AI research wiki
llms-foundation-models · 2026-10-29 · Tier 2

Gemini 4 Carbon and GPT-6.1 Terra: a frontier release and a cheap rung

Gemini 4 Carbon and GPT-6.1 Terra: a frontier release and a cheap rung

Sources: Google DeepMind, Gemini 4 Carbon (lab watcher), The Decoder, Arena, AI Breakfast on GPT-6.1 Terra, Interconnects on AA v5. raw: raw/labs/2026-10-28-230451-labs.md, raw/gmail/2026-10-29-newsletters.md.

TL;DR.

  • Gemini 4 Carbon (two weeks after Gemini 4 Argon) ranks #1 on Artificial Analysis Intelligence Index v5 at 59, #1 in Text Arena, #2 in Agent Arena (+16.2% net, on the cost/time/tokens Pareto frontier) and #4 in WebDev Arena (+89 Elo). Cyber: 88.2% SEC-bench Pro (finding vulns), 54.8% ExploitGym (exploitation), 71.0% CWE-bench (patching), all evaluated without guardrails. Safety claims: cheating on Terminal-Bench 4.0 reduced to 0.0% (0.4% attempts), ~80% fewer high-severity misalignment alerts in internal coding, ~2x fewer unsupported claims than Argon, overreach -26% and under-service -29%, 0.0% attack success on Gray Swan's indirect prompt injection benchmark. Access starts with cyber defenders via the Fairwind program, with US pre-release government testing. Price: $2/$10 per million input/output tokens introductory (95% off cached input), then $4/$20.
  • GPT-6.1 Terra (OpenAI mid-tier, visual reasoning and computer use): $1/$5 per million tokens ($0.05 cached), more than 3x faster than GPT-6 Astra and GPT-6.1 Sol. Matches Astra on Agents' Last Exam at medium effort at ~1/10 the per-token price; Terminal-Bench 4.0 50.3% (from 21.5% for GPT-5.6 Terra); factual error rate 4.3% at max effort. OpenAI's Ted Sanders confirmed the size ladder: Luna < Terra < Sol < Astra.

Relation to prior wiki

  • Cost lens: both launches push the price-performance frontier from opposite ends. Carbon is priced like Argon despite the jump; Terra makes the cheap rung of an effort/model router (see effort escalation) far stronger.
  • Benchmark caveat: Interconnects argues AA v5 tilts toward frontier "general agentic collaborator" tasks and private test sets, which widens the open-closed gap (Kimi K3.1 Pro fell five places) and puts total trust in the benchmark builder.
  • Distillation pressure: The Information reports DeepSeek and Moonshot relied heavily on GPT-6 Astra outputs after Anthropic's limits cut distillers' harvest from Claude by 90-99%. Frontier releases now double as training data for open models.
  • Cyber gating matches the 10-28 AISN "Black October" thread: Carbon is the first frontier model to launch to defenders first.