Gemini models and capabilities

The Gemini family in 2026 spans three generations on the Developer API plus the Vertex-only enterprise mirror: the current Gemini 3.x generation (3.5 Flash GA + 3.1 Pro preview + 3 Flash preview + 3.1 Flash-Lite), the still-recommended Gemini 2.5 generation (Pro / Flash / Flash-Lite), the 2.0 line (Flash + Flash-Lite), and the 1.5 legacy line (Pro / Flash) which is on a deprecation track. The Gemini Ultra brand was retired after the 1.x era; “frontier” capability shifted to Pro at the 2.5 tier and to Flash at the 3.x tier (per ai.google.dev 2026-05, Gemini 3.5 Flash is positioned as “most intelligent model for sustained frontier performance on agentic and coding tasks”). Native modality coverage — text, image, audio, video, PDF, code — is the line’s signature differentiator: every current Gemini model accepts all six.

1. Current generation (Gemini 3.x)

Per ai.google.dev/gemini-api/docs/models and ai.google.dev/gemini-api/docs/pricing as of 2026-05-25.

Feature	Gemini 3.5 Flash	Gemini 3.1 Pro	Gemini 3 Flash	Gemini 3.1 Flash-Lite
Status	Stable / GA	Preview	Preview	Stable + Preview tracks
API ID	`models/gemini-3.5-flash`	`models/gemini-3.1-pro`	`models/gemini-3-flash`	`models/gemini-3.1-flash-lite`
Vertex AI ID	`gemini-3.5-flash`	`gemini-3.1-pro`	`gemini-3-flash`	`gemini-3.1-flash-lite`
Input price (USD / MTok)	$1.50	(preview pricing varies)	(preview pricing varies)	$0.25
Output price (USD / MTok)	$9.00	(preview pricing varies)	(preview pricing varies)	$1.50
Context window	1,000,000 tokens	1,000,000+ tokens	1,000,000 tokens	1,000,000 tokens
Max output tokens	≥ 65,536	≥ 65,536	≥ 65,536	≥ 65,536
Modalities (input)	text + image + audio + video + PDF	same + extended reasoning	same	same
Modalities (output)	text + optional structured JSON	same	same	same
Thinking control	`thinkingLevel` (minimal / low / medium / high)	`thinkingLevel` defaults to high	`thinkingLevel`	`thinkingLevel`
Caching	implicit + explicit; min 1,024 tokens	implicit + explicit	implicit + explicit	implicit + explicit
Function calling	unique `id` per call required (Gemini 3 change — per `ai.google.dev/gemini-api/docs/function-calling`)	same	same	same
Knowledge cutoff	(observed late-2025)	(observed late-2025)	(observed late-2025)	(observed late-2025)

Selection heuristic (per the docs’ positioning):

3.5 Flash — current general-purpose default. Best price/performance on agentic coding, tool use, structured tasks. Replace Gemini 2.5 Flash for most workloads.
3.1 Pro (preview) — for the hardest reasoning tasks where thinkingLevel: high is needed and 3.5 Flash isn’t enough.
3.1 Flash-Lite — high-throughput, low-cost endpoint for classification, extraction, simple Q&A. Replaces Gemini 2.5 Flash-Lite for most workloads.

2. Gemini 2.5 generation (still recommended)

Per ai.google.dev/gemini-api/docs/models and …/pricing. All stable / GA as of 2026-05-25; many production deployments are still on 2.5 because the API surface is identical and the 3.x generation hadn’t fully GA’d.

Feature	Gemini 2.5 Pro	Gemini 2.5 Flash	Gemini 2.5 Flash-Lite
API ID	`models/gemini-2.5-pro`	`models/gemini-2.5-flash`	`models/gemini-2.5-flash-lite`
Vertex AI ID	`gemini-2.5-pro`	`gemini-2.5-flash`	`gemini-2.5-flash-lite`
Input price (USD / MTok)	$1.25 (≤200k tokens) / $2.50 (>200k)	$0.30	$0.10
Output price (USD / MTok)	$10.00 (≤200k) / $15.00 (>200k)	$2.50	$0.40
Context window	1,048,576 tokens (~1M)	1,048,576 tokens	1,048,576 tokens
Max output tokens	65,536	65,536	65,536
Thinking control	`thinkingBudget` 128 - 24,576 (cannot disable)	`thinkingBudget` 0 - 24,576 (set 0 to disable)	not supported
Caching minimum	4,096 tokens	1,024 tokens	1,024 tokens
Modalities (input)	text + image + audio + video + PDF	same	text + image + PDF (no audio / video)
Free-tier rate	5 RPM (limited)	higher	higher
Knowledge cutoff (observed)	Jan 2025	Jan 2025	Jan 2025

Important pricing notch: Gemini 2.5 Pro has a >200k-token surcharge — request input over the 200,000-token watermark doubles to $2.50 / MTok input and $15 / MTok output. The watermark is on the input size, not the context cap. Workloads near 200k should consider chunking + caching to stay under, or accept the surcharge as an explicit budget item.

