Meta’s Muse Spark Outruns Llama 4 with 10× Less Compute, 86 % Reasoning Score

TL;DR

• Meta Launches Muse Spark, Its Most Powerful AI Model, Achieving 52 on AI Index v4.0 and Powering 3B+ Users Across Meta Apps
• AI Model Mythos Preview Achieves AGI Milestone, Anthropic Locks Access
• Meta shuts down internal 'Claudeonomics' leaderboard after employees burned 60 trillion tokens in 30 days

🤖 Meta’s Muse Spark: 10× Leaner Compute, #5 Global AI Rank, 3B Users Next

Meta’s new Muse Spark just trained on 10× LESS compute than Llama 4 yet vaults to on the global AI Index—86 % reasoning score, 74 % agentic search. 🤯 That’s like a Prius outrunning a V12. Who wins when 3 billion users wake up to this tomorrow—devs or regulators?

Meta Superintelligence Labs flipped the switch on Muse Spark last week, a home-grown large-language model that scored 52 on the AI Index v4.0—nearly triple the 18 points its predecessor Llama 4 Maverick managed. Trained with one-tenth the compute and already live inside Facebook, Instagram, WhatsApp and Messenger, the closed-source engine is answering 3 billion users while narrowing the gap with OpenAI’s GPT-5.4 (57) and Anthropic’s Claude Opus 4.6 (53).

How did Meta close the gap so fast?

• July 2025 – April 2026: Ground-up rebuild of the AI stack for native multimodality, tool-use and visual chain-of-thought.
• Training tokens: 57–120 million output tokens, on par with Google’s Gemini 3.1 Pro but at >10× lower compute cost than Llama 4.
• CharXiv reasoning: 86.4 %, best-in-class; DeepSearchQA agentic search: 74.8 %; HealthBench hard medical questions: 42.8 %.
• Coding lag: Pass@5 Python score of 33 % trails GPT-5.4’s 41 %, signaling the next upgrade target.

What changes for users, rivals and regulators

Reach: 3 billion people now tap a sharper assistant—equal to adding a Canada-sized population of new brains to every chat thread.
Competition: Meta’s 8 % share-price pop added ~$120 billion in market cap in a day, pressuring Google and OpenAI to defend their developer turf.
Transparency: Closed API sparks researcher pushback; ARC-AGI 2 score of 42.5 % fuels calls for external audits.
Regulation: FTC and EU already probe data pipelines; closed models heighten antitrust risk.

Outlook: faster, bigger, (maybe) opener

• Q2 2026: Full US rollout complete; EU & Asia-Pacific follow in Q3-Q4.
• Late 2026: “Avocado” API exits beta; subscription pricing expected to lift Meta’s non-ad revenue above $1 billion run-rate.
• 2027: Pass@5 coding benchmark targeted >40 %; ARC-AGI 2 >60 % via RLHF cycles.
• 2028: Open-source “Muse 2” planned; 100 B-parameter scale on custom “Hyperion” chips; projected AI-services revenue >$5 billion.

Meta’s nine-month sprint shows AI leadership can still be bought with focus, cash and data. If the company keeps shortening the coding gap and loosens the source code, Muse Spark won’t just serve 3 billion chats—it could set the standard the rest of the industry is forced to copy.

🤯 97.6% Math Genius AI Spots 1,000 Zero-Days in 12 Minutes—US Elite-Only, China Goes Open

97.6% on USAMO & 1,000+ zero-days in 12 min 🤯—Claude Mythos spots bugs faster than you can order coffee. 30% false+ still need human eyes. US elites only, China gets open code. Who wins the cyber race?

On Tuesday Anthropic’s “Mythos Preview” became the first model to top 97 % on the 2026 U.S.A. Math Olympiad proof set and 93.9 % on the SWE-Bench coding suite—clear AGI territory. The catch: ordinary developers can’t touch it. Forty pre-cleared partners (AWS, JPMorgan, NVIDIA, et al.) get the keys, while Beijing-aligned labs are promised open-source drops under “Project Glasswing.”

