Sometime in April 2026, Meta circulated an internal leaderboard called “Claudeonomics.” It ranked 85,000 employees by AI token consumption, with top performers averaging 281 billion tokens per month. Burning more tokens was treated as a proxy for productivity. The name was not ironic.
By late June 2026, Lindy CEO Flo Crivello had switched 100% of his company’s API traffic from Claude to DeepSeek. His reported explanation: the cost curve “crashed to the ground.” Vercel’s AI Gateway Production Index confirmed the pattern at scale — DeepSeek’s token share jumped from under 1% in April to 17% in May. Its spend share stayed near 1%.
Something inverted. Understanding what — and what it means for your stack — is the point of this article.
What Tokenmaxxing Was
Tokenmaxxing treated AI token consumption as a performance metric. The logic had surface appeal: if AI tools make developers more productive, developers using more AI tokens should be more productive. Encourage more usage.
The problems were predictable in retrospect.
Amazon’s case is the clearest failure mode. Once token usage became a performance signal, Amazon developers ran meaningless tasks to inflate their numbers. Goodhart’s Law, applied to LLMs: when a measure becomes a target, it stops being a measure.
Uber is the cost version. The company exhausted its entire 2026 AI coding budget within four months and was forced to impose $1,500-per-employee-per-month caps. The metric it had optimized — tokens consumed — had no connection to the budget it forgot to protect.
The Jellyfish study is the productivity version. Heavy AI users in that research were approximately 2× more productive than non-users. They also consumed 10× the tokens. The return on additional spend above a baseline was nearly zero. Separately, high-AI-adoption engineering environments showed bugs up 54% and code churn up 861% — unconstrained generation created review debt that erased the apparent speedups.
The Market Signal: Vercel AI Gateway, May 2026
Vercel publishes a monthly AI Gateway Production Index tracking real API traffic across its infrastructure. The May 2026 edition is worth reading in full; the highlights relevant to this article:
DeepSeek’s emergence
DeepSeek V4 Flash launched at $0.14 input / $0.28 output per million tokens — roughly 20–50× cheaper than comparable Anthropic models. The market responded immediately: token share went from under 1% to 17% in a single month. Spend share stayed near 1% because the price differential is that large.
Anthropic held its premium position — in spending, not volume
Anthropic’s token share rose from 26% to 32%. Its spend share rose from 61% to 65%. This divergence — tokens growing slower than spend — means Anthropic is concentrating into higher-value, higher-cost use cases while lower-stakes workloads route elsewhere. Its strongest position is in high-stakes back-office and coding agent work, where it holds 70–80% spend share.
Price sensitivity exposed Gemini’s miscalculation
When Google launched Gemini 3.5 Flash at a higher price than version 3.0, adoption stalled. By month-end, 3.5 Flash held only 7% of the Flash family’s tokens while 3.0 held 90%. Compare this to Gemini 3.1 Pro, which launched at a competitive price in February and achieved 30% adoption immediately. The market is now price-elastic in ways it was not six months ago.
Total tokens grew 20% month-over-month; total spend grew 43%
The efficiency narrative does not mean AI usage is declining. Spend is growing faster than volume because the mix is shifting toward premium tasks. The market is not spending less — it is spending less per task and more on the right tasks.
What Changed for Enterprise Buyers
The June 26 CNBC report captured the inflection point. Companies that pushed developers to maximize AI usage throughout 2025 and early 2026 are now asking for ROI analysis before renewing or expanding contracts. The phrase their procurement teams are using: “cost per task,” not “cost per token.”
The Lindy switch crystallizes this. Flo Crivello did not switch because Claude is worse — his team built on Claude because it was the best model for their use case. He switched because DeepSeek V4 Flash was good enough for enough of those use cases that the cost gap became indefensible to a board. A 20–50× price difference is not a line item negotiation. It is a rearchitecting decision.
