Agentics / Tech Things: Tokenmaxxing is dead, long live tokenmaxxing

theahura · article · 2026-06-27 Reader 🖥️

mixed (theoretical-explanatory/predictive with a thin practical-advisory layer) · argumentative opinion essay (contrarian) plus an appended 'Tech Things' news roundup — The dominant mode is explanatory and predictive — why deliberate token burn rose, why it died, why it returns — not a how-to. The practical payload (use generalist platforms, run cheap open models, avoid provider lock-in, don't build brittle single-task agents) is real but thin and partly self-promotional (the author runs the Nori/Agentics AI-adoption community). Snarky register; the argument is serious. Scope is strictly inference / agentic-loop token spend, NOT training or RL compute.

Structure

Key terms

• tokenmaxxing Token Throughput
  Deliberately maximizing AI inference-token consumption as a policy or strategy — spending (or encouraging spend of) large sums on tokens as an end or a forcing function, not merely as a byproduct of doing work. The author splits it two ways: temporally (v1 = subsidy-era incentive-gaming to force adoption, now dead; v2 = ROI-driven maximal spend justified by compounding correctness, being reborn) and by target ('tokens for your developers' via Claude Code + loops, vs 'tokens for your pipelines' via brittle custom agents).
• compounding correctness
  The claimed current regime in which the more tokens you spend iterating on a task, the more likely a good outcome — errors get corrected rather than accumulating. Explicitly about inference-time / agentic-loop spend (running an agent longer, restarting the prompt), not training or RL.
• compounding error
  The prior regime: small model errors and hallucinations 'take on a life of their own' over long autonomous runs and become irreversibly embedded, so running agents longer produces worse work and there is no point spending more tokens.
• loop (Ralph Wiggum loop)
  Run an agent to the end of its turn, then automatically restart the same prompt; split a heavy spec into parts and solve it over iterations with no human supervision. Old (since ~last July) but formerly hard to tune; now 'easy' because compounding correctness lets even sloppy prompting improve each iteration.
• software factory / dark factory
  A codebase that autonomously writes, reviews, tests and fixes code from a human-supplied spec with no supervision (spec in -> app out). The 'dark factory' is the lights-out extreme.

