A first-principles counterargument to Nate B. Jones' token optimization framework — with hard pricing data from 2023-2026 showing frontier model prices are falling, subscriptions invert the economics, and the human is always more expensive than the model.
A response to Nate B. Jones — "Your Claude Limit Burns In 90 Minutes Because Of One ChatGPT Habit"
By Brian Edwards — April 2, 2026
I am willing to look dumb in order to state the ugly truth about how to actually use AI effectively in 2026. This essay is my response to Nate B. Jones' video and Substack post on token efficiency. I respect Nate. I watch his content. I think he is one of the best voices in the AI strategy space. But on this topic, he is wrong — not because his individual tips are bad, but because the framework he builds around them rests on axioms that do not survive contact with real data or first-principles reasoning. And some of what he is implying without stating will actively hurt the people who follow it.
Every argument has axioms — foundational assumptions the speaker treats as self-evident and therefore never examines. Nate's token efficiency framework depends on seven of them. None are self-evident. Most are wrong.
Axiom 1: API pay-per-token is the dominant pricing model. Nate's entire cost analysis — the $8-10 session, the $40-50 weekly burn, the team cost projections — assumes you are paying per million tokens at API rates. He never once mentions subscriptions. Claude Pro is $20 a month. Claude Max is $100 or $200 a month. On Max, the marginal cost of your next token is zero. His 8-10x cost reduction is a reduction from $10 to $1 on the API. On a subscription, it is a reduction from $0 to $0.
Axiom 2: Human time is free. Nate tells you to convert PDFs to markdown before feeding them to Claude. He tells you to split research across three separate models and three separate conversations, then manually synthesize the results. He tells you to prune your system prompt line by line every two weeks. He tells you to audit your plugins. He tells you to scope every agent's context by hand. He never once asks what your time costs. A senior developer in the United States bills at $150 to $250 an hour. The 20 minutes you spend pre-processing a document to save $0.47 in tokens just cost your employer $60 in salary. The hour you spend pruning a system prompt saved maybe $2 in cached tokens and cost $200 in lost engineering output. This math does not work.
Axiom 3: Frontier model prices are going up. This is the emotional engine of the video. The fear. "Mythos will be 10x Opus pricing." "$50 in, $250 out." "Your mistakes scale with the price of intelligence." There is exactly zero official data supporting this. Here is the actual data: Claude 3 Opus launched in March 2024 at $15 input and $75 output per million tokens. Claude Opus 4.6, the current flagship — a dramatically more capable model — costs $5 input and $25 output. That is a 67% price reduction in two years. GPT-4 launched in March 2023 at $30 input and $60 output. GPT-5.4 in 2026 costs $2.50 input and $15 output. That is a 92% reduction on input and 75% on output in three years. The Gemini 2.5 Flash-Lite model handles tasks that would have required GPT-4 in 2023 and costs $0.10 per million input tokens — a 300x reduction from GPT-4's launch price. Every single data point in the historical record shows frontier model prices falling, not rising. Nate's Mythos pricing scenario is not a prediction. It is speculation built on the word "expensive" from a Fortune article about a leaked model name.
Axiom 4: Context windows are scarce and must be rationed. In early 2023, most models had 4,000 to 8,000 tokens of context. By late 2023, we had 128K. By 2024, Gemini hit 1 million. Today, Claude Opus 4.6 has a 1 million token context window with no long-context surcharge — a 900K-token request is billed at the same rate as a 9K request. Auto-compaction lets sessions run indefinitely. Prompt caching gives you a 90% discount on stable context that re-enters the window on every turn. Extended thinking tokens are stripped from subsequent turns, so deep reasoning does not eat your context. The context scarcity that justified Nate's "start fresh every 10-15 turns" advice was real in 2024. It is not real in April 2026.
Axiom 5: Single-turn, pre-computed prompts are superior to interactive conversation. Nate says you should have two modes: gathering information and getting work done. He says these should never mix. He says your prompt should be "so clear that the AI needs to do nothing else and it just goes and gets the work done and comes back." This is the prompt engineering mindset from 2023 applied to 2026 models. It was correct when models were dumb and context was expensive. It is wrong now. The best work happens in interactive conversation — not because the human is correcting the model, but because the human and model are thinking together. A developer reviewing a diff with Claude, asking "why did you choose this approach," hearing the reasoning, pushing back, arriving at a better solution — that is not "conversation sprawl." That is the entire point. The models are good enough now that the conversation is the product.
