✦ Fast Mode for Opus 4.6 — Up to 2.5× Faster Output

Launched on 7 February 2026, Fast Mode for Claude Opus 4.6 delivers up to 2.5× faster output token generation via a new speed parameter in the Messages API. Previously you had to choose between Opus-level intelligence and sub-second responsiveness; Fast Mode lets you get both — at premium pricing — for latency-sensitive agentic applications.

How to enable Fast Mode

import anthropic

client = anthropic.Anthropic()

response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=1024,
    speed="fast",          # ← enable Fast Mode
    messages=[{
        "role": "user",
        "content": "Summarise this report in bullet points."
    }]
)
print(response.content[0].text)

When to use it

• Real-time agents — coding assistants, live chat, interactive workflows where users are watching the stream.
• Iterative pipelines — multi-step chains where every node adds latency; Fast Mode compounds the saving across each hop.
• High-volume burst workloads — faster throughput means you can serve more concurrent requests within your rate limits.

✦ Compaction API — Effectively Infinite Conversations

The Compaction API (launched in beta on 5 February 2026) provides server-side context summarisation for Claude Opus 4.6 conversations. When your conversation grows beyond the model's practical context limit, the API can automatically distil the history into a compact summary, then continue as if no context was lost. Long-running agents, multi-session workflows, and persistent assistants can now operate without manual context management.

How compaction works

• You set a compaction policy on your request — the API monitors token usage and triggers summarisation when a threshold is approached.
• The server generates a structured summary of prior turns, retaining key facts, decisions, and outstanding tasks.
• The summary replaces the raw history in subsequent turns — significantly shrinking the effective prompt size without losing meaningful context.
• A compaction_metadata object in the response tells you when and how many tokens were compacted.

Minimal usage

response = client.messages.create(
    model="claude-opus-4-6",
    max_tokens=2048,
    compaction={
        "enabled": True,
        "threshold": 0.85   # compact when 85 % of context is used
    },
    messages=long_conversation_history
)

✦ Claude Code: Three New Productivity Features

Recent Claude Code releases (v2.1.74 – v2.1.76) have shipped several quality-of-life improvements aimed at reducing wasted context and giving developers finer control over effort and speed. Here are the three most impactful additions.

1. /context — Actionable Optimisation Suggestions

Running /context in any Claude Code session now shows not just how many tokens are in use, but actionable suggestions identifying the biggest optimisation opportunities in your current context window — for example, large tool results that could be summarised, or conversation branches that are no longer relevant. Use it before a long coding task to ensure you're not burning context on stale scaffolding.

2. /effort — On-the-Fly Effort Control

The new /effort slash command lets you adjust Claude's effort level mid-session without restarting. Higher effort means deeper reasoning (and more tokens); lower effort is faster and cheaper for mechanical tasks like simple edits or reformatting. This maps to the API's effort parameter (which replaced budget_tokens for Opus 4.6).

/effort high    # deep reasoning for architecture decisions
/effort low     # fast mode for search-and-replace style edits

3. worktree.sparsePaths — Sparse Worktree Checkouts

In large monorepos, running Claude Code in an isolated git worktree previously checked out the entire repo — slow and token-expensive. The new worktree.sparsePaths setting in ~/.claude/settings.json lets you specify which directories to include, giving Claude a lean, relevant slice of the repository.

// ~/.claude/settings.json
{
  "worktree": {
    "sparsePaths": ["src/api", "src/shared", "tests/integration"]
  }
}

✦ Custom Subagents & Agent Teams in Claude Code

Claude Code now supports custom subagents — specialised Claude instances defined in your project's .claude/agents/ directory — and agent teams where multiple parallel sessions coordinate through a shared task list. These two features let you decompose complex, long-horizon work in a way that keeps each agent's context window clean and focused.

Custom Subagents

A subagent is a Markdown file in .claude/agents/ with a YAML front-matter block. Claude can delegate to it automatically (based on context) or on demand via /agent <name>. Each subagent runs in its own isolated context window, so tool results and file reads don't pollute your main session.

# .claude/agents/test-runner.md
---
name: test-runner
description: Runs the test suite and reports failures with root-cause analysis
model: claude-haiku-4-5         # fast, cheap for mechanical work
allowedTools: [Bash, Read, Glob]
---

You are a test specialist. When invoked, run `npm test`, capture failures,
read the relevant source files, and return a structured failure report.
Do not fix the code — only diagnose.

Agent Teams

For large parallel workloads, assemble a team: one lead agent breaks the problem into tasks, spawns worker agents to handle them concurrently, and synthesises the results. Workers communicate through a shared task list (created via TaskCreate / TaskUpdate) and message each other through the lead's context.

• Ideal for codebase-wide refactors, bulk migrations, and multi-file analysis.
• Each worker operates with the minimum tool set needed — use allowedTools to enforce least-privilege.
• The lead's context stays small: it only sees task summaries, not each worker's full file read history.