There is no free AI: the economics of inference and the window of opportunity
A budget meeting in a French sales department, early 2026. The CFO asks the simple question: "how much does AI cost us in pre-sales?". The IT lead does the math: "30 euros per user per month; on an average pre-sales effort, the team mobilizes eight people for three months -- so 720 euros per tender. On a large tender mobilizing fifteen people for six months, we reach 2,700 euros." The CFO notes the figures. The sales director nods. The ticket looks reasonable, almost virtuous: it feels like a genuine investment, calibrated to the size of the tender and the length of the cycle.
The calculation is wrong, and it is wrong in an interesting way. The ticket stays reasonable; it simply weighs, against the total cost of a serious pre-sales effort, an invisible fraction -- for a functional benefit that is heavily sub-optimized. The 30-euro tool does not handle serious tenders; the previous article traced the mechanics of economic throttling that render it powerless on long corpora. The tool that genuinely handles them carries a different price tag, and that price tag has been masked for three years by the venture capital flooding the generative-AI ecosystem. When the CFO asks "how much does AI cost," the honest answer is: "it depends whether you want the illusion or the work."
This article makes visible the real cost of AI usage sized for real work, sets it beside the visible cost of consumer subscriptions, and defends a counterintuitive thesis: the current window is the cheapest we will see for a long time.
The raw cost of serious inference
Anthropic's pricing on Claude Opus 4.7, released in April 2026, is public: 5 dollars per million input tokens, 25 dollars per million output tokens. This is the price of a SOTA-class model -- State Of The Art, meaning the class of models that define, in real time, the frontier of what the machine can do in reasoning, in analysis of long corpora, in coherence across long inference chains. The SOTA class today comprises a handful of models: Claude Opus 4.7 at Anthropic, GPT-5.5 and GPT-5.5 Pro at OpenAI, Gemini Pro at Google, and a narrow circle of challengers. It is the toolset of serious intellectual work, sharply distinct from that of consumer chatbots.
On 23 April 2026, OpenAI did exactly that: it released GPT-5.5 and doubled its API pricing relative to GPT-5 -- input rising from 2.50 to 5 dollars per million tokens, output from 15 to 30 dollars. Google holds Gemini Pro slightly below, but the slope is identical. No SOTA-class model is dropping significantly, and for the first time since 2023, the slope has reversed: prices are rising.
A complete tender file weighs, on the read, between 200,000 and 400,000 tokens -- CCTP, RC, BPU, DQE, DPGF, AE, the consultation rules, lots, technical appendices, and the technical proposal of the previous incumbent obtained through public channels. This raw ingestion represents between one and two dollars of input. Yet reading a file is a long way from answering it.
A real response demands multiple operations: read, analyze, map the requirements, identify the zones of strategic divergence, simulate the weighting formula, cross-reference competitive references, write a first proposal, challenge it, rewrite it. Each operation consumes tokens both in reading the prior context and in producing new text. An agent that orchestrates these steps properly does not make a single pass over the DCE: it makes ten to thirty, each one re-reading all or part of the prior context at every turn. A first automatic generation of a fifty-page technical proposal, with no human intervention, typically consumes between twenty and sixty dollars of tokens on Opus 4.7.
That is the floor cost. The real cost lies elsewhere.
Why a tender costs 400 dollars, not 50
The first generation is rarely the final file. In a real production chain, the bid manager intervenes repeatedly over a cycle of eight to twelve weeks.
They dialogue with the AI to reorient the strategy after the first pass. They progressively introduce information that was not initially available -- the internal pricing grid, the HR arbitration on the available team, the history of the relationship with this client. At each step, the agent's context swells, sometimes up to the million-token mark, and each conversational turn replays that context at the input rate.
The DCE evolves: the buyer publishes an update, adds an appendix, redefines the scope. The agent must redo part of the work. The published questions and answers trigger a new revision. The human-AI challenge -- "this section does not hold, propose an alternative that assumes the risk on the schedule but secures the scope" -- sets off five to ten rounds of refinement. Successive reviews by the sales director, the legal director, the executive sponsor each impose their share of localized rewrites. The last-minute changes, in the twenty-four hours before submission, are among the most expensive because they operate on the full mature context.
