Google Gemini for Automation

Key takeaways

As of 12 October 2026:

• Gemini 3.8 Flash is the production default, and its price doubles on 1 January 2027. Generally available (GA) since 2 September 2026, it costs $0.75 input and $3.75 output per 1M tokens and takes 1M tokens of input. From 1 January 2027 it costs $1.50 / $7.50, and cached input doubles too, from $0.075 to $0.15. Google's model card calls $1.50 / $7.50 the "regular" price, so the 2026 rate is a launch discount.
• Google's most capable models are gated. Gemini 4 Argon (announced 30 September 2026; $2 / $10 introductory, then $4 / $20) goes to vetted cyber defenders first. The only API Pro model is still gemini-3.1-pro-preview, in preview since 19 February 2026.
• The open parts of the agent stack need no contract. Next '26 rebuilt Vertex AI as the Gemini Enterprise Agent Platform. ADK (Apache-2.0, v2.11.0), A2A (Linux Foundation, 150+ organisations) and AP2 (donated to the FIDO Alliance on 28 April 2026) are free; Gemini Enterprise seats are quote-only.
• Developer and computer-use tools lag. Gemini CLI had 1.74M npm downloads in a month, against 56.3M for Claude Code and 98.0M for Codex. Computer use has been in public preview since 24 June 2026.
• Google's benchmark claims are its own, but independent boards put Gemini in the leading group. Gemini 4 Argon was a preliminary #1 on LMArena's 8 Oct 2026 text snapshot (1525 ±9) and scores 52.6 on Artificial Analysis's index, against 57.6 for Opus 5.5. Google puts Argon #1 on Zapier's AutomationBench (51.3%, vendor-reported), and Arena's agent list puts it #7. On the Terminal-Bench 4.0 board, 3.8 Flash scores 19.09% (mini-SWE-agent harness), against 64.85% for Opus 5.5, 61.82% for Sonnet 5.5 and 58.18% for GPT-6 Astra (max effort, vendor CLIs); Argon is not on that board. Anthropic's own 70.6% for Sonnet 5.5 is not reproduced there.
• Scale is large. Alphabet reports 950M Gemini app users and 22B API tokens a minute (Q2 2026). Google Cloud revenue grew 82% to $24.8B at a 35.6% margin.
• TPUs are now sold outside Google. Anthropic's October 2025 deal covers up to 1M TPUs, and its April 2026 deal adds multiple gigawatts of TPU capacity from 2027. TPU-system sales began in Q2 2026; Meta was only in talks.
• India has a big hub but little in-country processing. The Visakhapatnam AI hub broke ground around 28 April 2026. Google Cloud's data-residency page lists India (asia-south1, Mumbai) ML processing only for Gemini 3.5 Flash, Gemini 2.5 Flash (128K) and text-embedding-005. Gemini 3.6, 3.7 and 3.8 Flash are US and EU multi-region only, and Delhi is not listed. The Gemini app supports 13 of India's 22 scheduled languages, and Chirp 3 speech-to-text is GA for Hindi and Indian English.

How this page labels evidence

This page is current as of 12 October 2026. It draws on research collected on 11 October 2026; residency, benchmark, pricing-label and language-support facts were re-checked against primary sources on 12 October 2026. The main primary sources are Google's Gemini API pricing page (last updated 9 October 2026), the API changelog, the Cloud Next '26 wrap-up and Alphabet's SEC filings. Some claims carry labels. Vendor-reported means Google's own claim, not independently checked. Secondary means the claim comes from news, Wikipedia or an aggregator, and single-source means only one source was found. Snippet marks a claim seen only in a search-result summary, and HN listing one where only a Hacker News title and date were seen. Where earlier and later checks disagree, the later check against primary sources wins, and the conflict is shown. Ten decision-critical facts were re-read on the live primary pages on 11 October 2026 and are marked "re-read 11 October 2026". Related pages: Claude for automation, OpenAI, Top 10 models compared, Agent frameworks, Hardware, Costs, Governance and India.

Google sells every layer of an automation stack, but its strongest models are gated

Google has a product at every layer of an automation stack: chips, models, a managed agent platform, open-source frameworks and protocols, and developer agents. The layers differ in maturity. The table shows which parts are production-grade.

Sources: Gemini API pricing, Google Cloud Next '26 wrap-up and Agent Platform pricing.

