic retired the “Opus” product line and launched Claude Fable 5 publicly, with Claude Mythos 5 in restricted access under Project Glasswing. The naming choice is doing structural work. “Opus” was the reference anchor in every procurement spreadsheet, every benchmark comparison, and every competitor teardown. Opus 4.8 at $5 per million input tokens and $25 per million output tokens was the number enterprise buyers used to evaluate Anthropic spend. Fable 5 and Mythos 5 are new products with new names. The old comparison does not compute.^[19]

Fable 5 and Mythos 5 ship with a new tokenizer and a new sticker price: $10 per million input tokens, $50 per million output tokens. Opus 4.8 was $5 input, $25 output. Fable 5 is exactly 2x Opus 4.8 on both dimensions. That is the base increase. The tokenizer adds more on top. Identical source text generates roughly 30 percent more tokens on the new tokenizer. A workflow that cost $0.50 on Opus 4.8 at 100,000 tokens costs $1.30 on Fable 5 at 130,000 tokens, a 2.6x effective multiplier. The price doubled. The bill more than doubled.^[19,22]

The timing is not incidental. Anthropic filed its S-1 on June 1, 2026. OpenAI filed on June 8. Fable 5 and Mythos 5 launched on June 9. Two frontier labs on simultaneous IPO roadshows made their major pricing and product announcements in the same week. An S-1 requires disclosure of revenue trajectory, growth rates, and customer concentration. A price list showing premium products above $50 per million output tokens tells the prospectus reader that the frontier tier commands a durable premium, that switching costs work, that the revenue multiple is defensible. The enterprise customer reading the same price list on June 9 has a different set of interests than the institutional investor the roadshow begins pitching in August.^[20,21]

On Anthropic’s own announcement post, user reactions were immediate. Four comments, verbatim:

Penn, Anthropic’s VP of Developer Experience, addressed the pricing at the Fable 5 launch. Cost was “very top of mind,” Penn said. The response to concerns: “you just get a higher ROI by having more intelligent models.” That argument holds at the task level when the capability jump is real. It is also the argument every enterprise software vendor makes at every renewal. Oracle made it. SAP made it. Salesforce still makes it. The ROI argument does not resolve the structural question of whether the customer has any credible alternative if they disagree.^[20]

“The moat is not the model. It is the harness. Enterprises are not paying for any model’s benchmarks. They are paying for the cost of not having to rebuild everything they have built around that model’s output patterns.”

The lock-in forces cut in both directions for any frontier lab’s long-term trajectory. Favorable: harness-dependent customers rarely churn. The enterprise that has built six months of proprietary harness configuration into its engineering workflow is not going to non-renew because the per-token rate rises five percent. The total cost of staying is almost always lower than the total cost of leaving. Net dollar retention in harness-dependent businesses tends to be high. Unfavorable: the same forces that protect near-term revenue invite scrutiny from procurement departments, regulators, and investors who have seen this before. Oracle and SAP built the deepest enterprise lock-in of the 1990s and 2000s. They also attracted the most aggressive legislative attention and demands for interoperability standards once the switching cost became a matter of public concern rather than private annoyance.

A mid-sized enterprise software company published its token cost story in early 2026. It had been routing 120,000 LLM API calls per day entirely to Claude Sonnet, because that was the model the engineering team had tested against. An audit found that 45,000 were simple form-text extraction, 30,000 were support ticket classification, and only 10,000 required frontier-tier reasoning. After deploying LiteLLM with a 30-line rule-based classifier, daily cost fell from $1,420 to $594: a 58 percent reduction. Monthly savings: $30,000. Total implementation time: one afternoon.^[16]

The routing layer has scale now. OpenRouter reported 8 million developers in May 2026, up from 2.5 million a year earlier: a 3.2-fold increase. Weekly volume went from 5 trillion tokens in November 2025 to 25 trillion in May 2026. The company raised $113 million in a Series B led by CapitalG at a $1.3 billion valuation. 1.5 quadrillion tokens per year, up 15-fold from twelve months earlier, processed for 8 million developers. Andrej Karpathy described OpenRouter as “the de facto ‘transfer switch’ of AI today.”^[13]

LiteLLM abstracts 140-plus providers behind a single OpenAI-compatible API. RouteLLM achieves 85 percent cost reduction on conversational benchmarks by directing simpler requests to cheaper models. The economics: DeepSeek V3 costs $0.27 per million input tokens. Kimi K2, released June 2026 by Moonshot AI, costs $0.15. Claude Sonnet 4.6 costs $3.00. A company routing classification tasks to Sonnet by default pays 20 times the necessary rate on the majority of its calls. The routing layer does not make the AI smarter. It makes the billing rational.^[14]

The case studies that follow show what the numbers looked like for organizations that actually ran this experiment in 2025 and 2026.

Cloudflare routes 241 billion tokens per month across 3,683 internal users. For security workloads, routing to open-source models on Workers AI achieves 77 percent cost savings. A separate configuration service handles model selection per agent, so models can be swapped without changing the CI template. The routing decision is encoded in policy, enforced automatically, and auditable after the fact.^[17]

The second approach is deliberate multi-harness architecture: running two harness environments in parallel. Not because both are needed simultaneously. Because the option to switch only exists if the second harness is operational. A team that has been running a project in both Claude Code and Gemini CLI for six months, even at reduced intensity on the second, has already done most of the migration work. When a contract renewal conversation happens, that team has a credible alternative. A team with no fallback does not.

