Comprehensive October 2026 technical reference on Google DeepMind & Gemini Enterprise Agent Platform models: Gemini 4 Argon (periodic-element frontier naming, unified 1M token window, long decodes, implicit-only caching, Fairwind cyber tier), the Gemini 3.x lineup (3.8 Flash, 3.5 Flash-Lite, 3.8 Live, 3.8 Flash TTS, 3.5 Transcribe), Model Garden & MaaS vs. self-hosted serving, multi-model agentic routing, Interactions API vs. GenerateContent, generative media (Nano Banana 2.1, Gemini Omni 1.1 Flash, Veo 3.1, Lyria 3.5), Chirp 3 & Gemini Embedding 2, Gemma 4, SynthID & Frontier Safety Framework, and agentic discovery systems (AI Co-Scientist, AlphaEvolve, CodeMender, AlphaFold 3).
Google’s Gemini foundational models are built on native multimodality (early fusion). Instead of bolting separate vision or audio encoders onto a pretrained text-only LLM, all modalities — text, code, images, audio, video, and PDFs — project into a unified token space during pretraining and pass through a single Sparse Mixture-of-Experts (MoE) transformer backbone. Conditional routing activates only a specialized subset of expert networks per token, decoupling massive total knowledge capacity from active per-token FLOPs and Time-to-First-Token (TTFT) (replacing legacy dense PaLM 2).
ASR → LLM → TTS), and causes text-dominant hallucinations when modalities conflict.Choose native multimodal (gemini-3.8-live, gemini-4-argon GA agentic video understanding) when sub-second conversational latency, prosody/visual nuance, or dynamic video zooming is critical. Choose a modular perception pipeline (Chirp 3 / gemini-3.5-transcribe or Document AI Layout Parser → Sensitive Data Protection PII redaction → text LLM) when compliance requires auditable verbatim transcripts, deterministic PII scrubbing before reasoning, custom domain phrase adaptation, or caching extracted text across many downstream passes.
Starting with Gemini 4, Google transitioned frontier model naming to periodic-table elements, led by Gemini 4 Argon (gemini-4-argon, GA October 13, 2026) and Fairwind-allowlisted Gemini 4 Argon Cyber Permissive (gemini-4-argon-cyber-permissive) (note: periodic naming starts with Argon — there is no "Gemini 4 Carbon" model).
gemini-3.5-flash-lite) engineered for front-door intent routing, inline guardrails, entity extraction, classification, and bulk enrichment.gemini-3.8-flash, plus 3.7 / 3.6 / 3.5 Flash) trained via knowledge distillation for long-horizon coding, rapid MCP tool calling, and default CodeMender execution.gemini-3.8-live & gemini-3.8-live-extended-thinking for sub-second voice/video agents; gemini-3.8-flash-tts (130 languages) & flash-lite-tts; gemini-3.5-transcribe & 3.5-live-translate-preview.gemini-4-argon delivers peak System-2 reasoning across unified 1M context, long decodes up to 1M tokens, GA agentic video understanding, computer use, formal math/science, legal/finance, and defensive cyber.antigravity-preview-09-2026 (isolated Linux coding/browsing sandbox), deep-research-preview-04-2026 / deep-research-max, and CodeMender.Lifecycle status: Gemini 2.5 (2.5 Pro, 2.5 Flash, 2.5 Flash-Lite) remains in legacy support for existing workloads and managed SFT/preference tuning on Agent Platform, while new projects target 3.5 Flash-Lite, 3.8 Flash, and 4 Argon. Gemini 2.0 Flash / Flash-Lite, gemini-2.5-computer-use-preview-10-2025, and Gemini Robotics ER 1.6 are shut down.