The platform layer also hosts other vendors' models. Anthropic lists Opus 5.5, Fable 5.1, Sonnet 5.5 and Haiku 5.5 on Google Cloud Vertex AI (Claude models overview). xAI's documentation lists Grok as a partner model in Vertex Model Garden (xAI Vertex guide, vendor claim). A business with a Google Cloud contract can therefore run more than one vendor's models on a single bill. For Claude, Google's data-residency page lists Singapore ML processing for Claude Opus 5.5, Sonnet 5.5, Sonnet 5, Opus 5, Fable 5.1 and Fable 5, but its locations page does not show those models in Singapore (a conflict, unresolved). No Claude model has India processing, and Anthropic's inference_geo setting does not apply on Google Cloud (Claude data residency). For Gemini, asia-south1 (Mumbai) has ML processing only for Gemini 3.5 Flash, Gemini 2.5 Flash (128K) and text-embedding-005. Gemini 3.6, 3.7 and 3.8 Flash are US and EU multi-region only, and Delhi (asia-south2) is not listed (Google Cloud data residency; Google Cloud locations, both read 12 October 2026).

Gemini 3.8 Flash carries production work while the Pro tier stays in preview

Google's pricing page was last updated on 9 October 2026. On it, the Flash models are GA and the Pro model is not. The table gives US dollars per 1M tokens on the standard paid tier (Gemini API pricing; release dates from the Gemini API changelog). "n/r" means the value was not recorded.

OpenRouter records the same prices for 3.8 Flash and 3.1 Pro Preview, and lists a 1,048,576-token context for both (OpenRouter models API). Google's 3.8 Flash model card gives 1M input tokens, 64K output tokens and a March 2026 knowledge cutoff (Gemini 3.8 Flash model card).

The list price is only part of the bill. Batch text costs 50% of the standard rate and cached input 10% of the standard input rate. Context-cache storage adds $0.50 per 1M tokens per hour on 3.8 and 3.6 Flash (rising to $1.00 on 1 January 2027), $1.00 on 3.5 and 3.1 Flash-Lite, and $4.50 on 3.1 Pro Preview and 2.5 Pro (re-read 11 October 2026). Thinking tokens are billed at the output rate: the output column of the Agent Platform pricing table reads "Output: text (response and reasoning)" (primary, read 12 October 2026). On 3.1 Pro Preview, once a query's input exceeds 200K tokens, all of its input and output tokens bill at the long-context rate, twice the input rate and 1.5 times the output rate (primary) (Agent Platform generative AI pricing). The 2.5 Pro row also prices prompts above 200K tokens higher. Google Search grounding costs $14 per 1,000 requests on Gemini 3 and newer after 5,000 free requests a month, and $35 per 1,000 grounded prompts on Gemini 2.5 after 1,500 free a day (Gemini API pricing). The grounding terms were read through a summarising tool.

The scheduled rise is the most important number on the page for any contract that runs past 2026. Gemini 3.8 Flash costs $0.75 / $3.75 through 31 December 2026 and $1.50 / $7.50 from 1 January 2027. Both the API pricing page and the Agent Platform pricing page confirm this (Agent Platform generative AI pricing). A re-read of the API pricing page on 11 October 2026, through a summarising tool, shows that cached input also doubles, from $0.075 to $0.15, and batch from $0.375 / $1.875 to $0.75 / $3.75. The same dated rise appears on 3.6 Flash, the 3.8 Flash TTS and 3.8 Flash-Lite TTS models and Robotics-ER 2. The 3.5 Flash-Lite, 3.1 Flash-Lite and 3.1 Pro Preview rows show no 2027 change. Google's 3.8 Flash model card lists $1.50 / $7.50 as the "regular" price, so the 2026 rate is a launch discount rather than a list price that is going up (Gemini 3.8 Flash model card). Argon's $2 / $10 is also introductory; Google's post says $4 / $20 applies afterwards.

Google releases and retires models fast. Four Flash generations shipped in 106 days: 3.5 Flash on 19 May 2026, 3.6 Flash and 3.5 Flash-Lite on 21 July, 3.7 Flash on 13 August and 3.8 Flash on 2 September. On 8 October, Google deprecated 3.7 Flash and routed its ID to 3.8 Flash, and routed 3.5 Flash to 3.6 Flash. Code that still calls the old IDs now gets a different model without any code change.

The same changelog records other breaking changes (Gemini API changelog). gemini-3-pro-preview shut down on 9 March and now points to 3.1 Pro Preview, and the gemini-pro-latest and gemini-flash-latest aliases changed on 21 January. Between January and October there were at least eight shutdowns or deprecations, including Veo 2.0 and 3.0 (30 June) and Imagen 4 (17 August). On 18 September, access to the 2.5 models was limited to recent users, with the advice: "For any new projects, use our latest models: 3.5 Flash-Lite or 3.8 Flash".

Gemini 4 Argon and Deep Think sit behind access gates

Google announced Gemini 4 Argon on 30 September 2026. It is "rolling out to a set of trusted cyber defenders through our Fairwind Program". Release to "developers, enterprises, and consumers" comes later, "starting with paid API customers and Google AI Ultra subscribers". Google says it is "actively engaged in the U.S. government's voluntary process for pre-release model access" (Google blog: Gemini 4 Argon). That process is the voluntary framework set under Executive Order 14409 (signed 2 June 2026), under which developers give the government up to 30 days of pre-release access. Sec. 3(c) rules out mandatory licensing or preclearance, and no implementation output was found (Federal Register: EO 14409). No GA date has been given.