3. Gemini 2.0 generation (deprecating)

Feature	Gemini 2.0 Flash	Gemini 2.0 Flash-Lite
API ID	`models/gemini-2.0-flash`	`models/gemini-2.0-flash-lite`
Status	Stable, deprecating in favor of 2.5 Flash	Stable, deprecating in favor of 2.5 Flash-Lite
Input / Output	(legacy pricing, see archived docs)	(legacy)
Context	1,048,576	1,048,576
Modalities	text + image + audio + video + PDF	text + image + PDF

Migrate to 2.5 Flash / Flash-Lite (same surface, better behavior). Per ai.google.dev 2026-05, 2.0 Flash receives only critical-fix updates.

4. Gemini 1.5 legacy (retired soon)

Feature	Gemini 1.5 Pro	Gemini 1.5 Flash	Gemini 1.5 Flash-8B
API ID	`models/gemini-1.5-pro`	`models/gemini-1.5-flash`	`models/gemini-1.5-flash-8b`
Context	up to 2,000,000 tokens (Pro is the only 2M model)	1,000,000	1,000,000
Modalities	text + image + audio + video + PDF	same	same
Status	Retiring — migrate to 2.5 Pro	Retiring — migrate to 2.5 Flash	Retiring

2M context survives only on 1.5 Pro; the 2.5 and 3.x generations cap at ~1M. Workloads that genuinely need 2M (e.g. analyzing 20+ hours of video transcripts in a single call, ingesting 1500-page legal corpora) must either stay on 1.5 Pro for now or chunk + use context caching on 2.5/3.x.

5. Multimodal sister models

Distinct model series accessible from the Gemini API surface but not part of the Gemini-family chat completion path:

Family	Models	Purpose
Nano Banana	Nano Banana 2, Nano Banana Pro	Native image generation + editing (replaces / supplements Imagen)
Imagen	Imagen 4 (and predecessor lines)	High-quality image generation; $0.02-$0.06 per image
Veo	Veo 3.1, Veo 3.1 Lite	Video generation; $0.05-$0.60 per second
Lyria	Lyria 3 Pro, Lyria 3 Clip, Lyria RealTime	Music generation
Live API	Gemini 3.1 / 2.5 Flash Live	Real-time bidirectional voice + video
TTS	Gemini 3.1 / 2.5 Flash TTS	Text-to-speech variants
Agents	Deep Research, Antigravity Agent, Computer Use	Hosted agent endpoints

Per ai.google.dev/gemini-api/docs/models 2026-05.

6. Modality matrix

Every Gemini chat-tier model accepts the same six input modalities. The cell content is the model’s per-unit token consumption.

Modality	Input cost (tokens)	Notes
Text	1 token ≈ 4 chars	Standard SentencePiece-like tokenizer
Image (≤ 384 × 384 px)	258 tokens flat	Whole-image fits a single tile
Image (larger)	258 × tile_count	Image is split into 768×768 tiles
PDF (per page)	~258 tokens / page	Per `ai.google.dev/gemini-api/docs/vision`
Video (per second)	(model-dependent; ~263 tokens / frame for 1-fps sampling)	Sampled at native rate by the Files API
Audio (per second)	32 tokens / second	Per Gemini docs; ~115k tokens / hour

Practical: a 1,000-page PDF ≈ 258k tokens (fits in 1M context with 700k+ headroom). A 1-hour video ≈ 950k tokens (close to the 1M limit; chunk if longer). 9.5 hours of audio ≈ 1.1M tokens (just over 1M).

7. Pricing details beyond per-token

Free tier

Per ai.google.dev/gemini-api/docs/pricing: limited access to certain models with free input + output tokens, Google AI Studio web access, and standard rate limits. Content from free-tier requests may be used to improve Google products — explicit opt-out requires paid tier. Practical use: prototyping only.

Batch mode

50 % discount on input + output via the Batch API. JSONL input format, 24-hour SLA, max 2 GB per input file. See gemini-api-and-sdks section on batch.

Context caching

Three components:

Implicit cache (automatic on 2.5+) — best-effort, no cost-saving guarantee. Just works.
Explicit cache — client.caches.create(...), guaranteed cost reduction on hits, with hourly storage fee (~$1.00 per 1M tokens / hour).
Hit pricing — cache reads charged at a reduced per-MTok rate vs base input.