On a serious tender run with a serious human loop, total token consumption typically falls between 150 and 400 dollars. A complex file -- variants, options, multiple lots, an oral defense prepared with rehearsal simulation, an agent maintaining the context beyond a million tokens over the final weeks -- can exceed 600 dollars per tender, or even approach 1,000 dollars on the very large multi-lot tenders with an oral defense.
This range signals the tender that was genuinely worked, far more than a failure to optimize. A file that consumes thirty dollars of tokens is a file no one challenged in depth.
Why no long-context model is cheap
The objection always comes: "and open source?". Llama, DeepSeek, Mistral Large publish capable models under open licenses. The promise of inference at zero marginal cost remains written into the narrative.
It does not hold up for professional use. Three reasons.
Inference on a Llama 405B or DeepSeek-V3 class model over a million tokens in context demands several H100 or B200 GPUs allocated for the duration of the processing. The hourly cost of a cluster that sustains this load, operated in-house by an IT department, quickly exceeds the API price of a proprietary SOTA model -- without the quality advantage.
Third-party hosting (Together AI, Fireworks, Groq, Anyscale) makes deployment more accessible, but the rebilled price mechanically converges toward the inference cost of the proprietary operators. The gaps published in the comparisons do not hold at comparable power and effective long context.
Attention remains, in most architectures, quadratic in complexity over context length. Doubling the window quadruples the inference cost. Recent optimizations -- Flash Attention, Ring Attention, Sliding Window -- reduce the constants but do not change the asymptotic class for contexts of several hundred thousand tokens. The physics of memory does not vanish by migrating to open source.
A dry conclusion: there is no economic shortcut, in 2026, to performant 1M-context. When a provider offers a price that defies the market, the effective window is shorter, recall of the median tokens collapses (cf. Lost in the Middle, Liu et al. NAACL 2024), or generation quality regresses. What looks free never is, structurally.
The Claude Max paradox
Anthropic offers two tiers of individual-usage subscription: Claude Max 5x at 100 dollars per month, and Claude Max 20x at 200 dollars per month, which open up respectively five and twenty times the capacity of the standard Pro plan. A consumer's first reflex faced with the 200-dollar price is: "that's very expensive for a personal tool." The first reflex of a user who has lived through the API pricing is: "it's the most subsidized offer on the market."
Both reflexes are true. They are not addressed to the same person.
For moderate usage -- a few conversations a week, occasional tenders, incidental code -- the 200-dollar plan is wasteful. ChatGPT Plus at 23 dollars or standard Claude.ai at 20 dollars are enough. For deep usage -- a bid manager orchestrating two complete tenders a week, a senior consultant handling audits of sixty documents, a developer holding the agent in a long loop over code -- the consumption observed among the power users who have switched to Max corresponds, at equivalent API pricing, to several thousand dollars of tokens per month.
The plan subsidizes deep usage. It is expensive only for those who use it little. It is one of the very rare products where the inverse of the spontaneous perception holds true -- and where peer recommendations diverge radically depending on the depth of usage of the person recommending.
This paradox nonetheless has a flip side -- a double-edged value-for-money that surfaced in March-April 2026. Anthropic quietly tightened the session limits on its Max users during peak hours: five-hour sessions were being consumed in ninety minutes on Max 5x, certain prompts pushed a Max 20x gauge from 21% to 100% in a single pass. Official confirmation from Anthropic followed -- roughly 7% of users were now hitting limits they had not hit before. The reason given publicly stayed vague; the structural reason is clear: demand exceeds available GPU capacity, and the vendor arbitrates, without saying so head-on, by throttling the most intensive usage to preserve service quality at scale. The plan stays subsidized, but the subsidy becomes conditional.
The same phenomenon, more visible still, appeared with Claude Mythos, the frontier model Anthropic announced on 7 April 2026 -- a model on the order of ten trillion parameters, trained on Nvidia's Blackwell generations. Its distribution stayed restricted to some fifty hand-picked partners (the Project Glasswing program), with an API price of 25 dollars input and 125 dollars output per million tokens -- five times the price of Opus 4.7. The public justification foregrounds safety; the internal communications quoted in the trade press say something else. Anthropic openly acknowledges that Mythos is "very expensive for us to serve, and will be very expensive for our customers to use," and is working to make it more efficient before any wider release. The restricted distribution is, for a significant part, a distribution constrained by the cost of inference and the available industrial capacity, more than a simple precautionary measure.