Google's claims for Argon are all vendor-reported. The introductory API price is $2 input and $10 output per 1M tokens, with cached input 95% off. A footnote adds that "after the introductory period expires", $4 input and $20 output apply; no end date is given (re-read 11 October 2026). The output limit is 1M tokens, up from 64K. The post's summary calls this a "1 million token limit", which an earlier reading in this research took as a 1M-token context; the body of the post says it is the output limit. Google also claims leading scores on DeepSWE v1.1 (77.9%), the Vals Index, Zapier's AutomationBench (51.3%, #1), LVBench (91.7%), CWE-bench v1 (68%, tied first) and Gray Swan's indirect prompt-injection benchmark.

Earlier research doubted Argon existed because it was missing from OpenRouter and from Google's September blog round-up (Google AI blog round-up); the 30 September post settles that.

Gemini 3 Deep Think received an upgrade announced in February 2026. Google reports 48.4% on Humanity's Last Exam (HLE) without tools, 84.6% on ARC-AGI-2 (verified by the ARC Prize Foundation, per Google) and a Codeforces Elo of 3455. Google AI Ultra subscribers can use it in the Gemini app. API access is open only to "select researchers, engineers and enterprises" through an early-access sign-up (Google blog: Gemini 3 Deep Think). No API price was found.

Gemini 3.5 Pro appears neither on the pricing page nor in OpenRouter's catalogue. A July search snippet that claimed a June release is not supported (aggregator snippet). Wikipedia says the launch was "delayed due to ongoing testing" (Wikipedia: 2026 in AI, secondary).

Specialised models cover voice, images, video and robotics

Voice is the most automation-relevant of the specialised lines. gemini-3.8-live and gemini-3.8-live-extended-thinking became GA on 15 September 2026. 3.8 Flash TTS and 3.8 Flash-Lite TTS followed on 22 September, with 150+ voices and "regional dialects" (Gemini API changelog). The changelog gives no TTS language list. Elsewhere, the Gemini app supports 13 of the 22 scheduled languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu and Urdu (Gemini Apps Help: languages). This is an app list, not a model-capability list. Google Cloud's Chirp 3 speech-to-text is GA for hi-IN and en-IN and in preview for 11 other locales (as-IN, bn-IN, bn-BD, gu-IN, kn-IN, ml-IN, mr-IN, or-IN, pa-Guru-IN, ta-IN, te-IN) (Chirp 3 docs), though the data-residency page lists no India processing for it; Speech-to-Text V2 standard costs $0.016 a minute for the first 500K minutes (Speech-to-Text pricing). Cloud Translation covers 19 of the 22 languages, not Bodo, Kashmiri or Santali (Cloud Translation languages). The only independent 2026 Indic test found, the Indi-RomCoM preprint (a single study), has Gemini 3.5 Flash falling from 61.4 to 54.1 at 75% code-mix (arXiv 2606.30790). Hindi and other Indian languages still need testing on your own content before you rely on them.

Image generation is priced per image-output token. Gemini 3.1 Flash Image ("Nano Banana 2", GA 28 May) costs $0.50 input and $60 per 1M image-output tokens, Nano Banana 2.1 (GA 6 October) $1.50 / $30 and Nano Banana Pro $2.00 / $120 (Gemini API pricing). On 6 October Google also deprecated gemini-3.1-flash-image, with no shutdown date yet, and recommends moving to Nano Banana 2.1 (changelog, re-read 11 October 2026). Arena ranks Nano Banana 2.1 #6 in text-to-image (Arena leaderboard). For video, Gemini Omni Flash generates and edits clips for $1.50 input and $9.00 text output or $17.50 video output, and Veo 3.1 Lite entered preview on 31 March. gemini-robotics-er-2-preview (30 July) costs $1.00 / $5.00, rising to $2.00 / $10.00 on 1 January 2027.

The Deep Research agents have been in preview since 21 April. The older deep-research-pro-preview-12-2025 shuts down on 23 October 2026, eleven days after this page's status date.

Gemma 4 is the open-weight fallback

Gemma 4 was released on 2 April 2026 in two sizes: gemma-4-26b-a4b-it, a mixture-of-experts model with 4B active parameters according to its name, and the dense gemma-4-31b-it (Gemini API changelog). Google's API offers Gemma 4 on the free tier only. Third-party hosting is cheap: OpenRouter lists the 26B-A4B at $0.0675 input and $0.225 output and the 31B at $0.09 / $0.34, both with a 262,144-token context (OpenRouter models API).

