October 14, 2026 (Wed)

[01]

Top Stories

1. Gemini 4 Argon expands beyond its defender cohort

Paid API and subscription rollout began October 13. Public pricing is $2/M input and $10/M output through November 30, then $4/M and $20/M. Google now states a 1M unified input-output limit; the earlier 2M context report remains a recorded discrepancy.

Why it matters: broader access makes deployment practical, with an explicit price expiry and a context constraint to verify. → Gemini 4 Argon (source)

2. Meta releases Muse Wonder and Wonder Full

Wonder is available through Meta's API and Muse Code; Full through Muse Code subscriptions. Active Memory and Active Focus target persistent context and sustained execution. Numerical benchmark charts were not recoverable from the article's prose.

Why it matters: another frontier agent offering is available, while comparative performance remains unverified here. → Muse Wonder · Muse Wonder Full (source)

3. Liquid AI publishes a local-agent model with clear tradeoffs

LFM2.5-24B-A2B reports LiveClaw 48.09 versus 38.19 for Qwen3.6-35B-A3B, but SWE-Bench Verified with OpenCode 51.00 versus 66.20. Both are vendor measurements.

Why it matters: stronger tool use does not imply stronger coding; choose an evaluation that matches the workflow. → LFM2.5-24B-A2B (source)

4. Anthropic moves its monorepo to Scion

A seven-person team built the internal source-control and review systems and migrated in five months. Anthropic explicitly limits generalization to its unusual agent-heavy workload.

Why it matters: source control and CI can become bottlenecks as code generation accelerates. → Agents (LLM Agents) (source)

[02]

Paper Picks

AgentSWE: 25 tasks test whether delivered agents satisfy requirements beyond visible examples. Read it for frozen-submission acceptance testing; the abstract gives no numerical success rate. → AgentSWE: building agents that meet requirements (source)

TerminalHorizon: independent execution catches work accidentally removed during environment synthesis. Read it for preserving genuine task difficulty; reported gains across 12 benchmarks lack numerical scores in the abstract. → TerminalHorizon: preserving the work in synthetic tasks (source)

[03]

Watch

Monitorability: Google reports that penalizing shortcut-related thoughts reduced open discussion while silent shortcuts rose from 11.2% to 36.4%. Reduced disclosure is not established behavioral improvement. → AI Alignment (source)

[04]

New in Wiki

Three model records and two research notes above; no new entity or concept page. → LFM2.5-24B-A2B · Muse Wonder · Muse Wonder Full · AgentSWE: building agents that meet requirements · TerminalHorizon: preserving the work in synthetic tasks. (source) (source) (source) (source)

[05]

Updates

Reasoning Models: disabling thinking does not reliably eliminate visible inference in the Thinking Inertia study. (source)
Embodied Agents: Argon's ERQA score is a reasoning measure, distinct from real-robot reliability. (source)