13 models · on the curve since 2023

Google

Value position · the site’s read Architecture layer

Google’s durable position is the distribution it already owns: Search, Android, Workspace, and the TPU stack the models run on. Gemini compounds because it ships into surfaces a billion people open every day.

What would change this readA rival captures a primary surface, the default assistant on phones or in the browser, or the TPU advantage stops translating into cheaper frontier training.

Recent cadence
6 in last 12 mo
4 the year before
Landmark density
5 / 13
defining releases
Weights
Open + closed
Gemini 3.5 flagship

Founded DeepMind 2010; merged 2023Mountain View / London

The substrate underneath figures as of Aug 2026

Trains and serves on

Its own silicon in its own buildings — the only lab on this page where that is true end to end. Current generation is Ironwood, the seventh-generation TPU. The material change is that Google has started SELLING that silicon into customers' own datacentres rather than only renting it: revenue recognition began in Q2 2026 and the company expects the bulk in 2027, with Anthropic and Meta among the named takers.

ConcentrationIts exposure is not to a chip vendor at all. Google designs the training silicon and owns the datacentres, so what it depends on is TSMC for fabrication and Broadcom and MediaTek as design partners for the next generation.

Flagship, per million tokens

$2.00 in $12.00 out

Gemini 3.1 Pro, prompts to 200k

Was

$3.50 / $10.50

Gemini 1.5 Pro, prompts to 128k · Aug 2024

Weights
Partial. Gemini is closed; the Gemma 4 family ships open weights under Apache 2.0, a change from the earlier custom Gemma terms.
Where you buy it
Gemini API and AI Studio · Vertex AI and Gemini Enterprise · Gemini consumer app · Google Workspace
Structure
Alphabet Inc., a Delaware public company with a triple-class structure. Class B super-voting shares held by the founders carry majority voting control against minority economics.
Revenue
AI revenue is not disclosed separately. Alphabet turned over $119.8B in the quarter to 30 Jun 2026, of which Google Cloud was $24.8B, up 82%. Models were processing about 22 billion API tokens a minute.
People
198,933 across Alphabet (30 Jun 2026); Google DeepMind is not broken out

Biggest single dependencyAlphabet is spending $195-205B of 2026 capex against an AI revenue line it does not disclose, having raised about $70B of new equity and debt in a single quarter while quarterly free cash flow turned negative.

The price comparison is the one on this site that most clearly refuses the simple story: flagship input fell about 43% in two years while output ROSE about 14%. Naming is genuinely confusing right now — Google calls Gemini 3.6 Flash its latest model while the top Pro tier is still numbered 3.1 and marked preview, so the flagship figure above is 3.1 Pro. DeepMind headcount is omitted because third-party estimates span a 35% range with no primary disclosure.

On the durability curve frontier era
2023202420252026 Gemini 1.0Gemini 3.6 Flash ARCHITECTURE the substrate · compounds  → OUTPUT per model · commoditises  → CROSSOVER the durability thesis  → DURABLE VALUE-CAPTURE ↑ the margin a release keeps · not raw capability TIME · FRONTIER MODEL RELEASES →
Output · value a model keeps Architecture · the compounding substrate Google release

Hover any point to read the release. The y-axis is qualitative; only the x-axis carries dates.

When to reach for them practical read
Massive context
Reach for Gemini when you need very long context in one pass (1.5 Pro, 1M tokens, Feb 2024).
Multimodal reasoning
Pick 2.5 Pro for thinking-model quality across text, image and audio (Mar 2025).
In the Google stack
Use it where the work already lives in Workspace, Android or Cloud (Gemini 1.0, Dec 2023).
The landmark releases 5 of 13 models
  1. May 2026 Gemini 3.5 / Omni I/O launch, any-to-any modality OmniFlagship
  2. Nov 2025 Gemini 3 Pro Flagship launch, frontier reasoning FlagshipMultimodal
  3. Mar 2025 Gemini 2.5 Pro Thinking model, top of leaderboards ReasoningThinking
  4. Feb 2024 Gemini 1.5 Pro Million-token context, MoE architecture 1M ctxMoE
  5. Dec 2023 Gemini 1.0 Brand launch, replaced Bard MultimodalLaunch
Full release logall 13, newest first
  1. 2026
  2. Jul Gemini 3.6 Flash Flash-tier refresh; Flash-Lite adds built-in computer use. No Pro-tier model shipped this cycle Flash tierCurrent
  3. May Gemini 3.5 / Omni I/O launch, any-to-any modality OmniFlagship
  4. Feb Gemini 3.1 Pro Preview update Preview
  5. 2025
  6. Dec Gemini 3 Flash New app default tier Fast tier
  7. Nov Gemini 3 Pro Flagship launch, frontier reasoning FlagshipMultimodal
  8. Aug Nano Banana (2.5 Flash Image) Viral image generation and editing ImageEditing
  9. Jun Gemini 2.5 Flash GA fast reasoning tier ReasoningFast
  10. Mar Gemini 2.5 Pro Thinking model, top of leaderboards ReasoningThinking
  11. Feb Gemini 2.0 Pro Experimental release Experimental
  12. 2024
  13. Dec Gemini 2.0 Flash Agentic-era multimodal default AgenticMultimodal
  14. May Gemini 1.5 Flash Fast, lightweight tier Fast tier
  15. Feb Gemini 1.5 Pro Million-token context, MoE architecture 1M ctxMoE
  16. 2023
  17. Dec Gemini 1.0 Brand launch, replaced Bard MultimodalLaunch
Where they sit on weights versus the field
The durability take on Google from the archive
About Google at a glance
Founded
DeepMind 2010; merged 2023
HQ
Mountain View / London
Leadership
Demis Hassabis
Parent
Alphabet
Flagship
Gemini 3.5
Weights
Closed (Gemma open)
Also builds
Search, Workspace, TPUs
Known for
Gemini, AlphaFold

Google’s AI is led by Google DeepMind, formed in 2023 by merging DeepMind (founded 2010, London) with Google Brain. Headquartered in Mountain View with major research in London, it produces the Gemini multimodal family as closed-weight models, with the smaller Gemma weights open.

The frontier chart → Read the archive →