# AI Coding Agent Token Costs

> Current per-token prices for Claude, GPT-4.1, and Gemini 2.5 Flash, the output premium, prompt-caching discounts, and real-world session costs — all sourced from primary documentation and verified reporting.

Source: https://tokenade.net/en/stats/ai-coding-agent-token-costs · Updated: 2026-08-04

Running an AI coding agent in 2026 is a billing event driven by a small set of numbers: the per-million-token rates for the model you pick, the ratio between output and input prices, the caching discount on repeated context, and — crucially — how many tokens an agentic session actually burns. This page collects those numbers from primary sources so they are in one place.

All prices are quoted as of June 2026. Model versions, context windows, and rates change frequently; re-verify at the primary source before quoting commercially.

## Key Takeaways

<div className="my-8 p-6 bg-amber-50 border-l-4 border-amber-500 rounded-r-lg">
  <h3 className="text-lg font-bold text-amber-800 mb-4">Key numbers at a glance</h3>
  <ul className="space-y-3">
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span><strong>Claude Sonnet 5</strong> costs $2/MTok input and $10/MTok output — the most common coding default. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>Output tokens cost <strong>≈5× more</strong> than input tokens across all Claude model tiers. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>Prompt-cache reads cost <strong>10% of fresh input</strong> on Anthropic; cache writes carry a 1.25× premium for 5-min TTL or 2× for 1-hour TTL. <a href="#source-2" className="text-blue-600 hover:underline">[2]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>Agentic coding tasks consume <strong>up to 1,000× more tokens</strong> than standard code-chat — a figure reported across multiple industry sources as real-world usage patterns have become clear. <a href="#source-6" className="text-blue-600 hover:underline">[6]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>A typical 50-turn agentic session uses roughly <strong>1 million input tokens and 40,000 output tokens</strong> — a 25:1 ratio — and costs $3–$6 on Sonnet. <a href="#source-7" className="text-blue-600 hover:underline">[7]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>Uber burned its entire 2026 AI budget in four months on Claude Code: monthly spend ran <strong>$500–$2,000 per engineer</strong> at 84% agentic adoption. <a href="#source-10" className="text-blue-600 hover:underline">[10]</a></span>
    </li>
    <li className="flex items-start gap-2">
      <span className="text-amber-600 font-bold">•</span>
      <span>Re-sent context makes up <strong>≈62% of the typical agent bill</strong> — the largest single lever is what the model reads, not what it writes. <a href="#source-8" className="text-blue-600 hover:underline">[8]</a></span>
    </li>
  </ul>
</div>

## How much does a million tokens cost on Claude in 2026?

Claude's three production tiers — Opus 5, Sonnet 5, and Haiku 4.5 — span a 5× price range from the top to the bottom of the stack.

**Claude Opus 5** costs $5.00 per million input tokens and $25.00 per million output tokens — the same rates as the Opus 4.8 it succeeds. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a>

**Claude Sonnet 5** costs $2.00 per million input tokens and $10.00 per million output tokens — introductory pricing through 31 August 2026, after which it moves to the standard $3.00 / $15.00. It is the most commonly used coding-agent model, balancing capability and cost. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a>

**Claude Haiku 4.5** costs $1.00 per million input tokens and $5.00 per million output tokens — used for grunt-work steps where lower latency and cost matter more than peak reasoning. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a>

<div className="my-8 overflow-x-auto">
  <table className="min-w-full border-collapse border border-gray-300">
    <thead>
      <tr className="bg-gray-100">
        <th className="border border-gray-300 px-4 py-3 text-left font-semibold">Model</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Input ($/MTok)</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Output ($/MTok)</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Output/Input ratio</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Context window</th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">Claude Opus 5 / 4.8</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$5.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$25.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">5×</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
      <tr className="bg-gray-50">
        <td className="border border-gray-300 px-4 py-2 font-medium">Claude Sonnet 5</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$2.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$10.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">5×</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">Claude Haiku 4.5</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$1.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$5.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">5×</td>
        <td className="border border-gray-300 px-4 py-2 text-center">200K</td>
      </tr>
    </tbody>
  </table>
  <p className="text-xs text-gray-500 mt-2">Source: Anthropic, Claude API Pricing docs (platform.claude.com, June 2026) <a href="#source-1" className="text-blue-600 hover:underline">[1]</a></p>
</div>

Claude's current Opus and Sonnet models carry a **1,000,000-token context window billed at a flat per-token rate across the whole window** — in Anthropic's own words, "a 900k-token request is billed at the same per-token rate as a 9k-token request". Haiku remains at 200K. There is **no long-context surcharge today at any threshold**; the premium that once applied above roughly 200K tokens is gone, and prompt-caching and batch discounts apply at standard rates across the full window.

