What's new
The highlights from Anthropic's last few weeks of releases.
More detail on Fable 5's cyber safeguards and a proposed jailbreak severity framework
Anthropic published the specifics of what Fable 5's safety classifiers block versus allow across cybersecurity use cases, plus an early draft of a Cyber Jailbreak Severity (CJS) scale built with Amazon, Microsoft, and Google. A related HackerOne program now takes reports on potential Fable 5 jailbreaks.
Claude Fable 5 and Claude Mythos 5 access restored
Global access to Fable 5 returned today after a temporary suspension; Mythos 5 is back for select US organizations. Enterprise admins also gained model entitlement controls (beta) to manage which models and effort levels their users can access.
Claude Sonnet 5 launched
Anthropic's most agentic Sonnet yet, with real gains in reasoning, tool use, coding, and knowledge work over Sonnet 4.6. It's now the default model in Claude Code, ships with a native 1M-token context window, and carries promotional pricing of $2 / $10 per Mtok through August 31, 2026.
Claude Science workbench goes live
A customizable app for researchers that bundles the tools and packages scientists already use, produces auditable artifacts, and gives flexible access to compute.
Claude Tag for Slack
Team and Enterprise plans can now tag @Claude directly inside Slack threads and delegate tasks without leaving the conversation.
Edit Claude's drafts in place
Drafts written in chat or Cowork now open beside the conversation — highlight a section, describe the change, and Claude edits it right there instead of regenerating the whole document.
Claude certifications
Anthropic launched its first official technical certification in March 2026 as part of the Claude Partner Network. One exam is live today, with more planned for later in 2026.
Claude Certified Architect — Foundations
Anthropic's first technical certification. Tests whether you can design and ship production-grade Claude applications at enterprise scale: agentic architecture, MCP integration, Claude Code workflows, prompt engineering, and context management. Closed-book, no AI assistance.
Who it's for
Open to organizations in the Claude Partner Network. Solution architects, technical partners, and consultants building Claude implementations are the intended audience. The first 5,000 partner-employee exam attempts have the fee waived.
Retakes
Up to 4 attempts per exam in any rolling 12-month period. Wait 14 days after attempt 1, 30 days after attempt 2, and 90 days after attempt 3 before retesting.
Exam domain breakdown
Domain weights are the most commonly reported breakdown from exam-prep sources; treat them as directional rather than official, and confirm against Anthropic Partner Academy materials before studying.
Free courses & completion certificates
Separate from the paid, proctored CCA-F exam above: anthropic.skilljar.com hosts Anthropic Academy's full catalog of free, self-paced courses. Each one issues a completion certificate — no Partner Network membership or exam fee required.
Getting started
Building & development
Cloud platforms
Setting up CI/CD with Git and Claude Code
Claude Code has an official GitHub Action so you can mention @claude in issues and PRs, or run it headless on a schedule. Here's the fastest path from zero to a working pipeline.
-
Run the installer from inside Claude Code
In your repo, run
/install-github-app. It installs the Claude GitHub App and walks you through adding the workflow file and API key secret. Requires repo admin access. -
Or set it up manually
Install the app from
github.com/apps/claude, addANTHROPIC_API_KEYas a repository secret, then copy a starter workflow fromanthropics/claude-code-action's examples folder into.github/workflows/. - Point it at a CLAUDE.md A CLAUDE.md at the repo root defines coding standards, review criteria, and project rules — Claude follows it automatically when creating PRs or responding to mentions.
-
Trigger it
Tag
@claudein a PR comment or issue, or configure the workflow to run automatically onpull_requestevents for hands-free review on every PR. -
Control cost
Cap
--max-turns, set workflow-level timeouts, and use GitHub's concurrency controls so a runaway job doesn't burn API tokens or Actions minutes.
# Minimal setup: Claude responds whenever someone tags @claude
name: Claude Code
on:
issue_comment:
types: [created]
pull_request_review_comment:
types: [created]
jobs:
claude:
runs-on: ubuntu-latest
steps:
- uses: anthropics/claude-code-action@v1
with:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
.github/workflows/pr-review.yml — automatic review on every PR
# Runs on every new or updated pull request, no @mention needed
name: Auto PR Review
on:
pull_request:
types: [opened, synchronize]
jobs:
review:
runs-on: ubuntu-latest
steps:
- uses: anthropics/claude-code-action@v1
with:
anthropic_api_key: ${{ secrets.ANTHROPIC_API_KEY }}
prompt: "Review this PR for bugs, security issues, and style."
claude_args: "--max-turns 5"
Using Amazon Bedrock or Google Vertex AI instead of the direct API? Same action, different auth block — swap in use_bedrock: "true" or use_vertex: "true" with OIDC-based cloud credentials rather than a static API key. Full setup steps are in the Claude Code GitHub Actions docs.
Best prompting techniques
Straight from Anthropic's own prompt engineering documentation — the handful of techniques that move the needle most.
Put long documents first
Place reference material, transcripts, or code above your instructions and question, not below. This measurably improves performance across every Claude model.
Use XML tags to separate content types
Wrapping instructions, context, and input in their own tags stops Claude from blending them together.
Set a role in the system prompt
"You are a senior security engineer reviewing this diff" focuses tone and judgment far more reliably than describing the task alone.
Ask Claude to quote before it answers
For long-document tasks, have Claude pull the relevant quotes first, then answer using only those quotes. Cuts through noise and reduces hallucination.
Don't over-specify
Only add constraints that are directly necessary. Over-engineered prompts with excessive edge-case handling tend to make Claude more rigid, not more accurate.
Give examples, not just rules
Two or three well-chosen input/output examples usually beat a paragraph of abstract instructions, especially for formatting or tone.
Cool resources
The places worth going back to.
platform.claude.com/docs
The API and developer documentation — models, tool use, prompt engineering, agent SDK.
support.claude.com
Product help center and the release notes feed — the fastest way to see what shipped this week.
code.claude.com/docs
Claude Code documentation: GitHub Actions, hooks, slash commands, MCP servers, remote control.
github.com/anthropics
Open-source repos including claude-code, claude-code-action, and claude-agent-sdk.
anthropic.skilljar.com
Anthropic Courses — free, self-paced training on Claude, prompting, and building with the API.
claude.com/partners
Claude Partner Network — training, certification access, and the partner directory.
claude.com/blog & anthropic.com/news
Product announcements and the Anthropic newsroom for model launches and policy updates.
anthropic.com/engineering
Engineering blog — deep dives on how Anthropic builds agents, evals, and infrastructure.
claude.com/connectors
Directory of MCP connectors and partner-built skills you can plug into Claude.