These two signals converge. They indicate that the SOTA has moved, in 2026, close to the limit of what the industrial ecosystem can serve at the going price. The 200-dollar plan, the API price of Opus 4.7, the doubled price of GPT-5.5 -- all reflect this tension. Far from a point of arrival, they mark a step upward.
The CIO's dilemma
A CIO buyer, in 2026, faces a grid of three options.
Option A -- strict restriction. You deploy Microsoft Copilot, ChatGPT Enterprise or Gemini for Workspace at 20-30 euros per user. Governance is simple, the ticket is known, integration into the IS is eased. The previous article documented what collapses: on the complete tender, on the minutes of a long meeting, on transversal document analysis, these tools fall back to throttled RAG architecture and produce deliverables that are fluent but structurally insufficient. The field feedback, at scale, is uniformly "very, very disappointing" on the high-stakes files. The hidden margin -- time lost reworking insufficient AI outputs, tenders lost for lack of depth, legal exposure on minutes that are false by omission -- is masked by the simplicity of the visible ticket.
Option B -- premium with self-rationing. You deploy API access to Anthropic, OpenAI or Google, but governance imposes intermediate models "for the margin": Sonnet, Gemini Flash, Grok, GPT-4.1 mini. The unit price drops by five or ten times. The reasoning capacity drops too, but less visibly. The result on a serious tender is competent but median reasoning, which misses the strategic inflections that only a premium model identifies. This option is in fact more dangerous than option A. With Copilot, the user is wary -- the tool is public, the limit is known, you proofread before signing. With a deployed premium API and an intermediate model running in the background, the user has the sensation of having accessed a sophisticated infrastructure, their confidence in the output rises, their critical vigilance drops. They sign off on analyses that look solid because they are fluent, but that miss precisely the zones where genuine reasoning capacity would have made the difference. The final disappointment is heavier, because it comes with errors validated along the way.
The same logic holds for the other form of self-rationing, more discreet: the massive use of RAG across the whole corpus. You index the documents, you inject the retrieved fragments at each question, you save on a premium long context. The output looks informed, sourced, structured. But RAG plays on semantic proximity between the question and fragments -- it plays neither on logic, nor on judgment, nor on transversal connection. On a tender, the typical strategic question -- "what coherences does this file demand at the intersection of weighting formula, requested references, and schedule?" -- has no answer in any isolated fragment. RAG returns paragraphs relevant by keyword, the model composes a coherent answer on that basis, and the user receives a deliverable that seems considered but that never saw the file in its entirety. It is the same illusion as option A, disguised behind a more expensive infrastructure.
Option C -- the only one that pays off. A premium model (Opus, full GPT-5.5, full Gemini Pro) on the high-stakes acts -- strategic analysis of the DCE, reframing of the file, production of the critical sections, defense rehearsal. Optimized support models (Sonnet, Gemini Flash, Grok) on the medium-stakes acts -- requirement extraction, first chapter skeleton, spell-checking. Rigorous architectural control -- deciding which act goes to which model, and measuring consumption. Rigorous cognitive control -- an upstream human framing, epistemic operators set by hand, a critical review of the outputs. And demanding change management on the user side.
This option implies a leap in tooling that few organizations have crossed today: moving from the chatbot to a genuine agentic mode. An agent that interacts directly on the documents -- that opens them, reads them, compares them, executes write and exploration commands, structures its own steps, keeps the trace of its reasoning. Instead of the assistant into whose chat window you copy-paste excerpts, a system that operates autonomously on the corpus, under human control. The technology is mature, the transition is technically easy in 2026 -- but too few sales departments and IT departments are positioned on this tooling. It is precisely this gap that constitutes the competitive lever of the next two years.
The visible cost of option C is higher than that of option A. The total cost of ownership is markedly lower. Even a file that consumes 1,000 dollars of API over its production weeks remains a fraction of the total cost: it is the equivalent of one to two days of a senior consultant's work, on a pre-sales effort whose total budget -- bid manager time, technical expertise, sales support, oral defense -- runs into the tens or even hundreds of thousands of euros in a serious IT services firm. The real question is not about the absolute cost of premium AI, but about the capacity of the extra inference cost to buy back several man-days and to improve the quality of the deliverable. On a high-stakes tender, the answer is mechanically yes.
It is, however, a calculation that French finance departments struggle to make, because it pits a visible monthly expense against a diffuse creation of value -- conversion rate, margins on tenders won, cycle speed.
