Every feature Alex should know about.

Anthropic measured 4× reduction in cases where Claude wrote code with a flaw and reported success. Better self-checking, more willingness to say "I'm not sure," more pushback on bad plans.

You can now insert role: "system" entries at any position in the messages array. Preserves prompt cache hits when instructions change mid-session.

// Inject a new instruction mid-conversation
messages: [
  { role: "user", ... },
  { role: "assistant", ... },
  { role: "system", content: "From now on, use formal voice" },
  { role: "user", ... }
]

When Claude declines, response includes a stop_details.refusal_category. Lets you route different refusal classes (safety, capability, ambiguity) to different next steps.

On 4.8 the effort param defaults to high across all surfaces. Levels: low · medium · high · xhigh · max.

output_config: {
  effort: "xhigh"  // for coding + agentic
}

2.5× output speed at 2× standard pricing ($10/$50 per MTok). Down from prior 6× premium. Opus 4.6 fast mode deprecated — removal 30 days after launch (~June 27).

Pair a fast executor (Haiku) with a high-intelligence advisor (Opus). Most token spend at executor rates; advisor injects strategic guidance mid-generation.

Server-side memory: Claude stores and retrieves info across conversations. Anthropic handles persistence + retrieval. No infrastructure needed on our side.

Claude dynamically discovers + loads tools from a large catalog on demand. Solves the "30+ tools = giant system prompt" problem.

Bash + file manipulation + write code in any language. Replaces the original Python-only code-exec. FREE when used with web search or web fetch.

Native web search tool returns rich SEC filing data (May 18), supports dynamic filtering via code-exec (Feb 17). Memory notes Alex uses Brave LLM-context mode.

Claude calls tools from inside code execution. Reduces latency + token usage in multi-tool workflows.

Retrieve full content from web pages + PDFs. Different from web search (which returns snippets).

Single cache_control field + system caches the last cacheable block and moves it forward as conversation grows. No manual placement of cache breakpoints.

cache_control: { type: "ephemeral" }

Pass diagnostics.previous_message_id on a request — API returns cache_miss_reason explaining where the cache prefix diverged.

In addition to 5-minute default, set 1-hour cache TTL. Better for long-running cron jobs that fire multiple times per hour.

cache_control: { type: "ephemeral", ttl: "1h" }

Smaller prompts cache successfully on Opus 4.8 than 4.7. Means even short cron jobs benefit from caching.