<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Toolbit AI</title>
    <description>The latest articles on DEV Community by Toolbit AI (toolbit_ai).</description>
    <link>https://dev.to/toolbit_ai</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Forganization%2Fprofile_image%2F15600%2F3bfeaf77-f366-4f23-90bb-c0d842f9010f.png</url>
      <title>DEV Community: Toolbit AI</title>
      <link>https://dev.to/toolbit_ai</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/toolbit_ai"/>
    <language>en</language>
    <item>
      <title>Over 40% of Pull Requests Now Start With an Agent. Review Them Like It.</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Tue, 03 Nov 2026 04:35:00 +0000</pubDate>
      <link>https://dev.to/toolbit_ai/over-40-of-pull-requests-now-start-with-an-agent-review-them-like-it-3d44</link>
      <guid>https://dev.to/toolbit_ai/over-40-of-pull-requests-now-start-with-an-agent-review-them-like-it-3d44</guid>
      <description>&lt;p&gt;GitHub's new Octoverse numbers came out on 28 October, and one line stuck with me: "As of September, developers start more than 40% of pull requests with an agent." In March that number was 17%.&lt;/p&gt;

&lt;p&gt;So if you review code for a team, a big chunk of what lands in your queue probably wasn't typed by the person whose name is on it. That doesn't make it bad code. GitHub says 90% of agent PRs were merged within two weeks in September, against about 86% for PRs with no agent involved. But it does change what's worth checking.&lt;/p&gt;

&lt;p&gt;Here's what GitHub already does for you, what it just changed, and the checks I'd add on top.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkxlicsw2n69gp6eotxlh.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fkxlicsw2n69gp6eotxlh.png" alt="Bar chart: PRs started with an agent rose from 17% in March to over 40% in September; agent PRs merged within two weeks 90% vs about 86%; 87% of agent PRs change more than one file vs 79%" width="800" height="497"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Source: GitHub Octoverse 2026.&lt;/em&gt;&lt;/p&gt;
&lt;h2&gt;
  
  
  What GitHub's coding agent already protects
&lt;/h2&gt;

&lt;p&gt;If you use Copilot's cloud agent, a few guardrails are on by default. Here's what GitHub's docs say it handles, and what's still on you:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;What GitHub does&lt;/th&gt;
&lt;th&gt;Default&lt;/th&gt;
&lt;th&gt;What's still on you&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Only people with write access can start the agent&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Who has write access in the first place&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Agent pushes to a single branch (normally &lt;code&gt;copilot/...&lt;/code&gt;) and your branch protections still apply&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;That your protections and required checks are actually set&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The person who asked for the PR &lt;strong&gt;can't approve it&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;A real second review, not a quick thumbs-up&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Actions workflows don't run on Copilot's pushes until a maintainer clicks &lt;strong&gt;Approve and run workflows&lt;/strong&gt;
&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Reading any &lt;code&gt;.github/workflows/&lt;/code&gt; change before you click&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Secret scanning, CodeQL and an advisory-database dependency check run on its code&lt;/td&gt;
&lt;td&gt;On&lt;/td&gt;
&lt;td&gt;Whether new packages are the ones you meant&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;strong&gt;Copilot approvals&lt;/strong&gt; count toward required reviews&lt;/td&gt;
&lt;td&gt;
&lt;strong&gt;Off&lt;/strong&gt; (new, path-scoped)&lt;/td&gt;
&lt;td&gt;Which paths, if any, you let it approve&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;That's a decent baseline. The second-reviewer rule alone stops the "I asked for it, I approved it, it's merged" loop.&lt;/p&gt;
&lt;h2&gt;
  
  
  The new setting to decide on this week
&lt;/h2&gt;

&lt;p&gt;On the same day, GitHub made &lt;strong&gt;Copilot approvals&lt;/strong&gt; generally available. When it's on, Copilot can approve a PR and "that approval counts toward your repository's required approvals." It's off by default, and admins pick which file paths it's allowed to approve.&lt;/p&gt;

&lt;p&gt;So now you have an odd mix: the person who asked the agent for a PR can't approve it, but Copilot's own approval can count on the paths you allow. That might be fine for docs or test fixtures. I wouldn't let it near auth, billing, migrations or &lt;code&gt;.github/workflows/&lt;/code&gt; without a person looking too. Whatever you pick, write it down so the team knows which paths a bot can sign off.&lt;/p&gt;
&lt;h2&gt;
  
