元記事の説明文
<p>I didn't have early access to today's <a href="https://www.anthropic.com/news/claude-fable-5-mythos-5">Claude Fable 5</a> release, but I've spent the past ~5.5 hours putting it through its paces. My initial impressions are that this is something of a <em>beast</em>. It's slow, expensive and has been quite happily churning through everything I've thrown at it so far. As is frequently the case with current frontier models the challenge is finding tasks that it can't do.</p>
<p>First, let's review the key characteristics.</p>
<p>Anthropic claim that <a href="https://www.anthropic.com/news/claude-fable-5-mythos-5">Claude Fable 5</a> offers the same performance as Claude Mythos 5, except with much more strict guardrails in place to prevent it being used for harmful things. Those guardrails trigger often enough that the Claude API has new mechanisms for letting you know when you hit them, and even has a <a href="https://platform.claude.com/docs/en/build-with-claude/refusals-and-fallback">new option</a> to request it falls back to another model automatically if something gets rejected.</p>
<p>Claude Mythos 5 is out today as well, <a href="https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5">Anthropic say it</a> "Shares Claude Fable 5's capabilities without the safety classifiers".</p>
<p>The models have a 1 million token context window, 128,000 maximum output tokens and a knowledge cut-off date of January 2026.</p>
<p>They are priced at twice the price of Claude Opus 4.5/4.6/4.7/4.8: $10/million input tokens and $50/million output tokens. There's no increase in price for longer context usage.</p>
<p>Other than that the <a href="https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5">upgrade guide</a> is substantially thinner than <a href="https://platform.claude.com/docs/en/about-claude/models/whats-new-claude-4-8">the similar guide for Opus 4.8</a>.</p>
<h4 id="the-big-model-smell">The big model smell</h4>
<p>The best way to describe Fable is that it feels <em>big</em>. Not just in terms of speed and cost, but also in how much it knows.</p>
<p>Here's an example prompt, which I used to compare Fable's knowledge with Opus 4.8 (typo included):</p>
<blockquote>
<p><code>List all of Simon Willion's open source projects, most recent first, each with a rough date of when they were first released</code></p>
</blockquote>
<p>Opus 4.8's reply <a href="https://gist.github.com/simonw/e77abf645156ade8c5c610f524226dcb#response-1">started like this</a>:</p>
<blockquote>
<p>I don't have a reliable, comprehensive, and date-verified list of all of Simon Willison's open source projects, and I want to be honest about that rather than risk giving you inaccurate dates or fabricated entries.</p>
<p>Here's what I can tell you with reasonable confidence:</p>
<p><strong>Well-known projects by Simon Willison:</strong></p>
<ul>
<li>
<p><strong>LLM</strong> (~2023) – A command-line tool and Python library for interacting with large language models.</p>
</li>
<li>
<p><strong>Datasette</strong> (~2017) – A tool for exploring and publishing data, probably his most prominent recent project.</p>
</li>
<li>
<p><strong>sqlite-utils</strong> (~2018-2019) – A Python library and CLI tool for manipulating SQLite databases, often used alongside Datasette.</p>
</li>
<li>
<p><strong>Django</strong> (2005) – He is one of the co-creators of the Django web framework, along with Adrian Holovaty, originally developed at the Lawrence Journal-World newspaper.</p>
</li>
</ul>
<p>He also maintains a large ecosystem of <strong>Datasette plugins</strong> and many smaller utilities (often with names like <code>shot-scraper</code>, various <code>*-to-sqlite</code> tools for importing data, etc.).</p>
</blockquote>
<p>That's not bad. Here's <a href="https://gist.github.com/simonw/e77abf645156ade8c5c610f524226dcb#response">the start of the response</a> for the exact same prompt from Fable 5 - both models were running without access to search:</p>
<blockquote>
<p>I think you mean <strong>Simon Willison</strong> — the prolific open source developer, co-creator of Django, and creator of Datasette.</p>