Propositions

• explicitUnlike consumers (who happily spend on things that waste time or cause pain), serious businesses do not burn money for its own sake — so large deliberate token spend demands an economic explanation.conf: high
• explicitdevelopedTokenmaxxing's first wave was deliberate, not accidental: executives purposely incentivized employees to burn tokens (e.g. Meta tying performance reviews to per-person token usage) as a blunt-force tactic to force resistant senior engineers into using AI at all.conf: high
  supportsDesigning the AI-native engineering organization · p7 — This doc's contrarian claim (burning tokens on 'useless tasks' was purposeful) is a concrete instance of 'some token spend with no direct value is justified by a second-order benefit' — here, forced adoption / organizational learning.
  extends ←Reinventing team flow within AI uncertainty · p1 — Adds a pairing-specific failure mode (individual token consumption is inconsistent to measure when two people share a session) to the general tokenmaxxing-as-gamed-incentive critique.
  extends ←Radar Trends to Watch: July 2026 · p3 — Describes the correction phase (usage-based billing, cost observability) to the 'tokenmaxxing' dynamic that doc documents as the deliberate first wave of AI cost inflation.
• explicitcorroboratedTying a metric (token usage) to performance predictably produces gaming rather than value — Meta employees ran two agents chatting all day to inflate their token counts.conf: high
  supportsThis Week in AI: Production Viability · p7 — The Meta gaming anecdote (agents chatting to inflate counts) is direct evidence that tokenmaxxing incentivizes quantity of consumption over quality of product.
  extends ←The 8 software engineering metrics AI broke · p7 — extract-tokenmaxxing documents one concrete case of metric-gaming-as-Goodhart's-Law (Meta tying performance reviews to token usage, employees running agents to inflate counts). This article generalizes that single case into a sweeping claim that AI has made Goodhart's Law unavoidable across all engineering activity metrics, not just token-usage incentives.
  supports ←The Hidden Realities of AI Adoption · p12 — Independently corroborates, from enterprise AI governance, that ungoverned incentives around AI usage produce metric-gaming ('token maxing') rather than value — the same phenomenon Meta's token-tied performance reviews illustrate in the tokenmaxxing extraction.
• explicitcorroboratedThe first wave died because rising aggregate token spend coincided with OpenAI and Anthropic going public: they throttled subscriptions and raised API prices, subsidies vanished, the incentive disappeared, cost rose, and teams rolled back unlimited-spend policies.conf: high
  supports ←Are You Getting Your Money's Worth from AI? · p9 — Tokenmaxxing's own history (subsidies vanished, API prices rose, unlimited-spend policies got rolled back once companies went public) is a concrete precedent for the upward-price scenario Shore says organizations should model.
• explicitcontestedA regime shift has occurred from 'compounding error' (small errors accumulate irreversibly over long runs, so more tokens = worse work) to 'compounding correctness' (the more tokens you spend getting a task right, the likelier a good outcome) — a claim about inference/agentic-loop spend, not training.conf: low
  refinesLLMs Corrupt Your Documents When You Delegate · p9 — theahura explicitly concedes the 'compounding error' regime was real ('up until recently... we called this compounding error') and claims a regime shift has since flipped it, bounding the 'longer autonomous sessions are disproportionately dangerous' claim in time rather than denying it. Note: his evidence for the flip is a single security-domain data point, so weight the refinement accordingly.
  contradicts ←How Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks · p3 — Tokenmaxxing's 'compounding correctness' regime-shift thesis claims more token spend generally yields better agentic-loop outcomes; this paper's own 500-task, 8-model SWE-bench study finds the opposite at the high end -- accuracy peaks at intermediate cost and saturates or declines with more spend -- directly supplying the general coding-task benchmark evidence that extract-tokenmaxxing#g1 says is missing.
• explicitdevelopedBecause compounding correctness flips the calculus, maximal token spend becomes ROI-rational, so tokenmaxxing will return long-term — this time driven by genuine value rather than incentive-gaming.conf: medium
  refines ←What is the harder human-capital problem beneath token capital? · p11 — Tokenmaxxing #6 argues maximal token spend has become ROI-rational and value-driven (compounding correctness). This essay adds a prior structural condition on that value: token capital only converts to value when wielded by frontier-movers inside a porous structure; in a rigid corporate structure the same spend 'runs in circles,' so ROI-rationality is contingent on org design, not just on the inference math.
• explicitcorroboratedAgent 'loops' (run to end of turn, restart the prompt, split a heavy spec across iterations, no supervision) are not new (the 'Ralph Wiggum loop', since ~last July) but are now easy specifically because compounding correctness lets even sloppy prompting improve each iteration.conf: medium
  supports ←11 Tips For AI Coding With Ralph Wiggum · p1 — This is the how-to for the exact technique that piece names: it treats the 'Ralph Wiggum loop' (run to end of turn, restart the prompt, split the spec across iterations, no supervision) as an established pattern and supplies the concrete scaffolding — plan/progress files, feedback gates, capped iterations — that makes the loop workable.
• explicitcorroboratedCheap open models run in loops beat expensive frontier models on cost-per-outcome (Claude ~1.1x/iteration vs GLM 5.2 ~1.05x/iteration at ~5x lower cost -> run the cheap loop 5x more and win), so open-model platforms are the real winners and frontier-lab tokenmaxxing won't survive CFO scrutiny.conf: high
  supportsThe Pulse: a new trend, smart model routing · p5 — The cost-per-iteration case for running cheap open models in loops reinforces the finding that open models are sufficient for a large (~60%) and growing share of coding work by token spend.
  refines ←The Almedra Token-Efficiency Benchmark, Updated For GLM 5.2 · p5 — extract-tokenmaxxing#p8 claims GLM 5.2 beats Claude on cost-per-outcome (~5x cheaper) holding the task constant. This benchmark holds the model (GLM 5.2) constant instead and finds harness choice alone produces a comparable-magnitude (~5.4x) cost spread among coding agents — refining the model-level cost claim by showing harness design is at least as significant a confound as model choice.
  supports ←Downstream #1: Where Do Our Tokens Go? · p6 — Uber's own workflow migration to an open-weight model with cost falling and quality holding is a concrete corroborating case for the open-vs-frontier cost-per-outcome finding.
  refines ←The Pulse: Interesting AI coding stats from Cursor · p6 — tokenmaxxing#p8 claims cheap models beat expensive frontier models on cost-per-outcome. Cursor's cost-per-line-accepted data qualifies this: on the outcome-adjusted metric the expensive models' disadvantage shrinks (Opus 4.7 matches GPT-5.5; Composer drops from ~10x cheaper per request to only ~5x cheaper per accepted line), because higher acceptance offsets higher price. Cheap still wins, but by much less than raw request price implies.
• explicit'Pipeline' tokenmaxxing — hand-writing code to build brittle, non-deterministic single-task 'agents' (e.g. a data-labeling agent, then a quality-checker agent on top) — is a bad investment: more expensive AND less accurate than deterministic code, and stacking checker agents just multiplies token cost.conf: high
• explicitGeneralist model platforms skinned for a specific task now outperform purpose-built single-task agents; consultants selling 'build me an agent' for ~$2m (essentially a skill file) are arbitraging buyers' ignorance that generalist agents have gotten very good.conf: high
• explicitThe 'software/dark factory' and the $1000/token/day-per-engineer target are hype (the author's own factory runs ~$600/month), but with a kernel of truth: the incentive to spend ludicrously on tokens is latent, waiting for diffusion.conf: high
• implicitcorroboratedThe value/ROI of developer-side tokenmaxxing is asserted, not demonstrated: the author quantifies cost precisely (pricing, $12.5k/attempt, $125k, $600/month, $1000/day) but hedges the benefit ('ostensibly this is a good use of money') and never quantifies the productivity gain.conf: high
  supports ←How People Are Really Using AI in 2026 · p6 — Independent real-world corroboration of tokenmaxxing's gap that AI's value/ROI is asserted more than demonstrated: HBR reports explicitly quantified ROI is rare in the data and business growth is mostly claimed anecdotally.
• implicit'More tokens = better output' is precisely the narrative that benefits token-sellers and AI-adoption consultants — including the author's own Nori/Agentics community ('reach out, we can help!') — yet the essay never flags this incentive alignment.conf: high
• implicit'No diminishing returns across the token budgets tested' is treated as evidence of open-ended returns, but it only rules out a plateau within the tested range and within one domain (security exploit discovery); the generalization to unbounded coding-loop spend is unwarranted.conf: high
• implicitTokenmaxxing is fundamentally an ROI story, and the same CFO cost-scrutiny that kills expensive frontier-lab token burn is what redirects spend toward cheap open models — cost discipline does not end token maximization, it relocates it.conf: high

