Google Ships Gemini Decision 1: Single-Pass Decision Model at $0.10 per 1M Tokens
Latent Space · rss · 2026-11-12
At the AIE CODE summit's "System One Day," Google beat Anthropic to a production decision model, unveiling Gemini Decision 1 with Sundar Pichai and Logan Kilpatrick on stage.
Gemini Decision 1 specs
- Closed-set decisions: scores choices in a single pass instead of generating tokens; supports yes/no, multiple-choice (up to 255 options) and rubric scoring
- Accepts text, image, audio, video and PDF; up to 1,000 questions per request; 256K context
- 150–250ms end-to-end latency with a confidence score per answer
- Priced at $0.10 per 1M input tokens with $0 for output, roughly 5x cheaper than Flash-Lite per Google
- Started as a DeepMind prototype; a small team shipped model and API in about six weeks
Mixed evals
- Google's own tests: beats TypeSafe's Jev 1.14 on 5 of 7 shared benchmarks, 92.2% vs 90.0% on JevBench; Jev costs $0.042 per 1M input tokens
- Artificial Analysis is more mixed: Decision Index score of 66, tied with Elephant 9B and Jev 1.14, seven points behind Drex Max v1.1 and GLiDE v1.2; $345 per 1M tasks vs $103 for Drex Max; Brier score 0.18; 211ms average per task vs Jev's 92ms
- Early adopters: Stitch built a 5-second multi-screen product-flow experiment; LangChain shipped day-zero support
Frontier model evals
- Claude Fable 5.2 tops Agent Arena with +30.4% net improvement, 1.6x the #2 Opus 5.5 (+19.5%); API prices cut from $10/$50 to $8/$40
- Gemini 4 Carbon ranks #3 on Vals and is fastest/cheapest of the top five, but Theo measured 22m21s at low effort and 11s to first token, blaming the server side
- Meta Muse Ember posts +1.7% on Agent Arena at $0.02 per task, 4x cheaper and 12x faster than GPT-6 Luna
- Chinese models: Qwen-4-Max appears live but draws skepticism; Baidu's ERNIE 5.5 Flash Preview rises to #2 among open models on Code Arena WebDev (1668) with Apache 2.0 weights promised
AI for Science and robotics
- DeepMind's AlphaProof Code produced Lean proofs of libsignal security properties down to machine code; Gemini agents generated 6M+ lines of proofs, a five-person team finished in four weeks (vs 20 person-years for seL4)
- AlphaGenome 2 conditions variant-effect predictions on cell type and state, identifying 30% of causal variants on TenK10K (vs 16%)
- Runway released Praxis-1, a 27B MoE diffusion-transformer world action model under Apache 2.0, claiming 15.6x faster task completion and 51x lower cost than GPT-6 Astra
- Gemini Robotics 2.5 reports 30–100% higher success rates across 200 tasks
More from Models
- Fable 5.2 delivers 38% more per limit after price cut, still burns quota 2.9x faster than Opus 5.5 — FinanceYF5 · 2026-11-12
- Theo's logged requests reveal Fable burns Claude quota much faster than Opus — FinanceYF5 · 2026-11-12
- CogBattery benchmarks 54 open-weight LLMs against humans on 13 cognitive paradigms — Dengzhe Hou · 2026-11-12
- WMIP Probe Shows LLM Cumulative State Tracking Depends Heavily on Output Protocols — Dengzhe Hou · 2026-11-12
- Baidu ERNIE 5.5 Flash Preview: 280B MoE, 1M context, builds playable games from design docs — LearnWithBishal · 2026-11-12
- Community releases experimental diffusion LLM built on LFM2.5-24B-A2B — alexcovo_eth · 2026-11-12