Axiom 6: Third-party tools and intermediaries add value. Nate recommends Perplexity for search because it burns "10 to 50,000 less tokens per search" than Claude's native search. He promotes his Open Brain ecosystem for markdown conversion and token auditing. He is building a product that sits between you and the frontier model. This is exactly backwards. Every intermediary you add is a company that needs to make money by sitting between you and the intelligence. Perplexity takes your query, sends it to a model (often Claude or GPT), wraps the result, and charges you for the wrapper. Cursor takes a frontier model, adds a UI, and charges you $20 a month for the privilege of not using the model directly. These companies do not have the competitive pressure that Anthropic, Google, and OpenAI have. The frontier companies are in a three-way war for dominance. That war drives prices down and capabilities up. The wrapper companies are in a margin extraction business. Use frontier companies' products directly. Use free, open-source CLI tools hosted on GitHub for everything else. That is it.
Axiom 7: Token counting is a valuable professional skill. Nate calls token management "one of the most valuable skills on the planet." He frames it as a job skill. He built a "stupid button" to audit your token usage. He wants you to instrument every agent call and track input tokens, output tokens, model mix, and cost ratio. For API-based production workloads at scale, sure — measure your costs. For everything else, this is like telling a carpenter to weigh every nail. The model is cheaper than you are. Every minute you spend counting tokens is a minute you did not spend building something. The skill that matters is knowing how to think with AI, not how to ration it.
Here is what Nate's framework completely misses, and it changes everything.
Claude Max 20x costs $200 a month. Third-party analysis of Claude Code usage data shows heavy developers consume $35 to $53 per day in API-equivalent compute. That is $1,050 to $1,590 per month of compute for $200. A 5x to 8x subsidy. Anthropic is almost certainly running this at a loss to acquire developers during their growth phase. Inference costs continue to drop, which means the subsidy improves over time even if the subscription price stays flat.
On a subscription, Nate's entire optimization framework inverts. "Use Sonnet for execution and Haiku for polish" becomes "use Opus for everything, because it is free at the margin and produces the best results." "Start fresh every 10-15 turns" becomes "use the full context window for as long as the work requires, and /clear when you switch tasks." "Audit your plugins" becomes "give the agent every tool it might need — the model decides what to invoke, and tool definitions are cached at 90% discount." "Don't mix research and execution" becomes "have an interactive conversation with the best model on Earth, because the conversation is the work."
The distinction that actually matters is not "clean session vs. sloppy session." It is "subscription vs. API." And the rule is simple: use your subscription for everything you possibly can. Interactive development, research, code review, planning, writing, debugging — all subscription. Reserve API spending for production workloads that must run without a human in the loop: webhook handlers processing incoming tickets, scheduled enrichment pipelines, in-app chat for end users. Even some of that automation can run on subscription via Claude Code's /loop command, which executes prompts on a recurring interval within the subscription allocation.
graph LR
subgraph "Subscription — $0 marginal cost"
A[Claude Code TUI] --> B[Interactive dev]
A --> C[Code review]
A --> D[Research + planning]
A --> E[Writing + docs]
A --> F["/loop scheduled tasks"]
A --> G[Agent workers in worktrees]
end
subgraph "API — pay per token"
H[Vertex AI] --> I[Webhook resolution plans]
H --> J[Batch enrichment pipeline]
K[Google AI Studio] --> L[Embedding — FREE on Tier 1]
M[GitHub Actions] --> N[Cloud sync cron]
end
style A fill:#2d6,stroke:#000,color:#000
style H fill:#d62,stroke:#000,color:#000
style K fill:#26d,stroke:#000,color:#000
style M fill:#888,stroke:#000,color:#000
Everything on the bottom of that diagram is covered by a flat monthly fee. The optimization target on the bottom is not "minimize tokens." It is "maximize value per session." Use the best model. Use the full context. Have the conversation. Think with the machine.