<p>An important caveat upfront: a truly complete list isn't feasible. Simon Willison is famously prolific — he has published <strong>hundreds</strong> of repositories (Datasette alone has 100+ plugins, many written by him), and my knowledge has a cutoff date, so I'll be missing his most recent work. Below are his <strong>notable</strong> projects with approximate initial release dates, most recent first:</p>
<h2 id="notable-projects-approximate-dates-">Notable projects (approximate dates)</h2>
<ul>
<li>
<strong>files-to-prompt</strong> (April 2024) – concatenates files into a single prompt for LLMs</li>
<li>
<strong>datasette-extract</strong> (2024) – plugin for extracting structured data using LLMs</li>
<li>
<strong>LLM</strong> (May–June 2023) – CLI tool and Python library for interacting with large language models, plus a large ecosystem of plugins (llm-gpt4all, llm-mistral, llm-claude, etc., 2023 onward)</li>
<li>
<strong>symbex</strong> (June 2023) – search Python code for symbols</li>
<li>
<strong>ttok</strong> and <strong>strip-tags</strong> (May 2023) – CLI tools for token counting and HTML cleanup for LLM pipelines</li>
<li>
<strong>datasette-lite</strong> (May 2022) – Datasette running entirely in the browser via WebAssembly/Pyodide</li>
<li>
<strong>shot-scraper</strong> (March 2022) – automated website screenshots via Playwright</li>
<li>
<strong>s3-credentials</strong> (November 2021) – CLI for creating scoped AWS S3 credentials</li>
<li>
<strong>django-sql-dashboard</strong> (2021) – SQL reporting dashboards for Django</li>
<li>
<strong>Dogsheep suite</strong> (2019) – personal analytics tools: twitter-to-sqlite, github-to-sqlite, healthkit-to-sqlite, dogsheep-beta, etc.</li>
<li>
<strong>sqlite-utils</strong> (2018) – CLI and Python library for manipulating SQLite databases</li>
<li>
<strong>Datasette</strong> (November 2017) – his flagship project; tool for exploring and publishing data</li>
<li>
<strong>csvs-to-sqlite</strong> (2017) – convert CSV files to SQLite</li>
<li>
<strong>Various early tools</strong> (~2007–2010) – soupselect, json-head, geocoders, and others</li>
<li>
<strong>Django</strong> (developed 2003–2005, open-sourced July 2005) – co-created with Adrian Holovaty at the Lawrence Journal-World</li>
</ul>
</blockquote>
<p>(Here's <a href="https://gist.github.com/simonw/8d05ce836d44bd8543055c9614f7c478">GPT-5.5 for good measure</a>. It listed even more projects than Fable did!)</p>
<p>In the past I've stated that I don't care about how much models <em>know</em> - I want them to be able to manipulate text and code in useful ways and actively look up the information they need via search tools, not bake it into their weights.</p>
<p>But knowledge like this is a reasonably good proxy for model size - you can cram a whole lot more details about the world into a larger number of parameters.</p>
<p>Does <em>knowing more stuff</em> mean the model is better at the tasks we pose to it? I can certainly imagine how a coding model with deeper knowledge of modern libraries and patterns could crunch through coding tasks more effectively.</p>
<p>Is Fable really bigger than Opus? Anthropic haven't said anything about model size, so all we have are tea-leaves, but the speed, pricing and my own poking at its knowledge make me think that it's a large model. Maybe the largest yet from any vendor.</p>
<h4 id="using-fable-in-claude-ai">Using Fable in Claude.ai</h4>
<p>Anthropic made Fable 5 available across all of their surfaces - the <a href="https://claude.ai/">Claude.ai</a> chat interface, Claude Code for web, Claude Code CLI and Claude Cowork as well. The model is available "until June 22nd" on the subscription plans (I'm on $100/month Max at the moment), after which it will be billed extra.</p>
<p>Claude.ai is often under-estimated. Since <a href="https://simonwillison.net/2025/Sep/9/claude-code-interpreter/">September 2025</a> every chat has had access to a full container environment to run code, including the ability to install additional packages and even clone repositories directly from GitHub.</p>