Arguments

Narratives

• Meta tied performance evaluations to per-person token usage; employees gamed it by having two agents talk to each other all day to inflate token counts.
  → Canonical example of tokenmaxxing-as-incentive-gaming; grounds the claim that it was deliberate and that a consumption metric predictably gets gamed (Goodhart).
• AISI budgeted 100M tokens per Mythos attempt (~$12,500 each, ~$125k for ten runs); none of the models showed diminishing returns across the tested budgets — security reduces to a 'proof of work' where you win by paying more, not by being clever.
  → The single quantified data point for 'more tokens = better'; carries the entire 'compounding correctness' claim — but only in the security-exploit-discovery domain.
• Boris Cherny, creator of Claude Code, said 'loops' on stage and everyone freaked out — though loops (the 'Ralph Wiggum loop') have existed since ~last July.
  → Illustrates that the rebirth vehicle is not new capability but newly-easy execution; supports the 'loops are the operational face of tokenmaxxing v2' point.
• If Claude gives 1.1x improvement per iteration and GLM 5.2 gives 1.05x at ~5x less cost, you run the cheaper loop 5x more and end up better; GLM 5.2 (~$1.4/$4 per M) undercuts Opus 4.X ($5/$25) and even beats Haiku 4.5.
  → Concrete cost-per-outcome economics that make open models 'the real winners'; the essay's most data-anchored argument.
• A buyer asks a consultant to 'build me an agent'; the consultant essentially writes a skill file and charges $2m.
  → Illustrates market arbitrage / buyer ignorance in pipeline tokenmaxxing — value extracted from not knowing generalist agents already suffice.
• StrongDM's software factory pushes engineers toward spending $1000 in tokens per day; the author's own factory spends ~$600/month.
  → Hype-vs-reality anchor — the author polices an inflated claim with his own real number while conceding a 'kernel of truth.' Exactly the rigor a skeptical reader values.
• Executives saw custom-agent building, decided 'oh, this is just a more flexible Zapier workflow,' and demanded non-deterministic 'agentic' pipelines for one-off tasks (e.g. 'I need data labeling, so I'll build a data-labeling agent') instead of deterministic code.
  → Names the category error behind pipeline tokenmaxxing — using probabilistic agents where deterministic automation is cheaper and more correct.
• Roundup: GPT-5.6 (Sol/Terra/Luna) and Mythos are being released only via opaque government-gated 'trusted partner' processes; Cerebras serves ~750 tok/s.
  → Context notes — government picking winners/losers, and the aside that very fast inference could revert async agent operation back to synchronous use. Loosely coupled to the main argument.