## How does Claude pricing compare to GPT-4.1 and Gemini 2.5 Flash?

Claude is not the only option for coding agents. GPT-4.1 from OpenAI and Gemini 2.5 Flash from Google are the most commonly benchmarked alternatives.

**GPT-4.1** costs $2.00 per million input tokens and $8.00 per million output tokens, with a 4× output-to-input ratio. <a href="#source-3" className="text-blue-600 hover:underline">[3]</a> OpenAI applies automatic prompt caching on GPT-4.1 at a **75% discount** on cached input tokens — higher than Anthropic's 90% savings but different in structure: OpenAI caches automatically with no explicit API call, while Anthropic requires a `cache_control` breakpoint. <a href="#source-4" className="text-blue-600 hover:underline">[4]</a>

**Gemini 2.5 Flash** costs $0.30 per million input tokens and $2.50 per million output tokens for non-thinking output. <a href="#source-5" className="text-blue-600 hover:underline">[5]</a> Thinking tokens on Gemini 2.5 Flash are billed separately at $3.50 per million, meaning reasoning-heavy tasks cost substantially more than the headline rate suggests. <a href="#source-5" className="text-blue-600 hover:underline">[5]</a>

<div className="my-8 overflow-x-auto">
  <table className="min-w-full border-collapse border border-gray-300">
    <thead>
      <tr className="bg-gray-100">
        <th className="border border-gray-300 px-4 py-3 text-left font-semibold">Model</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Input ($/MTok)</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Output ($/MTok)</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Cache read discount</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Context window</th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">Claude Sonnet 5</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$2.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$10.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">−90% (10% of input)</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
      <tr className="bg-gray-50">
        <td className="border border-gray-300 px-4 py-2 font-medium">Claude Opus 5 / 4.8</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$5.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$25.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">−90% (10% of input)</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">GPT-4.1</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$2.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$8.00</td>
        <td className="border border-gray-300 px-4 py-2 text-center">−75% (automatic)</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
      <tr className="bg-gray-50">
        <td className="border border-gray-300 px-4 py-2 font-medium">Gemini 2.5 Flash</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$0.30</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$2.50 (non-thinking)</td>
        <td className="border border-gray-300 px-4 py-2 text-center">10% of input</td>
        <td className="border border-gray-300 px-4 py-2 text-center">1M</td>
      </tr>
    </tbody>
  </table>
  <p className="text-xs text-gray-500 mt-2">Sources: Anthropic pricing <a href="#source-1" className="text-blue-600 hover:underline">[1]</a>, OpenAI pricing <a href="#source-3" className="text-blue-600 hover:underline">[3]</a>, Google AI pricing <a href="#source-5" className="text-blue-600 hover:underline">[5]</a> (July 2026)</p>
</div>

Gemini 2.5 Flash undercuts every Claude tier dramatically on input, while GPT-4.1 now sits level with Sonnet 5 at $2. Whether a lower headline rate translates to a lower total bill depends on session structure, because the input-to-output token ratio and caching behaviour interact with price in ways that can reverse the ranking.

## Why does the output/input price ratio matter so much for agents?

Output tokens cost **5× more than input tokens** across the entire Claude model line — $25 vs $5 per million on Opus 5, $10 vs $2 on Sonnet 5, $5 vs $1 on Haiku 4.5. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a> At face value, this makes output the expensive side of the ledger.

The counter-intuitive reality for coding agents is that input is the higher-cost line item in practice. A 2026 analysis of production agentic sessions found that **input tokens drive approximately 85% of total session cost**. <a href="#source-7" className="text-blue-600 hover:underline">[7]</a> This is because the agent re-reads its full context — system prompt, tool definitions, conversation history, file contents, command output — on every single tool call. A 50-turn debugging session accumulates that context turn by turn, so by turn 30 the agent is paying to re-read 25,000–35,000 tokens of accumulated history on every request. <a href="#source-7" className="text-blue-600 hover:underline">[7]</a>

The output premium (5×) is real, but it applies to a much smaller volume. A typical 50-turn session generates roughly 40,000 output tokens and 1,000,000 input tokens — a 25:1 ratio. <a href="#source-7" className="text-blue-600 hover:underline">[7]</a>