This dynamic is playing out across enterprise contracts. Both OpenAI and Anthropic filed confidential S-1s with the SEC in early June (OpenAI on June 8, Anthropic on June 1), with Anthropic at a $965 billion post-money valuation. Their investors need growth rates that justify these valuations. But the customers most likely to drive that growth are now the ones most focused on cost discipline. The tension is not going away.
The Routing Solution
The production answer to the tokenmaxxing reckoning is heterogeneous model routing: default to cheaper models, escalate to frontier only when the task requires it.
RouteLLM (ICLR 2025) demonstrated what is possible. A router trained on preference data cut benchmark cost by more than 85% while preserving approximately 95% of flagship model quality. The mechanism: route easy queries to a cheap model, route failing cases to GPT-5.5 or Claude Sonnet. Most queries are easy.
This is not a theoretical exercise. The Vercel data shows this pattern running in production. Builders who have already implemented tiered routing are seeing the spend share stay at Anthropic (for the queries that need it) while token volume shifts toward DeepSeek (for the queries that do not).
The hard part is knowing which queries are which.
Builder Checklist: The Efficiency Era
1. Audit your model assignment by task type
Which tasks in your system genuinely require a frontier model? Which are pattern matching, summarization, or simple classification? If you cannot answer this, you cannot route intelligently. Map your task types before touching your stack.
2. Replace token-count metrics with cost-per-task metrics
If you are measuring success in tokens per day, you are measuring the wrong thing. Track cost-per-merged-PR, cost-per-completed-ticket, cost-per-resolved-support-query. These connect to the value you are generating. Tokens do not.
3. Implement explicit budget constraints in agents
Agents without stop conditions will run until they hit a wall you did not intend. Add explicit token budgets per agent run, with graceful degradation rather than silent overrun. The Uber case is a team-level version of what happens to every unbudgeted agent at production scale.
4. Evaluate DeepSeek V4 Flash for your non-critical paths
$0.14 input / $0.28 output per million tokens. If you have not benchmarked this against your specific workload, you do not know what you are leaving on the table. Do not switch on assumption — benchmark on your actual queries. But do the benchmark.
5. Treat context engineering as infrastructure, not optimization
The Corti analysis identified four practices that distinguish token-disciplined teams from tokenmaxxing teams: context compaction (summarizing prior turns rather than including them wholesale), structured note-taking (agents write to persistent state rather than re-reasoning from scratch), sub-agent isolation (each agent sees only what it needs), and just-in-time retrieval (fetch context at query time rather than loading everything at session start). These are architectural decisions, not prompting tricks.
6. Instrument quality alongside cost
The Jellyfish finding — 2× productivity, 10× tokens — was only visible because the team measured both. If you instrument cost without measuring output quality, you cannot distinguish “cheap and good” from “cheap and quietly broken.” A code churn metric (lines deleted / lines merged) is a low-cost quality proxy; bugs-per-commit is another. Pick at least one and track it alongside your spend.
The Questions to Ask Before Your Next Model Renewal
At your current usage, what percentage of queries could run at or below DeepSeek V4 Flash quality with acceptable outcomes?
What is the cost-per-completed-task for your highest-volume agent workflow today? What would it be with a tiered routing layer?
Do any of your agent runs lack explicit token budgets? What is the worst-case spend per run if an agent loops?
Is your team measuring AI productivity in tokens, completions, or outcomes? Which of those is actually correlated with your business metrics?
The tokenmaxxing era was not malicious — it was an honest attempt to measure something hard to measure. The measure was wrong. The efficiency era is about finding the right measures. That work is not optional.
Data in this article draws from Vercel’s AI Gateway Production Index (May 2026), the Corti analysis of enterprise AI engineering practices (June 2026), and public reporting from CNBC (June 26), The Neuron, and Exadel. The Jellyfish study is cited as reported in secondary analysis; the original research is from Jellyfish’s internal engineering dataset.