Agent Platform Model Garden (formerly Vertex AI Model Garden) is the single enterprise catalog of 200+ Google first-party, partner MaaS, and open-weight models governed by unified IAM, VPC-SC, org policies, and billing.
| Family | Active Examples | How You Consume It | What You Pay For |
|---|---|---|---|
| First-party (Google) | Gemini 4 Argon, Gemini 3.8 Flash, 3.5 Flash-Lite, 3.8 Live, Nano Banana 2.1, Gemini Omni 1.1 Flash, Veo 3.1, Lyria 3.5, Chirp 3, Gemini Embedding 2 | Fully managed serverless API only | Per token / media unit (PayGo) or reserved capacity (Provisioned Throughput GSUs) |
| Partner (MaaS) | Anthropic Claude, Mistral AI, DeepSeek (e.g., DeepSeek-V4.1-Flash 1M context preview) | Managed serverless MaaS API inside Google Cloud boundary | Per token (PayGo) or Provisioned Throughput where offered |
| Open-weight | Gemma 4 (E2B, E4B, 26B A4B MoE, 31B), Gemma 3/3n, MedGemma, ShieldGemma 2, Llama, Qwen | Either curated serverless MaaS API or self-deployed on dedicated Agent Platform endpoints / GKE | MaaS: per token. Self-deployed: VM + GPU/TPU compute for every hour deployed |
AnthropicVertex): Serverless API inside Google Cloud's security boundary with IAM, VPC-SC, regional/global endpoints, Cloud Audit Logs, and GCP billing.| Criterion | MaaS (Managed API) | Self-Deployed Agent Platform Endpoint | Self-Deployed on GKE |
|---|---|---|---|
| Ops burden | None. Google autoscales accelerators. | Low–medium. Prebuilt vLLM / TGI / SGLang containers, autoscaling, idle undeploy. | High. You operate Kubernetes, node pools, and serving stacks. |
| Custom / LoRA weights | Base catalog checkpoints only | ✅ Custom weights, LoRA adapters, custom containers | ✅ Full control + dynamic multi-LoRA serving |
| Traffic shape | Bursty, unpredictable, prototyping | Steady high volume (lower TCO at sustained QPS) | Steady high volume; many models packed across shared GPU/TPU clusters |
| Billing trap | Pay strictly per token | Billed while deployed, even at 0 QPS (set idle TTL or undeploy) | Billed for provisioned nodes / reservations |
Model choice solves for: matching each agent node to the cheapest option that still clears its quality, latency and compliance gates. Three axes move independently: scale (frontier LLM vs. efficient SLM tier), serving (managed API vs. self-hosted capacity) and weights (proprietary vs. open).
| Option | Data residency | Single-tenant compute in your network | Latency | Cost model | Customization | Best agent role |
|---|---|---|---|---|---|---|
Frontier managed LLM (gemini-4-argon) | Regional, us/eu multi-region, or global (no residency guarantee) | No: multi-tenant Google API, reached privately through VPC-SC and Private Google Access / PSC for Google APIs | Highest (thinking tokens add time before output) | Per token (PayGo tiers) or Provisioned Throughput GSUs | Prompting, grounding, implicit caching; no managed tuning on Argon | Coordinator, planner, ambiguous tool selection, escalation path |
Efficient managed tier (gemini-3.8-flash, gemini-3.5-flash-lite) | Same endpoint choices as above | No: same managed API path | Low; Flash-Lite lowest TTFT in the Gemini lineup | Lowest per-token rate; PT, Flex and Batch available | Prompting, caching, SFT/LoRA on eligible versions | Tool-loop worker, router, classifier, extractor, inline guardrail |
| Partner MaaS (Claude on Agent Platform) | Regional, multi-region or global depending on the model card | No: serverless partner endpoint governed by Google Cloud IAM, VPC-SC and audit logs | Medium to high (extended thinking adds time) | Per token; PT ordered through a Google Cloud account team | Prompts and prompt caching; no weight access | Long-horizon coding or synthesis worker where the model is already standardized |
| Self-hosted open weights, Model Garden endpoint (Gemma 4 on a dedicated private endpoint) | Single region you pick for the endpoint | Yes: dedicated endpoint exposed only through a PSC endpoint in your VPC | Low to medium (set by accelerator shape and replica count) | Node-hours while deployed, even at zero traffic | Full weights, LoRA or full fine-tunes, custom vLLM containers | Specialist model with tuned open weights, served privately by a team that does not want to run Kubernetes |
| Self-hosted open weights, GKE + vLLM (Gemma 4 on GPUs or TPUs) | Region or zone of your node pools | Yes: pods in your private cluster; weights can load from your own storage instead of a public hub | Lowest achievable when tuned (continuous batching, prefix caching) | Node-pool or reservation cost; cheapest per token only at sustained high utilization | Full control of weights, quantization, serving flags, multi-LoRA | High-QPS fleet of fine-tuned specialists, strict isolation, existing GKE platform team |
Latency and cost columns are directional rankings, not benchmarks. Validate with your own load test and golden evaluation set before committing to capacity.