Change management is half the investment
Giving Opus to untrained users amounts to giving a concert piano to a beginner. The cost of the instrument looks absurd against the sound produced. The answer lies in training the pianist, rather than in scaling the instrument back down.
The cognitive discipline to transmit is precise.
State the intention clearly. A user who prompts "write me a technical proposal" wastes the model's capacity. Explicit framing -- "the client is a public administration, the scope is limited to lot 2, the differentiation strategy is securing the schedule, the expected tone is reassuring on operational risk and assertive on the quality commitment" -- redistributes the model's completion distribution toward the useful zone. It is the most profitable human operation in the chain. An hour spent reformulating the frame is worth, in leverage, ten hours of iterative prompting on a standard frame.
Provide the exact context, no more, no less. A user who dumps the whole DCE into the window without hierarchy drowns the model. A user who provides only the CCTP misses the relevant internal references. The right dosage is a discipline that is learned. It can be measured: if the output does not hold, the context provided was either too poor or too flat.
Minimize interactions through structured cycles. A dialogue of twenty poorly structured turns costs more and produces less than a dialogue of five turns with clear checkpoints. The effective method alternates long generation, targeted human review, calibrated correction instruction, framed regeneration. It is transmitted; it is not discovered on one's own.
Set the critical operators by hand. The passages carrying contractual or strategic weight -- delivery commitments, pricing formula, reversibility clauses, qualifying references -- are written or proofread by hand. Prices and commitments are never left to the model.
In the first months, some users will blow through the budget by an extra 200 to 300 dollars per month per person, sometimes more. That is the learning curve, and it is normal. The return on investment is measured in the tender conversion rate, and in the upskilling of the employee who will progressively optimize their interactions -- stating the frame more precisely, providing denser context, structuring the dialogue more tightly -- before AI inference ceases to be cheap. The management that penalizes over-consumption during the learning phase kills the transformation it paid to set in motion.
TenderGraph TITAN: the agentic system that optimizes inference for you
An organization that poses the problem honestly soon reaches the same conclusion: letting its employees "figure it out" with agentics, without a system, without a method, without a frame, is to guarantee one of the two worst scenarios. Either adoption fails because the complexity of usage discourages -- the user goes back to Word and their old method after three failed attempts. Or adoption succeeds badly -- the user massively consumes premium tokens for median results, because they have neither the framing, nor the inference sequence, nor the cognitive discipline that performant usage demands. In both cases, the organization pays without reaping.
It is precisely this gap that TenderGraph addresses with TITAN. TITAN is a cognitive agentic system designed for pre-sales production: it operates directly on the documents of the DCE, executes the inference chain in the right order, asks the right questions at the right moments, applies the right analytical logics -- weighting formula, BPU/DQE cross-referencing, reading of the implicit frame of the CCTP, identification of the zones of strategic divergence. The benefit is twofold. On the time side, the agent automates the mechanical steps that the bid manager should not redo by hand. On the inference-cost side, the agent drastically optimizes token consumption -- a pre-structured inference chain, no human friction generating redundant conversational turns, framing better set upstream, which reduces downstream regenerations. On a serious tender, a well-designed cognitive agentic system typically consumes between 30% and 60% less than an untooled human loop for an equivalent or better deliverable.
TenderGraph also offers dedicated training to help bid managers, sales directors and executive sponsors optimize the quality-to-cost ratio of their AI interaction -- framing method, economy of context, dialogue structure, critical review posture. This is the other half of the transformation: an agentic tool without trained users underperforms; trained users without an agentic tool over-consume. The two together straighten the arithmetic.
Why now is the cheapest window
Three lines of upward pressure converge over the next twelve to twenty-four months.
The venture-capital subsidy has reached its peak. The first quarter of 2026 alone saw raises -- OpenAI 122 billion dollars, Anthropic 30 billion, xAI 20 billion -- that are historic. Cumulatively, OpenAI exceeds 110 billion dollars of committed capital (Stargate included), Anthropic reaches nearly 64 billion since 2021, xAI 42 billion since 2023. These raises were necessary precisely because the published balance sheets show revenue taking off faster than unit costs fall: the delta between revenue and inference cost has been covered by capital. In April 2026, funds still accept valuations at several dozen times revenue, but financial discipline is returning -- and the price trajectory already reflects it. OpenAI doubled its API pricing in moving from GPT-5 to GPT-5.5 on 23 April 2026. The subsidy on the token is no longer a durable promise.