<p>Last week <a href="https://simonwillison.net/2026/Jun/6/micropython-in-a-sandbox/">I released micropython-wasm</a>, a Python library that uses <a href="https://wasmtime.dev/">wasmtime</a> to run a custom build of MicroPython in WebAssembly to act as a sandbox for untrusted Python code.</p>
<p>I decided to see if Fable could upgrade that to running full Python instead. I started with this prompt:</p>
<blockquote>
<p><code>Clone simonw/micropython-wasm from GitHub and research how this could use a full Python as opposed to MicroPython</code></p>
</blockquote>
<p>Fable identified that it could use Brett Cannon's <a href="https://github.com/brettcannon/cpython-wasi-build/releases">cpython-wasi-build</a> builds for this, but was unable to download them itself due to environment restrictions.</p>
<p>So I grabbed the two zip files from that page and uploaded them to Claude:</p>
<blockquote>
<p><code>Here's the Brett Cannon builds</code> (<code>python-3.zip</code>, <code>_build-python-3.zip</code> as attachments)</p>
</blockquote>
<p>And that was that. It churned away for a few minutes and got the entire thing working.</p>
<p>Part of the response included:</p>
<blockquote>
<p>I tried the cleaner single-zip-stdlib approach to shrink the filesystem surface, but CPython's <code>getpath</code> bootstrap fails to find <code>encodings</code> from inside a zip without more prefix finessing — the directory-preopen approach works reliably, so that's what the PoC uses. The zip path is solvable but needs <code>_PYTHONHOME</code>/frozen-getpath work.</p>
</blockquote>
<p>So I said:</p>
<blockquote>
<p><code>Try a bit more at the single-zip-stdlib problem</code></p>
</blockquote>
<p>Then a little later:</p>
<blockquote>
<p><code>I want a wheel that has the whole system in it, the Python wrappers and the WASM files and the stdlibrary, so I can do uv run --with path-to-whl python -c "demo code"</code></p>
</blockquote>
<p>... and it gave me <a href="https://static.simonwillison.net/static/cors-allow/2026/cpython_wasm-0.1.0-py3-none-any.whl">this 13.9MB cpython_wasm-0.1.0-py3-none-any.whl</a> file. You can try running Python code in a sandbox using that wheel URL and <code>uv</code> like this:</p>
<div class="highlight highlight-source-shell"><pre>uv run --with https://static.simonwillison.net/static/cors-allow/2026/cpython_wasm-0.1.0-py3-none-any.whl \
cpython-wasm -c <span class="pl-s"><span class="pl-pds">'</span>print(45 ** 56)<span class="pl-pds">'</span></span></pre></div>
<p>Here's <a href="https://claude.ai/share/a73b8b8b-8ebc-4fef-9e5c-7438e5e7ae35">the full chat transcript</a>.</p>
<p>This was a <em>very</em> strong start.</p>
<h4 id="adding-features-to-datasette-agent-and-llm-using-claude-code">Adding features to Datasette Agent and LLM using Claude Code</h4>
<p>Before I'd realized it was Fable day, my stretch goal for today was to add a new feature to <a href="https://agent.datasette.io/">Datasette Agent</a>: I wanted tool calls within that agent software to gain the ability to pause mid-execution and request approval directly from the user.</p>
<p>This felt like a suitably meaty task to throw at the new model.</p>
<p>Over the course of the day Fable not only <a href="https://github.com/datasette/datasette-agent/pull/20">solved that problem</a>, it also identified and then implemented four issues in my underlying LLM library that would help support this kind of advanced pause-resume mechanism in tool calls.</p>
<p>It got everything working first using somewhat gnarly hacks, but the moment I told it that changes to LLM itself were in scope it set to work unraveling the hacks and turning them into supported features of LLM instead.</p>
<p>My stretch goal turned into <a href="https://llm.datasette.io/en/latest/changelog.html#a3-2026-06-09">LLM 0.32a3</a>, almost entirely written by Fable. Here are the release notes:</p>
<blockquote>