<div className="my-6 grid grid-cols-1 md:grid-cols-2 gap-4">
  <div className="p-6 bg-gradient-to-br from-blue-50 to-blue-100 rounded-lg border border-blue-200 text-center">
    <div className="text-4xl font-bold text-blue-700">25:1</div>
    <div className="text-sm text-blue-600 mt-2">input-to-output token ratio in a typical 50-turn agentic session</div>
    <div className="text-xs text-gray-500 mt-2">Source: Vantage, 2026 <a href="#source-7" className="text-blue-600 hover:underline">[7]</a></div>
  </div>
  <div className="p-6 bg-gradient-to-br from-orange-50 to-orange-100 rounded-lg border border-orange-200 text-center">
    <div className="text-4xl font-bold text-orange-700">≈85%</div>
    <div className="text-sm text-orange-600 mt-2">of total agent cost driven by input tokens, not output</div>
    <div className="text-xs text-gray-500 mt-2">Source: Vantage, 2026 <a href="#source-7" className="text-blue-600 hover:underline">[7]</a></div>
  </div>
</div>

The practical implication: optimising what the agent *reads* on each turn is more impactful than making it write shorter responses. Retrieve only relevant context, prune conversation history aggressively, and route simpler steps to lighter models.

For further background on the [context window](https://tokenade.net/en/glossary/context-window) mechanics that drive this accumulation pattern, see also the [prompt caching](https://tokenade.net/en/glossary/prompt-caching) glossary entry.

## How much does prompt caching save?

Anthropic's prompt caching bills cache-read tokens at **10% of the base input price** — a 90% saving on repeated context. <a href="#source-2" className="text-blue-600 hover:underline">[2]</a>

The cache-write cost is **1.25× the base input price** for a 5-minute TTL, or **2× for a 1-hour TTL**. <a href="#source-2" className="text-blue-600 hover:underline">[2]</a> The breakeven is favourable: with the 5-minute cache, a single re-read pays off the write premium; with the 1-hour cache, two re-reads are enough.

For a 50-turn session on Sonnet 5 where the system prompt and tool definitions (10,000 tokens) are stable across all turns:

- Without caching: 10,000 × 50 × $2 / 1,000,000 = **$1.00** just for that static context
- With caching (1 write + 49 reads): (10,000 × 1.25 × $2 / 1,000,000) + (10,000 × 49 × $0.20 / 1,000,000) = **$0.025 + $0.098 = $0.123** — a reduction of roughly 88%

The requirement is that the cached prefix must be byte-identical across requests. Any change to the cached block — including reordering instructions — invalidates the cache. One documented failure mode: a developer left Claude Code running overnight in an automation loop; Anthropic had silently reduced the default cache TTL from one hour to five minutes, so an 800,000-token context was rebuilt from scratch on every 30-minute cycle — 48 times overnight — resulting in a $6,000 bill by morning. <a href="#source-12" className="text-blue-600 hover:underline">[12]</a>

GitHub measured the impact of systematic token optimisation in their own production agentic workflows in 2026, achieving up to a **62% reduction in token spend** by pruning unused MCP tool registrations, maintaining stable tool definitions across turns, and running daily auditor agents. <a href="#source-9" className="text-blue-600 hover:underline">[9]</a>

For the underlying mechanics of how tokens are counted, the [tokenizer](https://tokenade.net/en/glossary/tokenizer) and [token](https://tokenade.net/en/glossary/token) glossary entries provide reference definitions.

## How many tokens does a real coding session consume?

Production usage data from 2026 makes clear that agentic coding costs are not theoretical.

**At the session level**, Vantage's analysis of production teams found a typical 50-turn coding session at approximately **1 million input + 40,000 output tokens**, with indicative costs ranging from $0.60 (lighter models) to $6.00 (Opus-class), and average monthly spend of **$400–$1,500** for developers running coding agents full-time. <a href="#source-7" className="text-blue-600 hover:underline">[7]</a>

**At the organisation level**, Uber burned its entire 2026 AI budget in four months after rolling out Claude Code to approximately 5,000 engineers. Monthly costs ran **$500–$2,000 per engineer** for power users; overall adoption reached 84% agentic and 95% of engineers using AI tools monthly. <a href="#source-10" className="text-blue-600 hover:underline">[10]</a> Microsoft's response was to cancel Claude Code licences across its Experiences and Devices division and redirect engineers to GitHub Copilot CLI ($39/seat flat vs Claude Code's $20 seat + usage), with a deadline of 30 June 2026. <a href="#source-11" className="text-blue-600 hover:underline">[11]</a>

