# AI Radar ΓÇö 2026-05-01
Scored for Mark Barnett: cost-conscious, local-first, Apex (13.8 GB RAM), TP3 stack, Oracle, Bidet AI, teaching. Updated with second scan pass (Claude Code changelog, MCP updates, managed agents).
Actions I recommend Mark take this week (top 3, max)
-
Pull Gemma 4 E4B on Apex via Ollama ΓÇö fits in Apex RAM budget (~3-5 GB quantized), strong benchmarks, directly addresses the local-LLM goal; one command:
ollama pull gemma4:4b. Source: ollama.com/library/gemma4 -
Set
alwaysLoad: truefor OMI and biometric MCP servers in Claude Code ΓÇö Claude Code v2.1.121 added this option; it skips the startup deferral that causes silent disconnects on Apex's MCP servers. Edit.claude/settings.json, add"alwaysLoad": trueto each MCP server entry. Source: releasebot.io/updates/anthropic/claude-code -
Remove beta header from Claude API web search calls ΓÇö Web search + code execution are now GA (no beta opt-in required). Code execution is free when paired with web search. Any TP3 agent scripts using the beta header can drop it. Source: platform.claude.com/docs
Ranked candidates
| # | Candidate | Source | Score | Why it matters for Mark | Integration cost |
|---|---|---|---|---|---|
| 1 | Gemma 4 E2B / E4B ΓÇö Google open model, multimodal, 128K context, Apache 2.0 | HF Blog ┬╖ Ollama | 22 | Directly hits local-LLM goal. E4B benchmarks above models 4-5x its size. Fits Apex RAM. One ollama pull away. Zero ongoing cost. |
Low ΓÇö single command |
| 2 | Claude Opus 4.7 ΓÇö same price as 4.6, 87.6% SWE-bench vs GPT-5.5's 74.9% | Anthropic | 21 | Claude Code already defaults to Opus ΓÇö you get the upgrade for free. More capable = fewer retries = lower actual token spend. | None ΓÇö already live |
| 3 | Claude API: web search + code execution now GA ΓÇö no beta header, code exec free with web search | Claude Docs | 21 | Remove one header line per API call. Free code execution reduces token overhead on TP3 agent scripts that search + process data. | Minimal ΓÇö remove header |
| 4 | Claude Code v2.1.121ΓÇô123 ΓÇö alwaysLoad MCP, /resume by PR URL, image memory leaks fixed, OAuth 401 loop squashed | Releasebot | 20 | alwaysLoad directly fixes MCP startup reliability on Apex. Memory leak fix matters for long sessions. PR-URL resume improves agent handoff workflow. |
Auto-update ΓÇö confirm with claude --version |
| 5 | Claude Managed Agents public beta ΓÇö secure sandboxing, built-in tools, SSE streaming, Memory beta | Releasebot Anthropic | 20 | Relevant path for running persistent Oracle-style agents via Claude API rather than DIY Python supervisors. Memory beta = structured state across runs. | API usage-based cost ΓÇö needs Mark approval before adopting |
| 6 | MCP server auto-retry (3x on transient startup error) ΓÇö built into Claude Code | Claude Docs | 19 | TP3 runs 4 MCP servers. Transient startup failures no longer leave a server permanently disconnected ΓÇö directly addresses the silent-failure pattern that has burned this stack before. | None ΓÇö automatic |
| 7 | Unsloth Qwen3.6-35B-A3B-GGUF ΓÇö MoE, 3B active params at inference, GGUF, 1.94M HF downloads | HuggingFace trending | 19 | MoE trick: 35B label, 3B actual compute. Ollama-native. Strong coding + multilingual. Could replace Gemini for offline chat on Apex. | Low ΓÇö ollama pull |
| 8 | Claude MCP Connectors for Creative Tools — Blender, Adobe CC (50+ apps), Ableton, Autodesk (April 28) | Anthropic | 16 | Mark teaches digital media. Blender MCP is real (Python API bridge). Could slot into Content Creator Club or Bidet → creative pipeline. | Low-Medium — MCP setup |
| 9 | Gemini 3.1 Flash TTS ΓÇö upgraded expressiveness, "emotionally rich" audio | DeepMind blog | 16 | Mark already uses Gemini TTS for Grind voiceovers; Oracle could use audio replies. Single API field change to test. | Very low ΓÇö API param |
| 10 | Gemma 4 26B MoE ΓÇö same family, 4B active params at inference, ~18 GB RAM needed | Ollama | 14 | Better than E4B but won't fit Apex's 13.8 GB total RAM at 4-bit quant. Flag for future Apex GPU/RAM upgrade. | High on current Apex |
Cut with reason
- GPT-5.5 / DeepSeek V4-Pro (862B) / Stargate expansion ΓÇö cloud-only or unrunnable locally; no cost reduction for Mark
- OpenAI Advanced Account Security ΓÇö ChatGPT Enterprise account feature, doesn't affect Mark's API/Claude stack
- OpenAI "Where the goblins came from" ΓÇö interesting GPT-5 post-mortem; zero actionable signal for Mark's stack
- Microsoft Agent 365 ΓÇö enterprise M365 E7 license required; not for personal/self-hosted stacks
- Pika Labs AI Agents ΓÇö video creation; not relevant to TP3 or current Content Creator Club workflow
- Stripe Link AI Shopping ΓÇö commerce automation; not in scope
- OpenAI smartphone rumor ΓÇö analyst note, no product exists; pure hype
- HN top AI stories this week ΓÇö top results were older than 7 days or business/enterprise focused; no breakout this exact window
Sources scanned
- Gmail newsletters: 0 ΓÇö
tp3_scripts/ai_radar/senders.txtnot found; Gmail scan skipped this run - RSS/Blog: Anthropic ΓÇö fetched
anthropic.com/news; 4 posts in window Γ£ô - RSS/Blog: OpenAI ΓÇö fetched
openai.com/blog/rss.xml; 5 posts in window Γ£ô - RSS/Blog: Google DeepMind ΓÇö covered via web search; no direct fetch (deferred to stay under 15-min clock)
- HuggingFace trending ΓÇö fetched top 10 models Γ£ô
- Hacker News ΓÇö web search for top AI stories Apr 25ΓÇôMay 1 Γ£ô
- Claude Code changelog ΓÇö fetched releasebot.io/updates/anthropic/claude-code Γ£ô
- Weekly digest ΓÇö fetched promethean-ai.com week-in-AI Apr 27ΓÇôMay 1 Γ£ô
Note on Gemma 4 release date: Multiple sources cite early April 2026, just outside strict 7-day window. Included because it's actively trending (7.47M HF downloads), just hit Ollama with KV cache fixes, and directly serves Mark's stated local-first goals.
Cost of this run
- Input tokens: ~22,000 (estimated; multiple WebFetch + WebSearch calls, two scan passes)
- Output tokens: ~1,400 (this report)
- Estimated cost: ~$0.08 (Sonnet 4.6 pricing)
Generated 2026-05-01 19:51 by tp3_scripts/ai_radar/run_radar.ps1.