How it works

Mythos is a 3.5-GW-cluster behemoth trained to chain abstract proofs into end-to-end exploits. Given a codebase, it averages 12 minutes from vulnerability detection to working payload—tasks that occupy human red-teamers for days. Internal logs show 83 % of exploits fire on first compile; 595 Tier-1/2 crashes were logged in sandbox tests.

Impacts

Security surface: ≥1 000 high-severity CVEs surfaced, including a 27-year-old OpenBSD kernel flaw; 500+ zero-days linger in earlier open-source releases.
Market leverage: Anthropic’s gated playbook lets the 40 insiders patch faster than rivals using public models (GPT-5.4 scores 80.6 %).
Geopolitics: Planned open-source release to Chinese state labs risks arming 30 already-documented intrusions against U.S. finance and tech firms.

Institutional response

The U.S. Defense Secretary flagged Anthropic as a “potential national-security threat”; a federal injunction is pending. CISA, NIST and allies will receive incident reports within 48 h, but no statutory guardrails yet cover AI-generated exploits.

Outlook

• Q3 2026: Partner-only drops add 200–300 fresh zero-days/month, forcing sprint-patch cycles.
• 2027: If export controls pass, Anthropic could lock 30 % of enterprise AI-security spend; Chinese open-source forks raise cyber-conflict heat.
• 2028: “Red-team-as-a-service” SaaS platforms emerge, pushing ISO standards on AI exploit handling.

The takeaway: an AI that aces Olympiad proofs now writes faster hacks than most nation-state crews. Until democratized access matches democratized risk, the security gap between the insider forty and everyone else widens daily.

💸 Meta Axes 60T-Token ‘Claudeonomics’ Board After $1.4M/Day Burn

60 trillion AI tokens burned in 30 days—like every person on Earth typing 7,500 pages 📄. Meta just killed its "Claudeonomics" leaderboard after top dev racked up $1.4 M/day. Your move SV: will budget caps spark smarter code or talent exodus?

Meta pulled the plug on “Claudeonomics,” the internal leaderboard that turned its own AI into a high-stakes arcade. Engineers chased 60 trillion tokens in April—enough text to refill Wikipedia 2,000 times—pushing the top user to burn $1.4 million of compute in a single day. When the numbers leaked Wednesday, leadership froze the board overnight; the score that once unlocked 200 % bonuses is now a blank page.

What the meter rewarded

Every prompt to Claude Opus 4.6 cost $5 per million tokens in, $25 out. Hitting 281 billion tokens meant little more than “prompt, cut, paste, repeat,” yet that raw count decided promotions. No quality filter, no revenue link—just volume.

Impacts inside Meta

• Budget: $180 million vanished in 30 days → Q1 AI opex jumped 18 %.
• Morale: 85 000 staff lost a visible status game → hallway buzz shifts from bragging to budget dread.
• Talent: Anthropic engineers already spend $150 k/month; Nvidia’s CEO warns value proof starts at $250 k/year → pressure mounts to justify every cent.

Industry ripples

Shopify and OpenAI run copy-cat scoreboards; expect rapid rule tweaks. Meta’s next policy caps each engineer at $250 k annually and ties bonuses to shipped features, not tokens burned.

Timelines

• Q2 2026: Private dashboards replace public boards; token burn projected to fall 30 %.
• Q3 2026: Firm-wide budget alerts switch on; YoY AI cost growth drops from 40 % to ~20 %.
• 2027–2028: “AI-value units” (cost × accuracy × latency) replace token counts industry-wide; Meta’s $600 B capex plan pivots to efficiency over scale.

Bottom line

Silicon Valley just learned that what gets gamed gets wasted. The race now is to squeeze real products, not token trophies, out of every pricey prompt.

In Other News

• Muun AI Secures $700K Pre-Seed Round to Turn Operational Data into Actionable Insights
• Darwin V6 Merges Gemma 4 and Qwen 3.5 Models, Achieving 90.0% GPQA Diamond Accuracy
• AlphaFold Breakthrough: AI Solves 50-Year Protein Folding Problem, Now Used by 3 Million Scientists
• Google Maps launches AI-generated captions for photos, expanding globally after U.S. rollout