  
  Five things I'd check on every agent PR
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;1. Did the tests change along with the code?&lt;/strong&gt;&lt;br&gt;
Researchers behind ImpossibleBench found agents "may delete failing tests rather than fix the underlying bug." So when a fix and its tests change in the same PR, read the test diff first:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git diff &lt;span class="nt"&gt;--stat&lt;/span&gt; origin/main...HEAD &lt;span class="nt"&gt;--&lt;/span&gt; &lt;span class="s1"&gt;'*test*'&lt;/span&gt; &lt;span class="s1"&gt;'*spec*'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;(It's a rough filter, so it'll also catch names like &lt;code&gt;latest.js&lt;/code&gt;. Good enough for a first look.)&lt;/p&gt;

&lt;p&gt;If an assertion got looser or a test disappeared, ask why before you look at anything else.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. Any new dependencies?&lt;/strong&gt;&lt;br&gt;
A USENIX Security 2025 paper found an average hallucinated-package rate of at least 5.2% for commercial models and 21.7% for open-source ones. Attackers can register those names. Check every new line in your manifest actually points at the package you think it does:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git diff origin/main...HEAD &lt;span class="nt"&gt;--&lt;/span&gt; &lt;span class="s1"&gt;'*package.json'&lt;/span&gt; &lt;span class="s1"&gt;'*requirements*.txt'&lt;/span&gt; &lt;span class="s1"&gt;'*pyproject.toml'&lt;/span&gt; &lt;span class="s1"&gt;'*go.mod'&lt;/span&gt; &lt;span class="s1"&gt;'*Cargo.toml'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;3. Anything under &lt;code&gt;.github/workflows/&lt;/code&gt;?&lt;/strong&gt;&lt;br&gt;
GitHub's own review guide calls this out. A workflow change runs with your repo's permissions, so it deserves a slower read than the feature code.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;4. Where did the task come from?&lt;/strong&gt;&lt;br&gt;
In 2025 Invariant Labs showed an agent could be hijacked "via a malicious GitHub Issue" and end up leaking private data in a PR it opened itself. If the agent worked from an issue that an outsider wrote, read the issue too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;5. Is it doing more than it was asked?&lt;/strong&gt;&lt;br&gt;
GitHub says about 87% of agent-started PRs touch more than one file, compared with 79% for PRs with no agent detected. Wide PRs hide small surprises. If the task was "fix the date parser" and the diff also renames a config key, split it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why not just trust it?
&lt;/h2&gt;

&lt;p&gt;Because the research still says "check." Veracode tested over 100 models in 2025 and 45% of the code samples failed security tests. Google's 2025 DORA report found 30% of developers have little or no trust in AI-generated code, and that AI use still has "a negative relationship with software delivery stability." None of that means agents are useless. It means review is where the value gets decided, which is pretty much what GitHub's own Octoverse post says: "Review is where developers shape what comes back."&lt;/p&gt;

&lt;p&gt;If you're looking at review tools to help with the load, Toolbit has a &lt;a href="https://www.toolbit.ai/category/code-review" rel="noopener noreferrer"&gt;page of AI code review tools&lt;/a&gt; you can compare.&lt;/p&gt;

&lt;p&gt;What's on your team's checklist for agent PRs that isn't on mine?&lt;/p&gt;




