Anthropic API
Claude Fable 5.2 costs 20% less, with four breaking changes
| Per 1M tokens | Fable 5.1 | Fable 5.2 |
|---|---|---|
| Input | $10 | $8 |
| Output | $50 | $40 |
| Batch input / output | Not shown | $4 / $20 |
- What changed
Claude Fable 5.2 is available on the Claude API, AWS, Google Cloud, and Microsoft Foundry with a 1M-token context window. Mythos 5.2, the same model with fewer cyber and biology safeguards, is limited to verified organizations.
- Why it matters
Four changes can break existing code: edits to earlier turns fail the prefix check on every account, only 5.2 models read 5.2 thinking blocks, the older computer use tool errors outside Bedrock, and the advisor tool needs a 5.2 advisor. Forced tool choice isn't supported.
- What to do
Check whether your code edits conversation history before you change the model ID. Rerun your evals at the default high effort and compare cost per accepted task with 5.1.
Gemini API
Gemini 4 Carbon opens to paid developers, and its price doubles December 1
| Per 1M tokens | Through Nov 30 | From Dec 1 |
|---|---|---|
| Input | $2 | $4 |
| Output | $10 | $20 |
| Context caching | $0.10 | $0.20 |
- What changed
Google opened Gemini 4 Carbon to paid Gemini API developers and Gemini for business on November 10. It was limited to selected cyber defenders when it was announced on October 28.
- Why it matters
The introductory rate ends on November 30. Any test you run this month is priced at half what production will cost from December 1. Cyber use of the model still goes through Google's Fairwind program.
- What to do
If you test Carbon now, cost it at the December rate. Compare it with your current model on a task you already score, and record total tokens, not just the per-token price.
VS Code security
Update VS Code to 1.145.3: two fixes involve MCP servers
| CVE | Score | Path |
|---|---|---|
| 2026-120779 | 9.6 | Remote MCP server |
| 2026-114171 | 8.8 | MCP install link |
- What changed
Microsoft fixed 16 Visual Studio Code vulnerabilities in its November security release. The most severe, CVE-2026-120779, lets an attacker who controls a remote MCP server run code when the user clicks a crafted notification link.
- Why it matters
MCP servers are now part of the editor's attack surface. Connecting to a server or clicking an install link can be enough to run code with the developer's permissions.
- What to do
Move every developer machine and dev container to VS Code 1.145.3. Review which remote MCP servers your team connects to, and treat MCP install links like executable downloads.
Git security
Git fixed a clone that can run an attacker's code
| Where | Fix |
|---|---|
| Git | 2.56.1 |
| GitHub Enterprise Server | Patch releases |
| github.com, GHEC | Already fixed |
- What changed
Git released fixes for six vulnerabilities. The worst, CVE-2026-111760, lets a malicious repository replace a submodule's
.gitfile with a symlink, so a recursive clone runs the attacker's hooks.- Why it matters
Coding agents and CI jobs clone repositories without a person watching. Desktop updates don't cover the Git inside agent sandboxes, build images, and runners. That exposure is DVx's inference; GitHub's post doesn't discuss agents.
- What to do
Update Git in agent and CI images, not only on laptops. Until they're patched, don't recursively clone untrusted repositories or merge untrusted tags.
Meta Model API
Meta's Muse Ember is a $0.10 model for agent subtasks
| Per 1M tokens | Up to 200K | Above 200K |
|---|---|---|
| Input | $0.10 | $0.30 |
| Cached input | $0.004 | $0.012 |
| Output | $0.50 | $1.50 |
muse-ember-1.0. The context window is 1M tokens with up to 512K output. View official source ↗- What changed
Meta released Muse Ember, its smallest Muse model, on the Meta Model API and in Muse Code. Meta says Ember is also available on Google Cloud and Oracle Cloud.
- Why it matters
Meta aims it at the work agents repeat: tool calls, routing, and extraction. Its benchmark results are Meta's own, and Meta's table shows it behind Claude Haiku 5.5 on FrontierCode.
- What to do
Route one high-volume subtask, such as classification or field extraction, to Ember on a sample you can score. Keep it only if accuracy holds at the lower cost.
Also since the last issue
Useful updates
Fable 5.2 in Copilot
GitHub added Fable 5.2 to Copilot on November 9 for Pro+, Max, Business, and Enterprise. Admins must turn it on, and it requires data retention; zero data retention is allowed only through the end of the year.
GPT Image 2 deprecated
OpenAI deprecated gpt-image-2 on November 10. It will be removed on February 10, 2027; the replacements are the 2.6 Sunburst and Flare models.
Grok text watermarking
xAI's current watermarking docs say Grok API text watermarking is off by default. Team admins turn it on per model, so don't assume Grok output is marked.
One workflow to try
Patch the tools your agents use
List every place an agent or CI job runs Git or VS Code: laptops, dev containers, runner images, and agent sandboxes. Update Git and VS Code in each, then rebuild the images and confirm the versions in a run log.
Term of the week
Prefix check
The prefix check confirms that nothing before a thinking block changed since the model wrote it: the system prompt, the tools, or an earlier message. On Fable 5.2, a changed prefix returns an error. Code that only appends to the conversation passes.