Anthropic launches Claude Fable 5.2 and Mythos 5.2, with fast mode up to 2.5 times faster
On Monday, November 9, Anthropic announced Claude Fable 5.2 and Claude Mythos 5.2, which the company describes as the same base model. Fable 5.2 is already available on all platforms, including AWS, Google Cloud and Microsoft Azure, and in the API under the id claude-fable-5-2, at US$ 8 per million input tokens and US$ 40 per million output tokens. Mythos 5.2 remains restricted: it reaches security professionals and life-sciences researchers through the verification programs, and Enterprise customers through Claude Security.
According to Anthropic, those who tested it saw responses up to 30% faster than version 5.1's, and at Medium effort, the default in the apps and in Claude Code, 5.2 solves hard tasks with fewer tokens than 5.1. The announcement's table puts Fable 5.2 ahead of 5.1 and Opus 5.5 on code benchmarks such as FrontierSWE v2 (69.9% against 56.3% for 5.1). The novelty with the most effect on the wallet is the optional fast mode: it generates tokens up to 2.5 times faster, costs twice as much (US$ 16 and US$ 80 per million) and works in Claude Code, Claude Platform, Google Cloud and Microsoft Foundry.
The announcement itself recommends starting with Opus 5.5 for most everyday coding tasks. Fable is for the heavier work, and on Pro plans and standard Team seats it counts toward up to 50% of the weekly usage limit.
For .NET developers: Anyone calling the Claude API from .NET only changes the model id. Before turning on fast mode, do the math: the price per token doubles.
Source: Anthropic
Microsoft measures it: the version label changes how carefully some agents check an update
Waldek Mastykarz, from Microsoft, tested whether coding agents take SemVer seriously. He created a fictitious npm library with a renamed function and delivered the same change, byte for byte, either as 1.4.3 (patch) or as 2.0.0 (major), for GitHub Copilot CLI to update. In the first experiment, Claude Sonnet 5.5 checked the update in 10 of 10 runs when it came as a major, but only in 2 of 10 when it came as a patch. Opus 5.5, GPT-6.1 Sol and GPT-6.1 Terra always checked.
In the second experiment, with six models and 540 runs, merges without a check were 13% under the patch label, 9% under minor and 4% under major. The clearest effect appeared with Claude Haiku 5.5 (37% of merges without checking on patch, against 7% on major). Four merges left main broken, all with a patch or minor label. The author admits the limits: the break was placed on purpose in a patch, the apps were small and there were no release notes.
His conclusion fits in one sentence: the version number tells the agent how much to worry, but only the tests tell whether it should.
For .NET developers: The same goes for NuGet: if an agent or Dependabot bumps packages for you, require green CI on every dependency branch, including the patch ones.
Source: Microsoft Developer Blogs
Copilot code review becomes configurable by API and uses three models at balanced effort
On Friday, November 6, GitHub announced that Copilot code review settings can be managed through the REST and GraphQL APIs, from the company policy down to each repository. On repository endpoints, GitHub Apps and fine-grained PATs with the Administration permission work; on enterprise and organization endpoints, they do not.
The review summary became more compact, with skipped files grouped in a "Files not reviewed" section. The review agent's sessions now show the AI credits spent, the automatic review rule accepts a list of exceptions and, at balanced effort, the review uses three reviewer models plus another one that merges the findings. Reviews that fail or time out are retried once, automatically. It applies to the Pro, Pro+, Max, Business and Enterprise plans.
For .NET developers: Anyone who standardizes .NET repositories with scripts can enable Copilot review through the API alongside the other branch rules, without clicking repo by repo.
Source: GitHub Changelog
Simon Willison suggests hiding agent-written PR descriptions behind a details block
On Monday, Simon Willison published a simple practice for PRs opened with coding agents. The descriptions they write are usually correct, but too long for someone to read. His proposal: open the PR with a short paragraph written by a person, and put the agent's full text inside a <details><summary> block, which GitHub's Markdown shows collapsed.
In the block's label he puts the model and effort level used, such as "Claude Opus 5.5 Medium". Whoever reviews reads the essentials in seconds, opens the details only if they want to, and the history records which tool produced the code.
For .NET developers: It works for any .NET repository on GitHub: just the HTML in the PR body, with no configuration.
Source: Simon Willison
My take
The new model will dominate the headlines, but the news I would take to Monday's meeting is the SemVer one. It shows, with numbers, that an agent can decide how much to check based on a label anyone can type. The defense does not change with the model: tests that run on every update, including the small ones. And this ties in with the calendar: support for .NET 8 and .NET 9 ends tomorrow, November 10, according to the .NET blog, and many people will let an agent handle the migration to .NET 10. Without a test suite that catches the break, the agent will trust the version number.
With Fable 5.2, the noise is the benchmark race; what matters is the price and the fast mode, which doubles the cost per token. Anthropic itself says to start with Opus 5.5 for everyday work, and I would follow that order: move up a model only when the task calls for it. As for Copilot code review by API and the details trick, they are small things that make work with agents easier to govern, and that is where the gain is.
Also read
- Claude Code with .NET: setup and skills (Part 1)
- .NET 10 and C# 14: the new features you need to know
Need help with .NET backend and applied AI? See the service.