&lt;p&gt;&lt;em&gt;This article was created with the help of AI and reviewed by the author for accuracy. Disclosure: I work at Toolbit, linked above.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Sources&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;GitHub Blog, Octoverse 2026 (28 Oct 2026): &lt;a href="https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/" rel="noopener noreferrer"&gt;https://github.blog/news-insights/octoverse/octoverse-commits-pushes-reach-new-highs-in-2026-as-developers-build-with-agents/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Changelog, Copilot code review Max and approvals (28 Oct 2026): &lt;a href="https://github.blog/changelog/2026-10-28-copilot-code-review-merge-with-confidence-with-max-and-approvals" rel="noopener noreferrer"&gt;https://github.blog/changelog/2026-10-28-copilot-code-review-merge-with-confidence-with-max-and-approvals&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Docs, Copilot cloud agent risks and mitigations: &lt;a href="https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations" rel="noopener noreferrer"&gt;https://docs.github.com/en/copilot/concepts/agents/cloud-agent/risks-and-mitigations&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;GitHub Docs, Reviewing Copilot pull requests: &lt;a href="https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent/review-copilot-prs" rel="noopener noreferrer"&gt;https://docs.github.com/en/copilot/how-tos/use-copilot-agents/coding-agent/review-copilot-prs&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;ImpossibleBench, arXiv 2510.20270: &lt;a href="https://arxiv.org/abs/2510.20270" rel="noopener noreferrer"&gt;https://arxiv.org/abs/2510.20270&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Spracklen et al., package hallucinations, USENIX Security 2025: &lt;a href="https://arxiv.org/abs/2406.10279" rel="noopener noreferrer"&gt;https://arxiv.org/abs/2406.10279&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Invariant Labs, GitHub MCP exploit (26 May 2025): &lt;a href="https://invariantlabs.ai/blog/mcp-github-vulnerability" rel="noopener noreferrer"&gt;https://invariantlabs.ai/blog/mcp-github-vulnerability&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Veracode, 2025 GenAI Code Security Report: &lt;a href="https://www.veracode.com/blog/genai-code-security-report/" rel="noopener noreferrer"&gt;https://www.veracode.com/blog/genai-code-security-report/&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Google Cloud, 2025 DORA report: &lt;a href="https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report" rel="noopener noreferrer"&gt;https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>ai</category>
      <category>github</category>
      <category>codereview</category>
      <category>security</category>
    </item>
    <item>
      <title>Prompt Caching Cost Is What Decides Your AI Coding Agent Bill</title>
      <dc:creator>Nirali</dc:creator>
      <pubDate>Mon, 26 Oct 2026 09:59:08 +0000</pubDate>
      <link>https://dev.to/toolbit_ai/prompt-caching-cost-is-what-decides-your-ai-coding-agent-bill-50ad</link>
      <guid>https://dev.to/toolbit_ai/prompt-caching-cost-is-what-decides-your-ai-coding-agent-bill-50ad</guid>
      <description>&lt;p&gt;A coding agent sends your whole conversation back to the model on every single turn. So the price that decides your bill is often the smallest number in the pricing table: the prompt caching cost.&lt;/p&gt;

&lt;p&gt;Max Woolf's post &lt;a href="https://minimaxir.com/2026/10/cheap-llms/" rel="noopener noreferrer"&gt;OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?&lt;/a&gt; is getting a lot of discussion on Hacker News right now. It walks through a summer of price cuts. One line stuck with me. He writes that when you use agents for coding, cache read costs are "the vast majority (96% - 99%) of the total token expenditure." I wanted to see what that means for a real AI coding agent cost, using today's pricing pages instead of launch headlines.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why agents live on cache reads
&lt;/h2&gt;

&lt;p&gt;Every time an agent calls the model, it sends the system prompt, the tool definitions, every file it has read and every message so far. Then it adds a little new content at the end: a tool result, or your next message.&lt;/p&gt;

&lt;p&gt;Providers cache that repeated prefix. Reading it back from the cache is much cheaper than sending it fresh:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On Claude Opus 5.5, a cache hit costs 5% of the base input price.&lt;/li&gt;
&lt;li&gt;On GPT-6.1 Sol, it's also 5% ($0.10 vs $2.00).&lt;/li&gt;
&lt;li&gt;On Claude Fable 5.1, it's 2.5%.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The new tokens at the end usually get &lt;em&gt;written&lt;/em&gt; to the cache. That costs a bit more than normal input: 1.25x on both Anthropic's default 5-minute cache and OpenAI's default 30-minute cache.&lt;/p&gt;

&lt;h2&gt;
  
  
  LLM API pricing 2026: the numbers that matter
&lt;/h2&gt;