**At the extreme end**, Peter Steinberger's team ran roughly 100 Codex instances in the cloud for a full month, working on the OpenClaw project, accumulating **603 billion tokens across 7.6 million requests** — a posted bill of **$1,305,088.81**. The figures come from a screenshot in his own post; press reporting says OpenAI covered the spend, though his post doesn't state that. <a href="#source-13" className="text-blue-600 hover:underline">[13]</a>

A research paper on agent context management (arXiv:2508.21433, presented at the NeurIPS 2025 DL4Code workshop) confirmed that re-read context is the dominant cost driver and found that simple observation masking — replacing old tool outputs with placeholders — halves agent cost while matching the solve rate of more expensive LLM-summarisation approaches. <a href="#source-6" className="text-blue-600 hover:underline">[6]</a>

LeanOps audited 30 engineering teams running agentic AI in production between March and May 2026, finding that **re-sent context accounts for ≈62% of the total bill** at those organisations. <a href="#source-8" className="text-blue-600 hover:underline">[8]</a>

<div className="my-8 overflow-x-auto">
  <table className="min-w-full border-collapse border border-gray-300">
    <thead>
      <tr className="bg-gray-100">
        <th className="border border-gray-300 px-4 py-3 text-left font-semibold">Session type</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Approx. input tokens</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Approx. output tokens</th>
        <th className="border border-gray-300 px-4 py-3 text-center font-semibold">Indicative cost (Sonnet 5, no cache)</th>
      </tr>
    </thead>
    <tbody>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">Typical 50-turn agentic session</td>
        <td className="border border-gray-300 px-4 py-2 text-center">≈1,000,000</td>
        <td className="border border-gray-300 px-4 py-2 text-center">≈40,000</td>
        <td className="border border-gray-300 px-4 py-2 text-center">≈$2.40</td>
      </tr>
      <tr className="bg-gray-50">
        <td className="border border-gray-300 px-4 py-2 font-medium">Complex single debugging session</td>
        <td className="border border-gray-300 px-4 py-2 text-center">500,000+</td>
        <td className="border border-gray-300 px-4 py-2 text-center">variable</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$1.00+</td>
      </tr>
      <tr>
        <td className="border border-gray-300 px-4 py-2 font-medium">Multi-agent runaway loop (11 days)</td>
        <td className="border border-gray-300 px-4 py-2 text-center">billions</td>
        <td className="border border-gray-300 px-4 py-2 text-center">variable</td>
        <td className="border border-gray-300 px-4 py-2 text-center">$47,000 total <a href="#source-14" className="text-blue-600 hover:underline">[14]</a></td>
      </tr>
    </tbody>
  </table>
  <p className="text-xs text-gray-500 mt-2">Sources: Vantage <a href="#source-7" className="text-blue-600 hover:underline">[7]</a>, dev.to postmortem <a href="#source-14" className="text-blue-600 hover:underline">[14]</a>. Indicative costs use $3/MTok input + $15/MTok output; no caching applied.</p>
</div>

## What discounts are available beyond prompt caching?

Three additional cost-reduction mechanisms are available on Anthropic's platform as of 2026.

**Batch API (50% discount).** The Anthropic Message Batches API processes requests asynchronously within 24 hours at exactly 50% off both input and output tokens. Claude Sonnet 5 drops from $2/$10 to $1/$5 per million; Haiku 4.5 from $1/$5 to $0.50/$2.50. <a href="#source-1" className="text-blue-600 hover:underline">[1]</a> There is no quality difference between batch and synchronous responses — only timing. This is suited to offline code-analysis pipelines, nightly evaluation runs, and bulk refactoring jobs that can tolerate hours of latency.

**Combining caching and Batch API.** The two discounts stack. A batch request that also hits the cache on a stable 10,000-token system prompt pays 50% off the write premium (1.25× × 0.5 = 0.625×) and 50% off cache-read tokens (0.1× × 0.5 = 0.05× — 5% of the base input price). <a href="#source-2" className="text-blue-600 hover:underline">[2]</a>

**Model tier routing.** Because Haiku 4.5 costs half of Sonnet 5 at every tier, routing simpler agent steps — file listing, output formatting, search-equivalent reads — to the lighter model is a compound saving. The LeanOps audit found that teams achieving 50–70% cost reductions within two weeks consistently combined per-user budget caps, prompt caching, model-tier routing, and context-window pruning. <a href="#source-8" className="text-blue-600 hover:underline">[8]</a>

