Verse VI · Overkill

Overkill

Claude Code and Codex are the two coding agents most engineers reach for in 2026, and neither one runs on a coding model. Both hook into the same trillion-parameter-class, AGI-chasing systems this manifesto has argued against since its first page, sold to a developer typing shell commands the same way they are sold to someone asking for a poem. Call it what it actually is: a general model wearing a coding tool as a costume, and the costume is expensive.

The tools are not what they are sold as

Claude Code runs on the same Opus, Sonnet, and Haiku models available in ordinary Claude chat, all sharing the same underlying architecture Anthropic ships to everyone else. Claude Code is a harness around that model, not a different model underneath it. OpenAI's Codex CLI runs on GPT-5-Codex, and GPT-5-Codex is, by OpenAI's own description, a version of GPT-5 optimized for agentic coding, a fine-tuned checkpoint of the frontier model rather than a distinct smaller architecture trained for code alone.^[1] OpenAI's own naming history makes the direction of travel explicit. Earlier Codex releases carried version numbers tying them to a specific GPT generation. The newest ones dropped that numbering, because the frontier general-purpose model had absorbed enough coding capability that a separate specialist tier stopped making sense to maintain as its own product. The industry is moving away from smaller, coding-focused models, not toward them, consolidating code into the same giant model that also writes marketing copy and argues philosophy, and billing every one of those uses at the same trillion-parameter rate.

What a coding-only model would actually buy

Definitions already established that inference cost tracks active parameters, not total capability. Evidence already showed a real, shipping model, DeepSeek V3, doing serious work at 37 billion active parameters against GPT-4's roughly 220 billion, a six-fold gap in the compute each query actually needs.

A coding assistant built on something in that range, rather than on a full frontier model, would need roughly a sixth of the compute per query, which means roughly a sixth of the GPUs for the same request volume, using this manifesto's own GB200 NVL72 figures from Costs. That is the hardware side of the assumption. The reliability side follows from it. Claude's own status history in 2026 counts 118 reported outages since January, and a cluster of ten separate disruptions inside a single twelve-day span in June alone.^[2] The worst of that run, on June 2, traced back to Claude Code's own sub-agent system: a bug let sub-agents spawn child sub-agents in a loop that never terminated, draining token quotas in minutes and taking claude.ai, the API, the developer console, and Claude Code down together for close to six hours.^[3] Anthropic has not published a post-incident engineering report for that outage or the ones that followed it in the same month. A smaller, purpose-built coding model, serving one workload instead of sharing capacity with every other use of the same giant model fleet, is a simpler system with fewer places for that kind of failure to start, and simpler systems are usually the ones that stay up.

Whether small models can actually do coding work is not the open question here. Chinese open-weight coding models already price the difference into the market today. DeepSeek V4 Flash runs fourteen cents per million input tokens.^[5] Claude Opus 4.8, the model underneath Claude Code, runs five dollars, and GPT-5.5, the model family underneath Codex, runs the same five dollars.^[5] This chapter's assumption describes pricing that already exists, for models already doing real coding work, that neither Claude Code nor Codex has adopted.

The subsidy that cannot last

The industry average price per million tokens fell from roughly ten dollars to roughly two dollars and fifty cents in a single year, and that collapse was not efficiency alone. OpenAI, Anthropic, Google, and Meta are, by their own investors' accounts, pricing inference below cost to capture market share, funded by venture capital and hyperscaler cross-subsidy rather than by revenue that covers what the compute actually costs.

OpenAI burned roughly 9 billion dollars in 2025, is projected to burn 17 billion in 2026, and is projected to burn 57 billion in 2027, a trajectory climbing faster than its own revenue.^[4]

Anthropic's numbers tell the same story from a stronger starting position. Annualized revenue reached roughly 30 billion dollars by April 2026, and the company still spent 2 dollars and 16 cents for every dollar that revenue brought in.^[4] A subsidy this size, run by two of the best-funded companies in the world, reads less like a permanent feature of the market and more like a bet that usage locks in before the bill comes due. When investors eventually demand the unit economics close, on Claude Code and Codex specifically as much as on chatbots, the honest way to close them is to stop running every coding query through a trillion-parameter model billed as if it might also write the next great novel. The dishonest way is to just raise the price and hope developers do not notice they are paying frontier rates for autocomplete. Chinese open-weight coding models, already priced at a fraction of Claude and GPT rates and already competitive on the benchmarks Evidence cited, are sitting there as the alternative the moment that price rises. A tool that gets more expensive while a cheaper one gets no worse is not a tool that keeps its market by default.

I wrote a meaningful share of this manifesto's prose using Claude Code, running Claude Sonnet 5, and I find that more amusing than reassuring. A tool built and marketed for writing and running code spent real compute drafting paragraphs of political argument about server costs and Ford's rehiring numbers, work with nothing to do with a compiler. The writing help itself was useful, no complaint there. What it proves is this chapter's whole point, in miniature. A model that drafts prose as comfortably as it drafts code was never specialized for either one. It is a generalist doing both jobs at a generalist's price, and I paid that price to write a manifesto arguing against paying it.

None of this argues that Claude Code or Codex are bad tools. Both are capable, and this manifesto has used one of them to help write itself. It argues that neither tool is built the way its own economics call for. A coding assistant priced and sized for coding would cost its maker less to run, fail less often, and charge its users less than either product does today. The frontier labs are not building that tool because the trillion- parameter model already exists and selling access to it is easier than building something smaller on purpose. That choice is a subsidy, not a strategy, and subsidies end.

Sources

1. OpenAI. "Building more with GPT-5.1-Codex-Max." 2026. openai.com. OpenAI's own description of Codex models as optimized variants of the GPT-5 frontier model, not a separate architecture.
2. TechTimes. "Claude Outage: Tenth Disruption in 12 Days Exposes Anthropic Infrastructure Strain." June 2026. techtimes.com. The June 2026 outage cluster and the 118-incidents-since-January figure.
3. ninetwothree. "The Slot Machine That Codes." June 2026. ninetwothree.co. The Claude Code sub-agent infinite-loop bug behind the June 2 outage.
4. AI After Hours. "OpenAI vs Anthropic: The $121 Billion Question." 2026. aiafterhours.substack.com. OpenAI and Anthropic's 2026 revenue, burn, and cost-per-revenue-dollar figures cited above.
5. Morph. "Best AI Model for Coding (June 2026): 12 Models Ranked by SWE-bench Pro Score and Cost per Task." June 2026. morphllm.com. The per-token pricing for DeepSeek V4 Flash, Claude Opus 4.8, and GPT-5.5, already cited in Evidence.