&lt;p&gt;Here is what the official pricing pages say today. Prices are USD per million tokens, standard tier.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Input&lt;/th&gt;
&lt;th&gt;Cache read&lt;/th&gt;
&lt;th&gt;Cache write (default)&lt;/th&gt;
&lt;th&gt;Output&lt;/th&gt;
&lt;th&gt;Note&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Claude Opus 5&lt;/td&gt;
&lt;td&gt;$5.00&lt;/td&gt;
&lt;td&gt;$0.50&lt;/td&gt;
&lt;td&gt;$6.25 (5 min)&lt;/td&gt;
&lt;td&gt;$25&lt;/td&gt;
&lt;td&gt;previous Opus&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Opus 5.5&lt;/td&gt;
&lt;td&gt;$4.00&lt;/td&gt;
&lt;td&gt;$0.20&lt;/td&gt;
&lt;td&gt;$5.00 (5 min)&lt;/td&gt;
&lt;td&gt;$20&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Claude Fable 5.1&lt;/td&gt;
&lt;td&gt;$10.00&lt;/td&gt;
&lt;td&gt;$0.25&lt;/td&gt;
&lt;td&gt;$12.50 (5 min)&lt;/td&gt;
&lt;td&gt;$50&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-5.6 Sol&lt;/td&gt;
&lt;td&gt;$4.00&lt;/td&gt;
&lt;td&gt;$0.40&lt;/td&gt;
&lt;td&gt;$5.00 (30 min)&lt;/td&gt;
&lt;td&gt;$20&lt;/td&gt;
&lt;td&gt;promotional price&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6.1 Sol&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$0.10&lt;/td&gt;
&lt;td&gt;$2.50 (30 min)&lt;/td&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$0.10&lt;/td&gt;
&lt;td&gt;not listed (implicit caching only)&lt;/td&gt;
&lt;td&gt;$10&lt;/td&gt;
&lt;td&gt;until 30 Nov 2026&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Checked 26 Oct 2026 on the &lt;a href="https://platform.claude.com/docs/en/about-claude/pricing" rel="noopener noreferrer"&gt;Anthropic&lt;/a&gt;, &lt;a href="https://developers.openai.com/api/docs/pricing" rel="noopener noreferrer"&gt;OpenAI&lt;/a&gt; and &lt;a href="https://ai.google.dev/gemini-api/docs/pricing" rel="noopener noreferrer"&gt;Google&lt;/a&gt; pricing pages. OpenAI prices are for prompts up to 272K input tokens.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;(I work at Toolbit, and we keep a running comparison of these prices &lt;a href="https://www.toolbit.ai/blog/opus-5-5-gpt-6-sol-luna-price-war" rel="noopener noreferrer"&gt;here&lt;/a&gt;.)&lt;/p&gt;

&lt;p&gt;Look at Opus 5 to Opus 5.5. Input and output dropped 20%. Cache reads dropped 60%, from $0.50 to $0.20. If you only read the headline, you'd plan for a 20% saving. An agent gets a lot more than that.&lt;/p&gt;

&lt;h2&gt;
  
  
  A worked example: one hour of agent work
&lt;/h2&gt;

&lt;p&gt;These are made-up but realistic assumptions for one coding session:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;160 model calls in about an hour&lt;/li&gt;
&lt;li&gt;the context averages about 100K tokens over the session (it grows, and the agent compacts it now and then), so each call re-reads about 100K cached tokens on average&lt;/li&gt;
&lt;li&gt;about 2K new tokens per call (tool output, your messages), written to the cache&lt;/li&gt;
&lt;li&gt;about 500 output tokens per call (mostly tool calls and short replies)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;That adds up to 16M cache-read tokens, 0.32M cache-write tokens and 0.08M output tokens. Cache reads are 97.6% of all tokens, right inside Woolf's range.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Model&lt;/th&gt;
&lt;th&gt;Cache reads&lt;/th&gt;
&lt;th&gt;Cache writes&lt;/th&gt;
&lt;th&gt;Output&lt;/th&gt;
&lt;th&gt;Session total&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Opus 5&lt;/td&gt;
&lt;td&gt;$8.00&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;$2.00&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$12.00&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Opus 5.5&lt;/td&gt;
&lt;td&gt;$3.20&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$6.40&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-5.6 Sol (promo)&lt;/td&gt;
&lt;td&gt;$6.40&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$9.60&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT-6.1 Sol&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$3.20&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon (until 30 Nov)&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;$0.64*&lt;/td&gt;
&lt;td&gt;$0.80&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$3.04&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gemini 4 Argon (from 1 Dec)&lt;/td&gt;
&lt;td&gt;$3.20&lt;/td&gt;
&lt;td&gt;$1.28*&lt;/td&gt;
&lt;td&gt;$1.60&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;$6.08&lt;/strong&gt;&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;em&gt;Google lists no cache-write price for Argon, so I priced the new tokens at the normal input rate. Its caching is implicit, so your hit rate isn't guaranteed.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Three things stand out.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Opus 5 to 5.5 is a 47% cut for this session, not 20%.&lt;/strong&gt; The 60% cache-read cut does most of the work. A headline-only estimate would have said $9.60. That lines up with Woolf's estimate of a 40-60% total cost cut.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Cache reads are the biggest line item, but not 96-99% of the money here.&lt;/strong&gt; They're 97.6% of the tokens. In dollars they're 50-67% of the bill, because writes and output cost far more per token. If your agent writes long answers, output matters more than you'd think.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Caching itself is the real saving.&lt;/strong&gt; Run the same session on Opus 5.5 with no caching and the input alone is 16.32M × $4 = $65.28. That's $66.88 in total, against $6.40 with caching.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faa1ies67vqduffw9z7u2.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Faa1ies67vqduffw9z7u2.png" alt="Stacked bar chart for an example Claude Opus 5.5 coding session: cache reads are 97.6% of tokens but 50% of the $6.40 cost; cache writes and output are 25% each" width="799" height="411"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  Which cuts are temporary
&lt;/h2&gt;