## What the numbers mean in aggregate

<div className="my-8 p-6 bg-white rounded-lg border border-gray-200 shadow-sm">
  <p className="text-lg font-semibold mb-6 text-center">Cost levers ranked by typical impact</p>
  <div className="space-y-4 overflow-hidden">
    <div>
      <div className="flex justify-between mb-1">
        <span className="text-sm font-medium">Reduce re-sent context (pruning, retrieval)</span>
        <span className="text-sm font-semibold">≈62% of bill addressed</span>
      </div>
      <div className="w-full bg-gray-200 rounded-full h-4 overflow-hidden">
        <div className="bg-blue-600 h-4 rounded-full" style={{width: '62%', maxWidth: '100%'}}></div>
      </div>
    </div>
    <div>
      <div className="flex justify-between mb-1">
        <span className="text-sm font-medium">Prompt caching on stable context (−90% on cached reads)</span>
        <span className="text-sm font-semibold">high leverage</span>
      </div>
      <div className="w-full bg-gray-200 rounded-full h-4 overflow-hidden">
        <div className="bg-green-600 h-4 rounded-full" style={{width: '55%', maxWidth: '100%'}}></div>
      </div>
    </div>
    <div>
      <div className="flex justify-between mb-1">
        <span className="text-sm font-medium">Model-tier routing (Haiku = 1/3 of Sonnet price)</span>
        <span className="text-sm font-semibold">medium leverage</span>
      </div>
      <div className="w-full bg-gray-200 rounded-full h-4 overflow-hidden">
        <div className="bg-purple-600 h-4 rounded-full" style={{width: '40%', maxWidth: '100%'}}></div>
      </div>
    </div>
    <div>
      <div className="flex justify-between mb-1">
        <span className="text-sm font-medium">Batch API for non-real-time workloads (−50%)</span>
        <span className="text-sm font-semibold">medium leverage</span>
      </div>
      <div className="w-full bg-gray-200 rounded-full h-4 overflow-hidden">
        <div className="bg-orange-500 h-4 rounded-full" style={{width: '35%', maxWidth: '100%'}}></div>
      </div>
    </div>
    <div>
      <div className="flex justify-between mb-1">
        <span className="text-sm font-medium">Reducing output length ("be concise")</span>
        <span className="text-sm font-semibold">low leverage</span>
      </div>
      <div className="w-full bg-gray-200 rounded-full h-4 overflow-hidden">
        <div className="bg-gray-400 h-4 rounded-full" style={{width: '12%', maxWidth: '100%'}}></div>
      </div>
    </div>
  </div>
  <p className="text-xs text-gray-500 mt-4">Sources: LeanOps audit <a href="#source-8" className="text-blue-600 hover:underline">[8]</a>, Vantage analysis <a href="#source-7" className="text-blue-600 hover:underline">[7]</a>, Anthropic pricing <a href="#source-1" className="text-blue-600 hover:underline">[1]</a> (2026). Bar widths are qualitative rankings, not a single computed metric.</p>
</div>

The figures above point to a consistent finding across multiple 2026 sources: what the model *reads* dominates costs, and reducing output length is the lowest-leverage optimisation available. The full practical playbook for [reducing AI coding agent token usage](https://tokenade.net/en/articles/reduce-ai-coding-agent-token-usage) covers these techniques in detail. [Tokenade](https://tokenade.net/en) applies them automatically — context compression, cache-stable prefixes, and model-tier routing — without requiring manual instrumentation.

## Methodology note

Per-token prices are from Anthropic's official pricing page (primary source, verified June 2026). Prompt-caching multipliers are from Anthropic's prompt caching documentation (primary source). OpenAI and Google prices are from their respective developer pricing pages. Session-volume figures are from the Vantage production analysis and the LeanOps industry audit of 30 production teams. The Uber and Microsoft figures are from The Information and Windows Central respectively. The OpenClaw figures are from Tom's Hardware and TheNextWeb, reporting on a public disclosure by Peter Steinberger. The $47,000 agent-loop and $6,000 overnight figures are from published postmortems on dev.to and MakeUseOf. The arXiv 2508.21433 findings are from a peer-reviewed paper presented at the NeurIPS 2025 DL4Code workshop. All figures should be re-verified at primary sources before citing in commercial contexts.

