Grok 4.8 Rumored at 2.5T Params on New C++ Stack as Gemini 3.8 Live Slugs Surface
testingcatalog · x · 2026-09-15
TestingCatalog's daily AI brief (Sept 15) rounds up unverified leaks and industry news:
Model forecast
- Gemini 3.8 Live and Live Extended Thinking slugs appeared on the Cloud quota page; not yet public.
- Grok 4.7 should land this week near Opus 5.0 level, not 5.1; multimodal still needs work.
- Testers report some Claude Code Opus 5 traffic routing to Opus 5.2 — faster and less lazy, unofficial.
- Grok 4.8 is rumored to be a 2.5T model on a new C++ stack; training wraps this week, then RL.
- Later: Grok 4.9 "probably Astra/Fable class"; Grok 5 "maybe better than anything."
AI slowdown debate: Dario Amodei argues for pacing the frontier rather than halting it — embedded evaluators, lab coordination, then a global deal. Elon Musk endorsed pre-release peer review of rival models on All-In, saying "Dario is right." Sam Altman agrees and will match evaluators, stressing pacing means slower, not stopped. Demis Hassabis says Amodei's essay "points towards the right path."
All leaks are unofficial.
More from Models
- Claude's hidden text watermark: how invisible fingerprints detect AI writing — Two Minute Papers · 2026-09-15
- User claims GPT 5.6 Luna cost $0.30 over 14 days, 3x cheaper than GPT 5 — Spiritual_Grape3522 · 2026-09-15
- Ant open-sources SingProbe, a runtime guardrail reading risk from hidden states at <0.5% overhead — aigclink · 2026-09-15
- Codex users report built-in deep-research skill vanishing from CLI overnight — Hot_Independence5160 · 2026-09-15
- Community-built "AI Race" page tracks which models lead in each category — ohhtthatguy · 2026-09-15
- DeepSeek says frontier models now lead open source by 18 months, up from 2 at R1's launch — Select_Dream634 · 2026-09-15