“Opus Advisor Cuts Agent Costs” — AlphaSignal

Why this is in the vault

Curation issue covering Anthropic’s new advisor tool pattern, OpenAI’s $100 Pro tier, and Meta’s Neural Computer paper. The advisor tool is directly relevant to RDCO’s agent architecture and cost management.

Topics covered

1. Anthropic Opus Advisor tool [RDCO-RELEVANT] — New API pattern: run Sonnet/Haiku as executor, escalate to Opus only when reasoning gets hard. Single API call, no external orchestration. Results: SWE-bench +2.7 points, cost -11.9% vs Sonnet-only. Haiku jumps from 19.7% to 41.2% on BrowseComp with advisor. Haiku+Advisor is 85% cheaper than Sonnet for comparable workloads. Controlled via advisor_20260301 tool definition with max_uses parameter.

2. OpenAI $100/month Pro tier — Higher Codex limits for sustained coding sessions. GPT-5.4 Pro reasoning model, ~400K token context. Plus stays at $20, rebalanced for frequent shorter sessions. Promotional 10x Plus capacity through May 31.

3. Lambda MFU optimization [SPONSORED] — Framework pushing Model FLOPS Utilization past 60% (vs industry 35-45%) on Blackwell GPUs. Llama 3.1 8B-405B benchmarks. Root causes: memory overhead, parallelism strategy, serialized communication.

4. Meta Neural Computer — Paper proposing models that act as computers rather than using them. Trains on screen recordings, reproduces desktop/terminal workflows. Fails on multi-step reasoning with cross-step dependencies. Research-stage, not production-ready.

5. Anthropic Monitor tool — Claude can now run background scripts and wake on events. Relevant to always-on agent patterns.

6. Perplexity + Plaid integration — Financial account connections for unified tracking.

7. Cursor PR demos — Agents attach screenshots/demos to pull requests for review.

RDCO mapping

• Opus Advisor pattern — directly applicable to our agent architecture. Currently we run Opus for everything; an executor/advisor split could cut costs significantly while maintaining quality on hard decisions. Worth prototyping.
• Monitor tool — relevant to our always-on channels agent architecture. Background script execution with event-based wake is the pattern we already use via LaunchAgent/tmux; native support simplifies it.
• Cost optimization — the 85% cheaper Haiku+Advisor vs Sonnet result is the kind of concrete ROI data useful for advisory client conversations about AI cost management.

Related

• 2026-04-12-alphasignal-claude-code-leak-harness-engineering — the harness architecture that the advisor pattern fits into
• 2026-04-11-garry-tan-thin-harness-fat-skills — advisor as a harness-level routing decision