Questions addressed

• ✅ Was tokenmaxxing an accident by dumb executives, or deliberate?
  Deliberate — a purposeful blunt-force change-management tactic to force AI adoption past resistant senior staff.
• ✅ Is tokenmaxxing dead?
  The first (subsidy-era, incentive-gaming) wave is dead; a second, ROI-driven wave is coming — 'dead, but we haven't seen the last of it.'
• ✅ What changed to make maximal token spend worthwhile again?
  A shift from 'compounding error' to 'compounding correctness' plus easy loops and cheap open models — though WHY the shift happened (the technical mechanism) is left unexplained.
• ✅ Who captures the value of the tokenmaxxing rebirth?
  Open-model platforms win on cost-per-outcome; the frontier labs (whose price hikes caused the death) lose under CFO scrutiny.
• ⭕ Does developer-side tokenmaxxing actually deliver net-positive ROI (value > now-unsubsidized cost)?
  Never quantified — asserted as 'ostensibly' good. Cost is measured precisely; value is only adjectival.
• ⭕ Is 'compounding correctness' an established general fact or a vendor/consultant-convenient narrative the author has a stake in?
  Not examined; the author sells AI adoption and never discloses that the pro-spend thesis benefits him.
• ⭕ If inference gets very fast (Cerebras ~750 tok/s), do we revert from async agents to synchronous operation?
  Raised as an open possibility in the roundup, with a demo pointer (chatjimmy.ai), but not argued.