More capable models are more expensive -- and the slope is now documented. The GPT-5 -> GPT-5.5 transition at OpenAI illustrates the mechanics: input price multiplied by two, output price multiplied by two, the launch of a GPT-5.5 Pro variant at 30 dollars input and 180 dollars output per million tokens. Anthropic proceeded differently -- the "Opus 4.5 / 4.6 / 4.7" nomenclature maintains a stable catalog price of 5 / 25, but the new tokenizer of Opus 4.7 inflates effective consumption by up to 35% on the same texts, which amounts to a silent increase. Google holds Gemini Pro slightly below, but the gradient is identical. The SOTA rises; the effective price of the SOTA rises; the price of entry-level models falls, but those models do not handle serious tenders. The gaps between classes will widen, far from narrowing.
Industrial capacity is constrained on four dimensions at once -- and none resolves with money in the short term.
Chip production. TSMC is the sole foundry capable of producing at scale Nvidia's Blackwell generations, AMD MI400, and the proprietary chips of Google (TPU v7), Amazon (Trainium 3) and Meta. The 3 nm and 2 nm etching capacity is saturated for 2026 and largely reserved for 2027. No actor, not even Microsoft or Google, can accelerate the etching cadence: the plants already run at full tilt, and building a new TSMC fab in Phoenix or Kumamoto takes four to six years. The queue to buy H200 or Blackwell is counted in months, and the hyperscalers consume the bulk of the allocations.
The cost and availability of energy. Large-scale inference became, in 2025-2026, one of the fastest-growing items of electricity consumption in the industrialized countries. Data center operators now pay for their electricity at rates that have doubled over two years in Northern Virginia, in Ireland, in Singapore. The energy bill is becoming a significant fraction of the inference cost -- and it rises with each wave of bringing high-density GPU clusters online.
The physical time of construction. All the money in the world does not produce a data center instantly. Acquiring the land, obtaining the permits, negotiating the grid connection with the network operator, building the structure, installing high voltage, cooling the racks, validating safety -- each step takes twelve to thirty-six months, not counting administrative appeals. The data centers under construction today were launched in 2023-2024; those that will cover the demand of 2027-2028 must be launched now. No financial shortcut erases this physical delay.
Connection to the electrical grid. AI-class data centers demand high-voltage connections of several hundred megawatts. Grid operators in the United States, in Ireland, in the Paris region, in Germany report queues that add eighteen to thirty-six months to projects, sometimes more. The electrical grid was not sized for this demand, and its reinforcement follows its own industrial and political timelines. Microsoft, Google and Amazon are securing ten-year nuclear contracts precisely because controllable electrical availability is becoming the limiting factor -- not the compute, the electron.
The reckoning is final: supply is already at its ceiling for several months, even several years, owing to the contracts already signed and the industrial chains already committed. Meanwhile, demand is exploding -- enterprise adoption taking off, agentics multiplying the volume of tokens consumed per active user, long contexts multiplying the cost per request, more capable models demanding more compute. The pass-through into inference prices is mechanical: when demand grows several times faster than supply, and supply cannot accelerate in the short term, prices can only rise.
Consequence: the cost of deep AI usage will rise before it falls again. The models will keep getting smarter, but at a higher price. The methodologies -- the ways of prompting, of structuring the dialogue, of setting the frame, of calibrating the operators -- are, for their part, durable assets. An organization that invests in 2026 in the cognitive discipline of its bid managers will reap, in 2027 and 2028, the benefits on more capable models. An organization that waits for "it to cost less" will wait a long time, and will arrive on a market where its competitors hold a two-year methodological lead.
The standard economic argument -- "wait for the technology to mature" -- rests, on generative AI, on an inverted reading of the curves. The tools are already mature; it is the price that is ceasing to be.
Operational consequence
For a sales director, a CIO, an executive sponsor, the decision grid holds in three lines.
The lowest visible cost -- Copilot, ChatGPT Plus -- is the highest real cost, because it funds structurally insufficient files and because it erodes trust in the tool. It is the option that produces the sentence "we tried AI, it's not convincing" when all that was tried was a throttled product on out-of-scope cases.