<p>Driven by the needs of <a href="https://github.com/datasette/datasette-agent">Datasette Agent</a>'s human-in-the-loop <code>ask_user()</code> feature, made the following improvements to how tool calls work:</p>
<ul>
<li>Tool implementations can declare a parameter named <code>llm_tool_call</code> in order to be passed the <code>llm.ToolCall</code> object for the current invocation. This allows them to access the current <code>llm_tool_call.tool_call_id</code>. See <a href="https://llm.datasette.io/en/latest/python-api.html#python-api-tools-llm-tool-call">Accessing the tool call from inside a tool</a>. <a href="https://github.com/simonw/llm/pull/1480">#1480</a>
</li>
<li>Every tool call is now guaranteed a unique <code>tool_call_id</code> - providers that do not supply one get a synthesized <code>tc_</code>-prefixed ULID. <a href="https://github.com/simonw/llm/pull/1481">#1481</a>
</li>
<li>Tools can raise a <code>llm.PauseChain</code> exception to cleanly pause the tool chain, useful for things like waiting for human approval. The exception propagates to the caller with <code>.tool_call</code> and <code>.tool_results</code> (completed sibling results) attached, and no model call is made with a placeholder result. See <a href="https://llm.datasette.io/en/latest/python-api.html#python-api-tools-pause">Pausing a chain from inside a tool</a>. <a href="https://github.com/simonw/llm/pull/1482">#1482</a>
</li>
<li>Failure semantics for concurrent tool execution: async sibling tool calls always run to completion before a pause or hook exception propagates. <a href="https://github.com/simonw/llm/pull/1482">#1482</a>
</li>
<li>Chains can now resume from a <code>messages=</code> history ending in unresolved tool calls: the calls are executed through the normal <code>before_call</code>/<code>after_call</code> machinery before the first model call, skipping any that already have results. The <code>execute_tool_calls()</code> method also accepts a new optional <code>tool_calls_list=</code> argument for executing an explicit list of <code>ToolCall</code> objects in place of the calls requested by the response. See <a href="https://llm.datasette.io/en/latest/python-api.html#python-api-tools-resume">Resuming a chain with pending tool calls</a>. <a href="https://github.com/simonw/llm/pull/1482">#1482</a>
</li>
<li>Fixed a bug where the async tool executor silently dropped calls to tools not present in <code>tools=</code> - these now return <code>Error: tool "..." does not exist</code> results, matching the sync executor. <a href="https://github.com/simonw/llm/pull/1483">#1483</a>
</li>
</ul>
</blockquote>
<p>I'm really impressed with the quality of API design, tests, code and documentation that Fable put together for this. I spent several hours on it today, but it feels like several days' worth of work.</p>
<h4 id="how-much-i-ve-spent">How much I've spent</h4>
<p>I recently started using <a href="https://agentsview.io">AgentsView</a> to help track my local LLM usage across all of the different coding agents. I published a <a href="https://til.simonwillison.net/llms/agentsview-custom-model-price">TIL today</a> about adding custom Fable pricing to that tool, which I expect will not be necessary in the very near future.</p>
<p>After setting the price, I ran this command to start a localhost web server to explore my usage:</p>
<pre><code>uvx agentsview serve
</code></pre>
<p>Here's the treemap showing the breakdown of my Fable usage across various projects today:</p>
<p><img alt="Screenshot of a cost tracking dashboard with two panels. The first panel is titled "Cost Attribution" with toggle buttons for Project / Model / Agent and Treemap / List, with Project and Treemap selected. Italic text reads "Click to hide from chart". A treemap shows a large red block labeled prod_datasette_agent $99.26 89.9%, with smaller blocks to its right labeled cloud (blue), datasette (teal), llm (red), and money (pink), plus a tiny orange sliver. A legend lists: 1 prod_datasette_agent $99.26, 2 cloud $3.98, 3 datasette $2.81, 4 llm $2.30, 5 money $1.92, 6 simon $0.15. The second panel is titled "Top Sessions by Cost" and lists nine sessions, each with a "Claude" badge, a prompt excerpt, a project name with a session UUID (omitted here), a token count, and a cost: 1. Review ./datasette-agent and ./datasette-apps - we are going to add a new feature to agent but you ... prod_datasette_agent, 78.2M, $99.26. 