<div className="my-8 p-6 bg-gray-50 rounded-lg" id="sources">
  <h2 className="text-xl font-bold mb-4">Sources and references</h2>
  <ol className="space-y-3 text-sm">
    <li className="flex gap-2" id="source-1">
      <span className="font-semibold text-gray-500">[1]</span>
      <span>Anthropic. "Pricing — Claude API Docs". platform.claude.com, 2026. <a href="https://platform.claude.com/docs/en/about-claude/pricing" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-2">
      <span className="font-semibold text-gray-500">[2]</span>
      <span>Anthropic. "Prompt caching — Claude API Docs". platform.claude.com, 2026. <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-3">
      <span className="font-semibold text-gray-500">[3]</span>
      <span>OpenAI. "OpenAI API Pricing". openai.com, 2026. <a href="https://openai.com/api/pricing/" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-4">
      <span className="font-semibold text-gray-500">[4]</span>
      <span>OpenAI. "Prompt caching — OpenAI Platform Docs". platform.openai.com, 2026. <a href="https://platform.openai.com/docs/guides/prompt-caching" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-5">
      <span className="font-semibold text-gray-500">[5]</span>
      <span>Google. "Gemini Developer API pricing". ai.google.dev, 2026. <a href="https://ai.google.dev/gemini-api/docs/pricing" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-6">
      <span className="font-semibold text-gray-500">[6]</span>
      <span>Lindenbauer et al. (JetBrains Research). "The Complexity Trap: Simple Observation Masking Is as Efficient as LLM Summarization for Agent Context Management". arXiv:2508.21433, presented at NeurIPS 2025 DL4Code workshop. <a href="https://arxiv.org/abs/2508.21433" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-7">
      <span className="font-semibold text-gray-500">[7]</span>
      <span>Vantage. "The Hidden Cost Driver in Agentic Coding Sessions in 2026". vantage.sh, 2026. <a href="https://www.vantage.sh/blog/agentic-coding-costs" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-8">
      <span className="font-semibold text-gray-500">[8]</span>
      <span>LeanOps. "AI Agents Burn 50x More Tokens Than Chats". leanopstech.com, 2026. <a href="https://leanopstech.com/blog/agentic-ai-cost-runaway-token-budget-2026/" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-9">
      <span className="font-semibold text-gray-500">[9]</span>
      <span>GitHub. "Improving token efficiency in GitHub Agentic Workflows". github.blog, 2026. <a href="https://github.blog/ai-and-ml/github-copilot/improving-token-efficiency-in-github-agentic-workflows/" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-10">
      <span className="font-semibold text-gray-500">[10]</span>
      <span>The Information. "Uber CTO Shows How Claude Code Can Blow Up AI Budgets". theinformation.com, May 2026. <a href="https://www.theinformation.com/newsletters/applied-ai/uber-cto-shows-claude-code-can-blow-ai-budgets" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-11">
      <span className="font-semibold text-gray-500">[11]</span>
      <span>Windows Central. "Microsoft cancels Claude Code licenses, shifting developers to GitHub Copilot CLI". windowscentral.com, 2026. <a href="https://www.windowscentral.com/microsoft/microsoft-cancels-claude-code-licenses-shifting-developers-to-github-copilot-cli-a-move-likely-driven-by-financial-motives" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-12">
      <span className="font-semibold text-gray-500">[12]</span>
      <span>MakeUseOf. "Someone left Claude Code running overnight, and it cost $6,000". makeuseof.com, 2026. <a href="https://www.makeuseof.com/someone-left-claude-code-running-overnight-and-it-cost-6000/" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-13">
      <span className="font-semibold text-gray-500">[13]</span>
      <span>Tom's Hardware. "OpenClaw creator burns through $1.3 million in OpenAI API tokens in a single month". tomshardware.com, 2026. <a href="https://www.tomshardware.com/tech-industry/artificial-intelligence/openclaw-creator-burns-through-1-3-million-in-openai-api-tokens-in-a-single-month" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
    <li className="flex gap-2" id="source-14">
      <span className="font-semibold text-gray-500">[14]</span>
      <span>Anhaia, Gabriel. "The Agent That Spent $47K on Itself: An Autonomous-Loop Postmortem". dev.to, 2026. <a href="https://dev.to/gabrielanhaia/the-agent-that-spent-47k-on-itself-an-autonomous-loop-postmortem-3313" className="text-blue-600 hover:underline ml-1" target="_blank" rel="noopener noreferrer">Link ↗</a></span>
    </li>
  </ol>
</div>