See context-caching-gemini for full mechanics.

Google Search grounding

On Gemini 3.x: per-query billing (each search the model decides to execute is one billable unit). Free tier: 5,000 grounded prompts / month on Gemini 3 models. Paid tier above free: $14 per 1,000 queries.

On Gemini 2.5 and older: per-prompt billing (any request that triggered grounding counted once, regardless of how many sub-searches happened). The 3.x change makes grounding-heavy agentic workflows materially more expensive — budget accordingly.

Code execution

No additional charge for enabling code_execution. The model is billed on:

Original prompt → input tokens
Generated code + execution result → intermediate tokens (billed as output)
Summary / final answer → output tokens

Flex inference (where supported)

50 % cost reduction with relaxed SLA. Useful for non-interactive workloads (data processing, periodic summarization) where 10-15 minute response is acceptable.

8. Benchmarks

Per Google DeepMind technical reports + Gemini 3 launch materials (2026-Q1). Treat as directional; vendor self-reports.

Benchmark	Gemini 3.5 Flash	Gemini 2.5 Pro	Gemini 2.5 Flash	Notes
MMLU	high 80s	mid 90s	low 90s	Knowledge breadth
GPQA Diamond	high 70s	high 80s	mid 70s	Graduate STEM
SWE-bench Verified	strong (positioned for agentic coding)	mid range	lower	Software engineering
MMMU (multimodal)	high 70s	high 70s	mid 70s	Multimodal reasoning
LiveCodeBench	leader	strong	mid	Live competitive programming

Per ai.google.dev model cards 2026-05; the DeepMind Gemini technical report carries audited numbers.

9. Endpoint types

Endpoint	Hostname pattern	Auth	Best for
Gemini Developer API (AI Studio)	`generativelanguage.googleapis.com`	API key (`x-goog-api-key`)	Prototyping, indie, sub-enterprise
Vertex AI (enterprise)	`aiplatform.googleapis.com` (regional)	GCP IAM service-account / ADC	Production, compliance, private VPC, audit logs
Vertex AI Express Mode	regional	API key (simpler, no IAM)	On-ramp from Developer API to Vertex without full IAM setup

The same model IDs work on both, but Vertex IDs typically drop the models/ prefix (e.g. gemini-2.5-pro vs models/gemini-2.5-pro). See vertex-ai-enterprise-gemini for the full delta.

10. Programmatic model discovery

from google import genai
client = genai.Client(api_key="...")
for m in client.models.list():
    print(m.name, m.display_name, m.input_token_limit, m.output_token_limit,
          m.supported_actions)

REST:

curl "https://generativelanguage.googleapis.com/v1beta/models?key=$GEMINI_API_KEY"

The supported_actions field tells you which methods the model accepts (generateContent, streamGenerateContent, embedContent, countTokens, batchEmbedContents, cachedContents, etc.). Use this for capability-aware routing rather than hardcoding lists.

11. Migration guide

Quick reference for choosing the migration target:

From	Recommended target	Why
Gemini 1.5 Pro	Gemini 2.5 Pro	Same context (1M is enough for most cases); much better quality + price; same API surface. Stay on 1.5 Pro only if you genuinely need 2M context.
Gemini 1.5 Flash	Gemini 2.5 Flash	Strict upgrade
Gemini 2.0 Flash	Gemini 2.5 Flash	Strict upgrade; same modality matrix
Gemini 2.5 Flash	Gemini 3.5 Flash	Better agentic + coding; 3.5 has `thinkingLevel` semantics — adapt any code that sets `thinkingBudget` directly
Gemini 2.5 Pro	Gemini 3.1 Pro (when preview lifts)	Wait for GA unless preview is fine for your use case

12. Further reading

Models — canonical model list with status
Pricing — per-model and per-feature pricing
Rate limits — tier-based RPM / TPM / RPD caps
DeepMind Gemini — technical reports + model cards
Vertex AI Model Garden — partner + open model catalog on Vertex
Gemini Enterprise Agent Platform — the new home for Vertex GenAI docs

Compendium

Explorer

Gemini models and capabilities

Gemini models and capabilities

See also

1. Current generation (Gemini 3.x)

2. Gemini 2.5 generation (still recommended)

3. Gemini 2.0 generation (deprecating)

4. Gemini 1.5 legacy (retired soon)

5. Multimodal sister models

6. Modality matrix

7. Pricing details beyond per-token

Free tier

Batch mode

Context caching

Google Search grounding

Code execution

Flex inference (where supported)

8. Benchmarks

9. Endpoint types

10. Programmatic model discovery

11. Migration guide

12. Further reading

Graph View

Table of Contents