Isolation note: VPC-SC and PSC keep managed-API traffic private and block exfiltration; inference still runs on Google's shared serving stack. Only self-hosted endpoints run on capacity dedicated to you.
Tuned Gemini caveat: tuned Gemini models deploy only to shared public endpoints, not to PSC endpoints, so a "fine-tuned and PSC-only" requirement points to an open model on a private endpoint or GKE.
| Endpoint | Where ML processing runs | Gains | Gives up |
|---|---|---|---|
Regional (us-central1, europe-west4, ...) | The named region | Strict residency; none of the global feature gaps; CMEK-protected regional resources | Lower burst headroom than global; more 429 risk under regional spikes |
Multi-region (us, eu) | Anywhere inside that jurisdiction | Jurisdiction-level residency with better availability than one region | Not single-region; CMEK unavailable for Agent Runtime on multi-region; Private Google Access not supported (use PSC endpoints or backends) |
Global (global) | Any supported Google Cloud location | Highest availability, fewer 429 errors | No residency guarantee; no tuning, no RAG corpus, no batch for Anthropic and open MaaS models; CMEK unavailable for Sessions and Memory Bank |
genai.Client(vertexai=True, project="my-proj", location="eu"); "eu" keeps ML processing in the EU, while "global" offers no residency guarantee.global first, then pick regional (single region) or multi-region (jurisdiction only).global location cannot hold CMEK-protected resources. Pair CMEK mandates with regional endpoints.constraints/vertexai.allowedPartnerModels limits which partner models a project can call.Open weights inside a public endpoint are no more isolated than a managed API; the isolation comes from the private endpoint, perimeter and egress controls around them.
Golden rule: pick the smallest, fastest, cheapest model that passes your golden evaluation set, and route dynamically across tiers per agent step rather than sending every turn to a frontier model.
Use Gemini 3.5 Flash-Lite (or ShieldGemma 2) at the front door for sub-200ms intent routing, slot extraction, and guardrails → run your core MCP tool-calling and coding loops on Gemini 3.8 Flash → escalate only hard multi-step planning, conflict resolution, or repository synthesis to Gemini 4 Argon (or enable Vertex AI Model Optimizer to route automatically across Gemini tiers by cost/quality target).
-latest; retired IDs return 404).| Customization Rung | Use When | Key Rule & Gemini 4 Constraint |
|---|---|---|
| 1. Prompt & Context Engineering | Baseline instructions, JSON schemas, dynamic few-shot (Example Store) | Start here; automate prompt iteration against evals with Agent Optimizer |
| 2. Grounding / RAG | Missing, private, or frequently changing knowledge + citations | Fixes what the model knows without retraining (Agent Search, RAG Engine, Search Grounding) |
| 3. Supervised Fine-Tuning (SFT) | Consistent behavior: strict syntax, domain jargon, replacing long few-shot prompts | Parameter-efficient LoRA adapters on eligible Gemini 2.5/3.x & Gemma models (not supported on gemini-4-argon; tuned models also cannot use explicit caching) |
| 4. Preference Tuning | Nuanced tone/quality trade-offs hard to label as single targets | Train on (prompt, preferred, dispreferred) pairs after SFT |
| Consumption Option | Best For | Key Behavior |
|---|---|---|
| Standard PayGo | Default variable traffic | Shared pool; subject to 429 contention during spikes |
| Priority PayGo | Latency-critical bursts without term commitment | Per-request priority header on shared capacity (supported on gemini-4-argon, 3.8-flash, nano-banana-2.1); absorbs PT overflow |
| Flex PayGo / Batch | Latency-tolerant evals, labeling, overnight bulk pipelines | Discounted token rates (Batch offers deepest async discount) |
| Provisioned Throughput (PT) | Production workloads needing guaranteed QPS/TPM & p95 SLA | Reserve non-cancelable GSUs (1 wk–1 yr, same publisher). Spills over to PayGo by default; set X-Vertex-AI-LLM-Request-Type: dedicated to cap spend strictly or shared to bypass PT |
A multi-agent design solves for: paying frontier prices only on the turns that need frontier reasoning. The coordinator owns ambiguity; workers own narrow, typed subtasks. A typical tree puts gemini-4-argon at the top (plans, delegates, merges) over a 3.5 Flash-Lite router or guard, a 3.8 Flash tool-loop worker, and a Gemma 4 + LoRA domain specialist on a private endpoint.