&lt;p&gt;Not every price in that table will still be there in January.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foye1rx0widl5mu7wtkgf.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Foye1rx0widl5mu7wtkgf.png" alt="Timeline of promotional API prices ending: GPT-5.6 Sol promo at least through 21 Nov; Gemini 4 Argon $2/$10 until 30 Nov, then $4/$20; Gemini 3.8 Flash $0.75/$3.75 until 31 Dec, then $1.50/$7.50" width="799" height="434"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;GPT-5.6 Sol:&lt;/strong&gt; OpenAI calls its current price promotional, "available at least through November 21, 2026."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemini 4 Argon:&lt;/strong&gt; $2 input / $10 output through 30 Nov 2026, then $4 / $20 from 1 Dec. Cache reads go from $0.10 to $0.20. The example session doubles.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Gemini 3.8 Flash:&lt;/strong&gt; $0.75 / $3.75 through 31 Dec 2026, then $1.50 / $7.50 from 1 Jan 2027.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Claude Sonnet 5&lt;/strong&gt; went the other way. Its $2 / $10 launch price was due to rise to $3 / $15 on 1 Sep 2026. Anthropic made $2 / $10 the standard price instead.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you pick a model on price, put those dates in your calendar.&lt;/p&gt;

&lt;h2&gt;
  
  
  How to estimate your own prompt caching cost
&lt;/h2&gt;

&lt;p&gt;You don't need my assumptions. Your usage logs have the real numbers.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Pull token counts per call.&lt;/strong&gt; Anthropic's API returns &lt;code&gt;cache_read_input_tokens&lt;/code&gt; and &lt;code&gt;cache_creation_input_tokens&lt;/code&gt; in the usage block. OpenAI reports &lt;code&gt;cached_tokens&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Split them into four buckets:&lt;/strong&gt; cache reads, cache writes, uncached input and output.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Multiply each bucket by its own price.&lt;/strong&gt; Don't use the input price for everything.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Check your hit rate.&lt;/strong&gt; Woolf says the cache tends to be hit around 90-100% of the time, depending on the provider. If yours is much lower, fix that before you switch models.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Watch for cache expiry.&lt;/strong&gt; Anthropic's default cache lives for 5 minutes and is refreshed each time it's read. Step away for a coffee and the next call writes the whole context again. On Opus 5.5, rewriting 100K tokens costs $0.50, while reading them costs $0.02. OpenAI's default is 30 minutes. Both offer longer lifetimes for a higher write price: 1 hour at 2x input on Anthropic, 12 hours at 2.5x on OpenAI.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Then redo the math for the date each temporary price ends.&lt;/p&gt;

&lt;h2&gt;
  
  
  The short version
&lt;/h2&gt;

&lt;p&gt;Input and output prices make the news. For coding agents, the cache-read column and your hit rate decide most of the bill. Cheaper cache reads are why Opus 5.5 costs about half as much as Opus 5 in the example above, even though its headline prices only fell 20%.&lt;/p&gt;

&lt;p&gt;&lt;em&gt;AI tools helped me research and draft this post. I checked every number against the linked sources on 26 Oct 2026.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;What does your agent's split look like? If you've checked your usage logs, I'd love to know how much of your bill is cache reads and how much is output.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>llm</category>
      <category>agents</category>
      <category>discuss</category>
    </item>
  </channel>
</rss>