Nate frames token waste as the primary cost. He is wrong. The primary cost is human opportunity loss.
A developer costs $150 to $250 an hour. Claude Opus 4.6 on the API costs $5 per million input tokens. One hour of developer time costs the same as processing 30 to 50 million tokens. On a subscription, that ratio is infinite — the developer's hour costs $150 to $250, and the tokens cost zero.
When Nate tells you to manually split your work across three models and three conversations, then synthesize the results yourself, he is asking you to substitute the most expensive resource in the system (your time and attention) for the cheapest one (tokens). When he tells you to "do the thinking in advance and bring that to the table," he is asking you to do the model's job for it. When he tells you to prune your system prompt line by line, he is asking a $200-per-hour human to do work that a $5-per-million-token model could do in seconds.
The correct approach is the opposite. Give the agent everything. Have the agent do the pre-processing. Have the agent convert the PDFs, chunk the documents, scope the context, synthesize across sources. Yes, do it once — do not re-process the same document on every call. But the agent does it once, more thoroughly, more consistently, and for a fraction of the cost of a human doing it. The human's job is to direct, evaluate, and decide. Not to save the machine from having to think.
This is especially true in a CLI-first workflow. Claude Code is a terminal application. You type a task. The agent reads files, searches code, runs commands, writes code, creates commits, opens pull requests. You review and approve. The agent is doing everything. The human cost of "doing nothing" is nearly zero — you are reviewing output, not producing it. The human cost of Nate's optimization regime — manually managing context, switching between models, pruning prompts, auditing plugins — is substantial and ongoing. It is a tax on every session that produces no value except slightly lower API bills that you are not even paying because you are on a subscription.
Here is what I actually use, and it costs less than Nate's "optimized" API workflow while producing better results.
Frontier models only. Claude Opus 4.6 for development (subscription). Claude on Vertex AI for production workloads (API). Google Gemini for classification and embedding (API, with free tiers). That is three providers from two frontier companies. No wrappers. No intermediaries. No Perplexity. No Cursor. No third-party SaaS that adds a margin on top of the model I am already paying for. Anthropic, Google, and OpenAI are locked in a competition that drives prices down every quarter. DeepSeek's $0.14-per-million-token model in December 2024 forced industry-wide repricing throughout 2025. That competitive pressure benefits me directly when I buy from frontier providers. It does not benefit me when I buy from a wrapper company that pockets the savings.
A note on OpenAI specifically: I do not use their products for ethical reasons. OpenAI is pursuing a strategy of becoming too big to fail. The $110 billion funding round at a $730 billion valuation, backed by SoftBank and the Stargate project with explicit United States government endorsement. The $500 billion infrastructure commitment. The systematic courting of government relationships to ensure that public money and policy backing flow to OpenAI specifically. This is not a competitive strategy. It is a capture strategy. It is the same playbook that created "too big to fail" banks in 2008 — build so much systemic dependency that taxpayers have no choice but to backstop your losses. If you have no ethical concerns about this, fine, use their models and their Codex GitHub app for code reviews. They make good models. But I choose to fund companies competing on merit, not political capture.
CLI and GitHub for everything else. GitHub Issues are model memory — persistent, searchable, linkable context that survives across sessions and conversations. Pull requests are code review infrastructure with two automated reviewers (Claude GitHub App and Gemini Code Assist) reading every diff before a human sees it. GitHub Actions run cloud workloads on schedule. The gh CLI drives all of it from the terminal. Git worktrees give agents isolated environments for parallel work. This is free infrastructure from a platform with a competitive incentive to keep developers on it. Use it.
The orchestrator/worker pattern. One interactive Claude Code session on main orchestrates. Background agents in worktrees do implementation on feature branches. Each agent gets a GitHub issue (the spec), a worktree (isolation), and a PR target (the deliverable). The orchestrator never writes code. It plans, delegates, reviews, and merges. This is not "conversation sprawl." It is a managed software development process where the AI does the work and the human does the judgment.