LlmAgent sets its own model, so a coordinator and its sub-agents can run different tiers in one agent tree.RoutedLlm wraps several models behind one agent and calls a router function per request (fallback on error, A/B tests, routing by input complexity); RoutedAgent does the same when instructions and tools also change. If the chosen model fails before emitting output, the router is called again with error context.| Role | Typical model tier | Why |
|---|---|---|
| Coordinator / planner | Frontier (gemini-4-argon) | Decomposes vague goals, picks among many tools and sub-agents, resolves conflicts |
| Router and inline guardrail | Flash-Lite or ShieldGemma 2 | Runs on every request, so TTFT and per-call cost dominate |
| Tool-loop worker | Flash (gemini-3.8-flash) | Many short MCP or function-call turns at high token velocity |
| Fine-tuned specialist | Gemma 4 with LoRA, self-hosted | Fixed schema or domain syntax; weights and data stay on private capacity |
| Dimension | Managed API, PayGo | Managed API, Provisioned Throughput | Self-hosted (Model Garden endpoint or GKE) |
|---|---|---|---|
| Unit of spend | Tokens consumed | GSUs reserved for a term | Accelerator node-hours |
| Idle cost | None | Full commitment, used or not | Full while replicas run |
| Burst behavior | Shared pool, 429 risk under spikes | Reserved throughput; overage spills to PayGo unless dedicated | Limited by your replicas and autoscaling speed |
| Ops owner | You (containers, scaling, patching, guardrails) | ||
| Model scope | Gemini, partner and open MaaS catalog | Supported Gemini and partner models (partner PT ordered through Google Cloud sales) | Any open or custom weights you can serve |
PT vs. self-hosting: both buy predictability. PT keeps the managed model and its update path; self-hosting buys weight control and private-only execution at the price of running the serving stack.
Anti-pattern: self-hosting a large open model for spiky, low-volume traffic. Idle node-hours usually cost more than PayGo tokens; start on MaaS and move once utilization is sustained.
us or eu endpoint: residency is defined at jurisdiction level and availability matters more than single-region pinning.gemini-4-argon) Architecture & Strict Constraints1,048,576 tokens): Shares a single 1M-token pool across input prompt + internal thinking + function calls + generated output.65,536 tokens (32,768 non-streaming HTTP), extendable up to the full 1,048,576 budget (minus input) via continuation_token on GenerateContent (v1beta1) or server-side background=true on the Interactions API.LOW | MEDIUM | HIGH): Defaults to MEDIUM. Passing legacy integer thinking_budget OR thinking_level="MINIMAL" on gemini-4-argon returns HTTP 400 INVALID_ARGUMENT! (Note: MINIMAL is also rejected on gemini-3.8-flash and gemini-3.7-flash, but is supported on gemini-3.6-flash and gemini-3.5-flash*; integer thinking_budget belongs to Gemini 2.5, where 2.5 Pro minimum is 128 tokens.)temperature, top_p, and top_k values are ignored on gemini-4-argon.cachedContents, Managed Tuning, Live API WebSockets, and direct RAG Engine generator binding are NOT supported on gemini-4-argon. Supports GA Agentic Video Understanding, Computer Use, Code Execution, MCP/Function Calling, and Grounding.gemini-4-argon-cyber-permissive): Allowlisted via DeepMind's Fairwind Program (global & us) for defensive vulnerability research, reverse engineering, and CodeMender; defaults to HIGH thinking and disables Computer Use.GenerateContent vs. Interactions API & CachingGenerateContent: Stateless client-managed turns (pass history + opaque Thought Signatures across tool loops; enable include_thoughts=True for UI summaries).Interactions API (client.interactions.create, Preview): Optional server-side state via previous_interaction_id + background=true for long-horizon decodes and Managed Agents.cachedContents): Available on Gemini 