Self-hosting becomes cost-competitive above roughly 700 million to one billion tokens per month in sustained, predictable workloads. Below that threshold, capital cost exceeds API cost. The frontier models that drive most enterprise AI value are not available for self-hosting at any price. The practical self-hosting opportunity is in the commodity tier: running DeepSeek V3, Kimi K2, or Llama 4 Maverick on internal infrastructure for classification and extraction, while keeping frontier-tier inference on provider APIs for the tasks that require it.

Enterprise AI spending splits into three categories by sensitivity to maturity. First: complex agentic coding, multi-step analysis, high-stakes document work that requires frontier capabilities. The ROI cases are documentable; the switching costs are high. This spending holds. Second: commodity work on frontier models by default, the classification and extraction tasks that 60 to 75 percent of API logs reveal were never frontier-worthy. The routing layer captures this category and redirects it at one-tenth to one-twentieth the cost. Third: experimental deployment without clear success metrics. Agents given open-ended mandates, chatbots added to every customer touch-point without termination conditions, coding tools deployed to every developer regardless of role fit.

Salesforce GM of CRM Applications Bill Patteson put the question at a Jefferies conference in late May 2026: “Are these tokens leading to yield? We’ve all heard about tokenmaxxing as a sort of strategy. Well, now the counter theory is, is this really output maxing? We know that we’re actually getting content, but is that content actually producing business yield?” That is what Uber’s COO was asking on May 22. It is the meme Sam Altman quoted back to his enterprise customers on June 4.^[18]

Wall Street Names the Bifurcation

On June 11, Citadel Securities published a macro note titled “Tokenomics.” The author was Frank Flight, Global Macro Strategy. He had been making this argument since February 2026, in a note called “Infl-AI-tion Risks” that responded to a viral scenario about a 2028 Global Intelligence Crisis. He was now saying the argument had crossed from thesis to observable behavior.

The primary exhibit was the Silicon Data LLM Token Expenditure Index, which tracks the price and mix of LLM token usage across enterprise accounts. The index had more than doubled from late 2025 through May 2026 as enterprises scaled up AI coding assistants and agentic workflows. In early June, it peaked and reversed. By the date of publication, it had fallen for seven consecutive sessions, the longest losing streak since January. The 21-day log annualized growth rate had gone negative.

Flight’s read of the data: the decline most likely reflects substitution toward cheaper models, not a collapse in aggregate demand. Silicon Data itself notes that the index “can fall when individual model prices decline, when users substitute toward more efficient model choices, or when the market diversifies away from expensive concentration.” That interpretation is consistent with everything happening in enterprise procurement simultaneously. The index was measuring the substitution signal; it had finally arrived in the data.

“Economic theory tells us that prices perform three basic functions: they signal scarcity, create incentives for substitution, and ration scarce resources toward their highest-value uses. These functions apply clearly to AI.”

Frank Flight, Global Macro Strategy, Citadel Securities, “Tokenomics,” June 11, 2026

The corroborating evidence Flight catalogued will read as familiar to anyone who followed the enterprise billing disclosures of May 2026. Amazon scrapped its internal token usage leaderboard after engineers began gaming it, optimizing for token volume as a proxy for productivity rather than for output quality. Meta pulled a similar internal leaderboard in April. Microsoft canceled several thousand Claude Code subscriptions once the bills cleared budget expectations, with engineers redirected to GitHub Copilot. Uber burned through its full 2026 AI coding budget inside four months. These are not isolated incidents. They are what substitution looks like in aggregate.

Flight’s conclusion was not bearish on AI. His conclusion was structural: “We do not think this implies that the frontier of inference-intensive AI will be abandoned, only that it is likely to be concentrated among a narrower set of firms with the balance sheets to absorb the compute cost, the research depth to deploy it effectively, and, most importantly, the operating domain to scale the rewards from solving genuinely hard problems. For the economy at large, simpler models may be the more cost-effective, productivity-augmenting pathway until physical constraints are eased. We hence see growing signs of a bifurcation in frontier vs ‘everyday’ AI usage.”

The physical constraints he named explicitly: compute, power, cooling, memory bandwidth, inference budgets. The narrative of frictionless AI scaling runs into each of these sequentially. This is not a capabilities argument. It is infrastructure economics, and it applies regardless of whether the next model is twice as capable as the last.

The bifurcation he described produces a two-tier market. A narrow top layer runs expensive frontier models on problems where the payoff justifies the cost: drug discovery, materials science, complex financial modeling. A broad base of everyday enterprise users runs cheaper, smaller, open-weight or fine-tuned models for coding assistance, summarization, classification, and the workflows that API logs reveal were never frontier-worthy in the first place. For asset prices, the question becomes which layer of the stack captures the margin in each tier. For enterprise technology teams, the question has already arrived: which tier are you actually in, and are you paying accordingly?^[23]

Frontier pricing is not going to fall the way commodity pricing fell. The structural incentives run the other direction. Both leading frontier labs filed S-1s this week. Their hyperscaler partners resell at margin. Every enterprise harness that runs six months without a migration attempt makes the next one more expensive. The system is designed to retain you at the current rate or higher.

That does not make you stuck. It makes the audit non-optional. Every team running AI at rational cost in 2026 has done the same thing first: they measured. They can tell you what a pull request review costs in tokens, what a support ticket classification costs, what a contract extraction costs, start to finish. They built that visibility before they built the routing layer. The routing layer paid for itself in the first month.

The question is not which model is best. It is whether you know your per-workflow token cost. Not your monthly invoice total. Not the blended API average. The cost to complete one named workflow, end to end. If you do not have that number, every renewal conversation happens on the provider’s terms, not yours. Pull the API logs. Build the denominator. The providers already know what you cost them to serve. You should know what they cost you.