2. issues.db is a copy of the Datasette issues database. There are a LOT of notes in there relating to... datasette, 826.8k, $2.81. 3. Consult fly-docs and then look at datasette.cloud (which launches fly machines) and datasettecloud-... cloud, 924.7k, $2.61. 4. simonwillisonblog.db is a copy of my blog, plus all my software releases and other interesting thin... money, 542.9k, $1.92. 5. Look in datasette.cloud and figure out all remaining steps and decisions that need to be made in or... cloud, 455k, $1.37. 6. Review PRs and issues filed against this repo within the last 4 weeks and see if any deserve to be ... llm, 323.3k, $0.95. 7. run mypy, llm, 320.9k, $0.76. 8. [Image #1] fix this in github actions, llm, 183.9k, $0.59. 9. simon, simon, 26.4k, $0.15." src="https://static.simonwillison.net/static/2026/agentsview-fable-full-day.jpg" /></p>
<p>I used $110.42 worth of tokens today, all as part of my $100/month subscription.</p>
<h4 id="and-some-pelicans">And some pelicans</h4>
<p>I ran "Generate an SVG of a pelican riding a bicycle" against all five thinking effort levels with Fable.</p>
<p>Here are <a href="https://tools.simonwillison.net/markdown-svg-renderer#url=https%3A%2F%2Fgist.github.com%2Fsimonw%2F94fde31c34a0400c1d29f57e6a708e6b">the results</a>, including the token cost for each one:</p>
<div style="display: flex; margin-bottom: 1em;">
<figure style="margin: 0;">
<img alt="low" src="https://static.simonwillison.net/static/2026/fable-low.jpg" style="width: 100%; height: auto;" />
<figcaption style="text-align: center;">low: <a href="https://www.llm-prices.com/#it=25&ot=1929&sel=claude-fable-5">1,929 out, 9.67c</a></figcaption>
</figure>
<figure style="margin: 0;">
<img alt="medium" src="https://static.simonwillison.net/static/2026/fable-medium.jpg" style="width: 100%; height: auto;" />
<figcaption style="text-align: center;">medium: <a href="https://www.llm-prices.com/#it=25&ot=2290&sel=claude-fable-5">2,290 out, 11.475c</a></figcaption>
</figure>
<figure style="margin: 0;">
<img alt="high" src="https://static.simonwillison.net/static/2026/fable-high.jpg" style="width: 100%; height: auto;" />
<figcaption style="text-align: center;">high: <a href="https://www.llm-prices.com/#it=25&ot=2057&sel=claude-fable-5">2,057 out, 10.31c</a></figcaption>
</figure>
<figure style="margin: 0;">
<img alt="xhigh" src="https://static.simonwillison.net/static/2026/fable-xhigh.jpg" style="width: 100%; height: auto;" />
<figcaption style="text-align: center;">xhigh: <a href="https://www.llm-prices.com/#it=25&ot=5992&sel=claude-fable-5">5,992 out, 29.985c</a></figcaption>
</figure>
<figure style="margin: 0;">
<img alt="max" src="https://static.simonwillison.net/static/2026/fable-max.jpg" style="width: 100%; height: auto;" />
<figcaption style="text-align: center;">max: <a href="https://www.llm-prices.com/#it=25&ot=14430&sel=claude-fable-5">14,430 out, 72.175c</a></figcaption>
</figure>
</div>
<p>It's interesting that high ended up using fewer tokens than medium for this particular run.</p>
<p>Here are the <a href="https://simonwillison.net/2026/May/28/claude-opus-4-8/#and-some-pelicans">Opus 4.8 pelicans</a> for comparison.</p>
<p>Tags: <a href="https://simonwillison.net/tags/ai">ai</a>, <a href="https://simonwillison.net/tags/generative-ai">generative-ai</a>, <a href="https://simonwillison.net/tags/llms">llms</a>, <a href="https://simonwillison.net/tags/anthropic">anthropic</a>, <a href="https://simonwillison.net/tags/claude">claude</a>, <a href="https://simonwillison.net/tags/llm-pricing">llm-pricing</a>, <a href="https://simonwillison.net/tags/pelican-riding-a-bicycle">pelican-riding-a-bicycle</a>, <a href="https://simonwillison.net/tags/llm-release">llm-release</a>, <a href="https://simonwillison.net/tags/claude-mythos">claude-mythos</a></p>