2.5 and supported 3.x base models (60-min default TTL + hourly storage fee); disable all caching when strict Zero Data Retention (ZDR) applies.gemini-nano-banana-2.1, GA + Priority PayGo): Flagship native Gemini image generator & mask-free multi-turn conversational editor (replaces retired standalone Imagen 4 on the developer Gemini API).gemini-3.1-flash-image), 2 Lite (gemini-3.1-flash-lite-image) & Pro (gemini-3-pro-image): 131,072-token context, up to 4K resolution, Google Search-grounded typography/infographics, and up to 14 reference images on Pro (5 character + 6 object + 3 style) for multi-turn identity preservation.gemini-omni-1.1-flash): Fast native multimodal video generation, instruction-based video editing, keyframe interpolation, and clip extension with native audio.Generate, Fast, Lite — GA on Agent Platform): Joint latent diffusion synthesizing 4K 16:9 / 9:16 video and synchronized lip-sync/Foley audio in one shared latent space with strong object permanence (succeeding silent Veo 2).lyria-3.5) for full-length structurally coherent songs, Lyria 3 Pro (184s), Lyria 3 Clip (30s), Lyria RealTime (lyria-realtime-exp), and Lyria 2 — accepting text/image prompts, BPM/structure controls, and multilingual vocals with SynthID-Audio watermarking.gemini-3.8-live): Default stateful WebRTC/WebSocket native speech-to-speech & video agent (16-bit PCM in → 24 kHz audio out) with server-side VAD barge-in, affective dialogue, and spoken function calling (71.5%+ ComplexFuncBench Audio).gemini-3.8-live-extended-thinking): Runs deep background reasoning asynchronously without stalling live spoken conversation.chirp_3): Built on the 2B+ parameter Universal Speech Model (USM) (12M+ hours, 100+ languages) for compliance ASR pipelines requiring StreamingRecognize / BatchRecognize, Model Adaptation (up to 1,000 custom phrases), denoise_audio=True, and word-level timestamps.gemini-embedding-2) & gemini-embedding-001: gemini-embedding-2 is Google's first native multimodal Gemini embedding model, mapping text, code, images, audio, video, and PDFs into one shared space. Both use Matryoshka Representation Learning (MRL) for truncatable L2-normalized dimensions (e.g., 3,072 → 768 / 256).E2B, E4B, 26B A4B MoE, 31B): Flagship open-weight family combining mobile-first Per-Layer Embeddings (PLE) on E2B/E4B (processing text, image, video, and native audio in a 2B–4B RAM envelope), an ultra-fast 26B Sparse MoE activating only 4B parameters per token (26B A4B) for self-hosted agentic loops, and a dense 31B checkpoint.T5Gemma) and hybrid Real-Gated Linear Recurrence (RG-LRU) models achieving constant fixed-size state memory on long edge streams.Paired with C2PA Content Credentials, SynthID embeds imperceptible statistical/cryptographic signatures directly into generated payloads so provenance survives metadata stripping:
gemini-4-argon-cyber-permissive requires Fairwind Program vetting with Computer Use disabled), Autonomous ML R&D, and Deceptive Alignment.SafetySetting thresholds + Vertex AI Model Armor + ShieldGemma 2.gemini-3.8-flash and gemini-4-argon (cm find / verify / fix) for vulnerability discovery, sanitizer verification, and automated patching.gemini-robotics-er-2-preview, succeeding retired ER 1.6 & RT-2) for 3D spatial pointing, grasp planning, and physical robot safety.Genie) and generalist 3D virtual game/environment agents (SIMA 2).GraphCast) and 15-day probabilistic diffusion ensemble (GenCast) global weather and extreme-event forecasting.us / eu)RoutedLlm / RoutedAgentTest your mastery of Gemini 4 Argon mechanics, Gemini 3.8 Flash/Live/TTS, multi-model agentic routing, Interactions API, Model Garden MaaS vs. self-hosted serving, Provisioned Throughput, Nano Banana 2.1, Gemini Omni 1.1 Flash, Veo 3.1, Lyria 3.5, Chirp 3, Gemini Embedding 2, Gemma 4, SynthID, Frontier Safety Framework, AI Co-Scientist, AlphaEvolve, and AlphaFold 3.