The intermediate visible cost -- self-rationed premium on Sonnet, Grok, Gemini Flash -- is the option of apparent sophistication without the performance. It disappoints just the same, more expensively.
The highest visible cost -- Opus on the critical acts, support models on the rest, serious architectural control and change management -- is the only one that pays off. It requires accepting that a tender handled in full consumes between 150 and 400 dollars of tokens, and that a user in the learning phase will overshoot the budget by 200 to 300 dollars per month. It also requires measuring the return at the right grain: conversion rate, margins on tenders won, cycle speed, quality of the oral defense.
No miracle solution
All the signals converge on the same conclusion. There is no shortcut. To genuinely benefit from generative AI in pre-sales, two conditions hold simultaneously: being ready to pay for inference at its fair price, and engaging the transformation now by training users to employ it better. Neither one suffices alone. And the second in fact implies the first: you do not learn to drive a Formula 1 in a city car, however state-of-the-art. A team trained on Copilot will keep reasoning Copilot -- short window, semantic RAG, deliverables that are fluent and structurally insufficient -- whatever you teach on top of it.
Being ready to pay, in practice, takes two forms. The ideal is an enterprise subscription with a SOTA vendor, under the most optimal conditions -- access to the premium model with no hidden rationing, long context available, full agentics -- and you look at the price as little as possible. This option is offered today to serious enterprises and remains accessible -- precisely because the window is subsidized. The alternative, for organizations that want to keep granular control, is to pay for the volume transiting through the API at the full rate, taking on the expense to secure technical predominance over the competitors still hesitating.
For very large enterprises, a third strategic lever is starting to emerge: owning their own data centers, their own chips, and mastering their consumption, their models, and their data. This is the path Microsoft, Google, Amazon are taking at scale for their own use, and that is progressively becoming accessible to large industrial and financial accounts as open-source models like DeepSeek V4 reach a quality comparable to the proprietary SOTA -- at the price of a considerable investment in infrastructure and in-house competencies. This strategic positioning is the subject of the next article, which examines what open source really changes, what it does not change, and the three-tier grid that follows from it for IT departments.
Whatever the lever chosen, the message holds in one line: the machine can handle the file. It is up to the human to decide which cost they accept to see, and which they prefer to keep paying hidden -- in lost margins, in failed tenders, in fluent answers that ground nothing, and in two years of accumulated lag behind the competitors who will have engaged the transformation while the window was still subsidized.
Primary sources: Anthropic, "Claude Opus 4.7 pricing and API documentation," platform.claude.com and anthropic.com, April 2026. OpenAI, "GPT-5.5 pricing and release notes," openai.com and platform.openai.com, 23 April 2026. Anthropic, "Max plan," claude.com/pricing/max, 2026. Anthropic, "Claude Mythos Preview / Project Glasswing," red.anthropic.com, 7 April 2026. PCWorld, "Anthropic confirms it's been adjusting Claude usage limits," March 2026. The Register, "Anthropic admits Claude Code quotas running out too fast," 31 March 2026. InfoWorld, "Anthropic throttles Claude subscriptions to meet capacity," 2026. MacRumors, "Claude Code Users Report Rapid Rate Limit Drain," 26 March 2026. GitHub issue anthropics/claude-code #41788, March 2026. Xaltius Academy, "The 10-Trillion Parameter Problem: Why Anthropic Locked Away Claude Mythos," 2026. Google, "Gemini API pricing," ai.google.dev, 2026. Crunchbase, "Foundational AI Startup Funding Q1 2026," news.crunchbase.com, April 2026. PitchBook / SiliconANGLE, "US venture funding surges to record $267B as OpenAI, Anthropic and xAI dominate AI deals," April 2026. Liu et al., "Lost in the Middle: How Language Models Use Long Contexts," NAACL 2024. Hoffmann et al., "Training Compute-Optimal Large Language Models" (Chinchilla scaling laws), NeurIPS 2022. Dao et al., "FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness," NeurIPS 2022. Liu et al., "Ring Attention with Blockwise Transformers for Near-Infinite Context," arXiv 2310.01889, 2023. Stanford HAI, "AI Index Report 2025," ch. 4 (Economy). Finout, "Claude Opus 4.7 Pricing -- The Real Cost Story Behind the Unchanged Price Tag" (Opus 4.7 tokenizer analysis), April 2026.