Claude Opus 4.6 has a 1 million token context window. No long-context surcharge. Auto-compaction lets sessions continue indefinitely. Prompt caching reduces the cost of re-sending stable context to 10% of the standard rate. Extended thinking tokens are stripped from subsequent turns and do not consume context on re-entry.
Epoch AI benchmarks from mid-2025 show that the input length at which top models reach 80% accuracy has grown by over 250x in nine months. Claude Opus 4.6 achieves 78.3% accuracy on million-token retrieval benchmarks — the highest of any frontier model at that length. Is "context rot" real? Yes. Models still degrade on complex reasoning tasks over very long contexts. But the correct response is awareness and good hygiene — put persistent instructions in CLAUDE.md, use /compact when switching focus, /clear between unrelated tasks — not the anxious rationing Nate prescribes.
The practical experience matches the benchmarks. Developers report running 500K-token Claude Code sessions that remain coherent. Full multi-file diffs that previously required chunking at 200K now fit in a single window. Agents run for hours without losing track of what they read at the start. Is the context window something to think about and keep in the back of your mind? Yes. Is it something to fear, ration, and structure your entire workflow around minimizing? In April 2026, no. That anxiety is a holdover from 2024, and it is being actively manufactured by people selling token-counting tools.
Nate says: "The next generation of models is likely to drop in the next one to two months... They will be more expensive, a lot more expensive."
Here is what actually happened, every time a next-generation model dropped:
| Transition | Input Price Change | Output Price Change | Capability |
|---|---|---|---|
| Claude 3 Opus → Opus 4.5 | $15 → $5 (-67%) | $75 → $25 (-67%) | Dramatically better |
| Claude 3.5 Sonnet → Sonnet 4.6 | $3 → $3 (flat) | $15 → $15 (flat) | Dramatically better |
| GPT-4 → GPT-4o | $30 → $5 (-83%) | $60 → $15 (-75%) | Better |
| GPT-4o → GPT-5.4 | $5 → $2.50 (-50%) | $15 → $15 (flat) | Better |
| o3 launch → o3 repriced | $10 → $2 (-80%) | $40 → $8 (-80%) | Same model |
Every row in that table is a price decrease or flat hold with increased capability. Not one is an increase. The macro trend from 2023 to 2026 is a 6x to 15x reduction in frontier input pricing and a 3x to 4x reduction in frontier output pricing. At the budget tier, the reduction is 200x to 300x.
Will Mythos cost more than Opus 4.6 at launch? Maybe. The o1-pro model from OpenAI launched at $150/$600 — a genuine premium tier. But that did not make Opus 4.6 more expensive. It did not make Sonnet more expensive. It created a new tier above the existing ones, which then came down in price within months (o3 dropped 80% after launch). The existence of a more expensive option at the top does not raise the price of the tier you were already using. It lowers it, because the previous flagship becomes the mid-tier. This has happened with every single model generation. There is no reason to believe it will stop.
Convert raw PDFs to markdown before processing. Yes. That is just good data hygiene and it applies everywhere — do it once, do it with the model, store the result.
Measure your API costs on production workloads. Yes. If you are running an agentic pipeline that makes thousands of calls a day on the API, instrument it. Know your per-call cost. Use prompt caching on stable context. Use batch API for async work at 50% discount.
Do not load 50,000 tokens of ChatGPT plugins you never use. Yes. Audit your tools. But the answer is not "use fewer tools." The answer is "use well-scoped tools and let the model decide which ones to invoke." An MCP server with five precisely defined tools is not overhead. It is infrastructure.
These are good tactical tips. What is wrong is the strategic framework built around them: the fear-based narrative that models are getting more expensive, that tokens are scarce, that you should ration your usage, that counting tokens is a professional skill, and that the way to get more from AI is to give it less. The opposite is true. Give it more. Give it everything. Use the best model. Have the conversation. Do it on a subscription. Let the frontier companies fight over who can give you the most intelligence for the least money. That is the actual trend, and it is accelerating.
The data in this essay comes from official pricing pages (Anthropic, OpenAI, Google AI), third-party pricing trackers (TokenCost, Artificial Analysis), Epoch AI research on context windows, and Chroma's context rot research. Mythos pricing claims are speculation — no official pricing has been announced.