Gaps

• incompleteNo general evidence for 'compounding correctness.' The single data point (AISI/Mythos: no diminishing returns across tested budgets) is narrow security exploit-discovery, where raw compute plausibly substitutes for correctness in a way it may not for open-ended coding loops. The regime-shift claim ('spending more tokens generally results in better results') is generalized from that one benchmark and from vibes.
  Read next: Look for benchmark data on whether iterative agent loops monotonically improve on general coding/knowledge tasks versus plateau or degrade; read the test-time / inference-time compute scaling literature (o-series scaling curves) and any controlled loop-vs-single-shot comparisons.
  closed byHow Do AI Agents Spend Your Money? Analyzing and Predicting Token Consumption in Agentic Coding Tasks — This paper supplies general coding-task benchmark data on the accuracy-vs-cost relationship the gap calls for -- across 500 SWE-bench-Verified problems and 8 frontier models, accuracy peaks at intermediate cost and saturates/declines at the high end, countering an unqualified 'more tokens = better' generalization. The match is not exact: the gap's 'compounding correctness' claim is about iterative agent loops (restarting/re-prompting across iterations), while this paper's cost-quartile comparison is across four independent repeated runs of the same task, not accumulating loop iterations -- so this closes the 'does more spend reliably improve outcomes on general coding tasks' half of the gap without directly testing the loop-iteration mechanism itself.
• incompleteCost is quantified, value is asserted. Every number in the piece is a cost (token prices, $12.5k/attempt, $600/month, $1000/day); the productivity or quality VALUE that would justify maximal spend is never measured, only hedged ('ostensibly'). The essay therefore cannot tell marginal from transformative benefit.
  Read next: Read HBR's 'How People Are Really Using AI in 2026' / the 'workslop'/'thinkslop' findings (marginal-not-game-changing benefits) and controlled productivity studies; ask whether the dev-loop value curve is transformative or merely marginal against unsubsidized cost.
• incompleteNo mechanism for the regime shift. The author asserts 'That's no longer true. We've entered a different regime' but never explains what technically flipped compounding error into compounding correctness — better base models? harness-level self-verification? evals in the loop? RL-from-verifiable-rewards? The rebirth's causal engine is a black box.
  Read next: Read on agent self-verification and eval-in-the-loop harness design, RLVR / verifier-based training, and what actually reduces error accumulation in long runs; distinguish model-capability gains from harness/scaffolding gains.
• incompleteVendor-vs-user value and the author's own incentive alignment are unexamined. 'More tokens = better' is exactly the narrative that benefits whoever sells tokens or AI-adoption services — and the author runs such a service, soliciting business throughout ('reach out, we can help!'). Who bears the cost (users, post-subsidy) versus who captures the value (vendors, consultants) is never separated, and the conflict is undisclosed.
  Read next: Read critically for supply-push vs demand-pull: who profits from maximal token consumption; compare against sources arguing 'most effective users are not the heaviest users'; watch for the token-seller framing in any pro-spend argument.
• incompleteThe two-kinds distinction is not integrated with the rebirth. The author says compounding correctness rescues developer loops but that pipeline agents are waste; he never reconciles whether compounding correctness would also rescue pipelines, or whether the generalist-platform shift moots custom pipelines regardless. The two 'this too shall pass' stories run in parallel without a unifying account.
  Read next: Trace whether reliability gains (compounding correctness) and consolidation gains (generalist platforms replacing bespoke agents) are the same phenomenon or two; read on why deterministic pipelines still beat agents for fixed high-accuracy tasks.

Closes gaps elsewhere

• Guest post: AI Inference Is Breaking Unit Economics · g1 — That gap noted the unit-economics piece treated inference cost purely as a supply-side engineering problem and ignored demand-side strategies. This doc supplies the demand side: subsidy withdrawal plus subscription throttling and API repricing as demand-shaping pricing levers, and open-model substitution / cost-per-iteration arbitrage (run a 5x-cheaper model 5x more) as ways to reduce cost per unit of user value rather than cost per inference.
• Baseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models · g2 — That gap asked whether there are diminishing returns to intelligence insertion or cost ceilings, versus infinitely elastic demand. This doc adds concrete discussion where there was none — a diminishing-returns data point (AISI's Mythos runs showed none across tested token budgets) and a named cost ceiling (CFO scrutiny driving substitution to cheap open models). Partial only: the data point is security-specific and actually reinforces the elastic-demand assumption rather than settling whether ceilings ultimately exist.

Related documents

• LLMs Corrupt Your Documents When You Delegate
• Designing the AI-native engineering organization
• Developer Newsletter: higher usage limits, dreaming for Claude Managed Agents, and Claude Security
  Agentic Loop Long-Running & Managed Agents Usage Limits & Rationing
• What is “loop engineering?”
  Agentic Loop Long-Running & Managed Agents Token Throughput
• The Agentic P&L: Beyond the Empire of Headcount
  AI Unit Economics Enterprise AI Adoption & Readiness
• AI Costs Are Surging and the Cheap Model Fix Might Not Last
  AI Unit Economics Open-Weight Models
• AI Root Cause Analysis Shifts from Model Reasoning to Context Engineering
  Agentic Loop Open-Weight Models
• All of AI's New Models and Tools
  Long-Running & Managed Agents Open-Weight Models
• Presentation: A Few Predicted Talks From QConAI 2030
  AI Unit Economics Token Throughput
• Baseten CEO Tuhin Srivastava on the AI Inference Crunch, Custom Models
  AI Unit Economics Open-Weight Models