Hugging Face data shows 11.7M downloads of the 26B-A4B and 9.6M of the 31B, under an Apache-2.0 licence (Hugging Face API, google). The licence comes from a single source, so check it before relying on it. For self-hosting, the research set estimates that a 32B-class dense model needs about 19 GB at 4-bit quantisation, which fits a 32 GB card (Hardware).

Per-task costs put Gemini 3.8 Flash in the middle of the price range

At $0.75 / $3.75, Gemini 3.8 Flash costs 7.5 times the per-token price of the cheapest tiers from Anthropic and OpenAI. Claude Haiku 5.5 (for prompts up to 100K tokens; $0.50 / $2.50 above) and GPT-6 Luna both cost $0.10 / $0.50 (Claude pricing; OpenAI pricing). Gemini 3.8 Flash is also far above DeepSeek V4.1 Flash at $0.15 / $0.60 off-peak (DeepSeek pricing).

The research set's cost model priced six common automation tasks at 2026 list rates. Its token profiles are assumptions, not measurements (method on Costs). The model applies Haiku 5.5's base rate throughout, but the code-change agent's last 8 turns pass 100K prompt tokens and bill at Haiku's higher rate. The Haiku cell for that task therefore shows about $0.28 instead of the model's $0.136 (calc, as on Claude).

On five of the six tasks, Gemini 3.8 Flash costs about two-fifths as much as Sonnet 5.5 and one-fifth as much as Opus 5.5. It also costs about seven to eight times as much as the cheapest tiers, except on lead research, where search fees narrow the gap, and against Haiku on the code-change agent (about 3.6 times), where Haiku's price step above 100K tokens applies. Capability does not justify that premium by itself. Artificial Analysis's own leaderboard (Intelligence Index v4.3.2, live 12 October 2026; primary) gives Gemini 3.8 Flash 40.9, against 43.4 for Claude Haiku 5.5 (Artificial Analysis leaderboard).

The case for 3.8 Flash therefore rests on what is distinctive about it: 1M-token multimodal input including video (LVBench 87.8%, vendor-reported), Google Search grounding, a free tier for prototyping, and integration with Google Workspace and Google Cloud.

The 2027 rise makes these numbers worse. Applying $1.50 / $7.50 to the same profiles (this page's arithmetic), a support reply rises from $0.0024 to about $0.0048. The code-change agent rises from $1.02 to about $2.04. Both figures assume the cached rate doubles to $0.15, as the pricing page now shows (re-read 11 October 2026).

In rupees, at the research set's rate of ₹96 per US dollar (X-Rates October 2026 average, 95.95), 10,000 support replies cost about $24, or ₹2,300, in 2026. In 2027 the same volume costs about $48, or roughly ₹4,600. Batch processing halves either figure for work that can wait.

Next '26 turned Vertex AI into the Gemini Enterprise Agent Platform

Google Cloud Next '26 ran in Las Vegas from 22 to 24 April 2026. Google's wrap-up, dated 25 April, counts 260 announcements and more than 32,000 attendees. The centrepiece is the Gemini Enterprise Agent Platform, described as the evolution of Vertex AI "to build, scale, govern and optimise agents" (Google Cloud Next '26 wrap-up). Agent Studio builds agents from prompts and exports their logic to ADK for full-code development. Agent Registry indexes approved internal agents, tools and skills. Agent Identity gives each agent a cryptographic ID, with an audit trail mapped to authorisation policies. Agent Gateway is the single control point for agent-to-tool connections and policy, and includes Model Armor protection against prompt injection and data leakage. Gateway integrations from Broadcom, Cisco, CrowdStrike, Palo Alto Networks, Zscaler and others are due in preview "soon". Around these sit ADK graph-based sub-agent networks, an Agent Runtime with sub-second cold starts, Agent Sandbox, Memory Bank, Sessions, Agent Simulation and Evaluation, and Agent Optimizer.

These pieces line up with the controls security frameworks ask for: one identity per agent, least-privilege tool access and audit logs (Governance).

Runtime pricing is published; seat pricing is not. Agent compute costs $0.085 per vCPU-hour and agent storage $0.30 per GiB-month. Memory Bank and Sessions have been billed since 1 September 2026, and Agent Gateway since 13 July 2026 (Agent Platform pricing). A re-read on 11 October 2026 adds the rest of the runtime schedule: the first 50 vCPU-hours a month per account are free, Agent Memory costs $0.009 per GiB-hour, and one vCPU-hour ($0.085) covers 15,000 Agent Gateway calls, or 1 million Memory Bank or Sessions writes, or 3 million reads. Model tokens are billed separately. Gemini Enterprise, the product for employees, is "contact Google Cloud sales". Its Plus edition can be tried free for 30 days "without sales assistance", and the same page advertises $300 in free credits and 20+ free products (Gemini for Business).

Google now supports MCP widely. The wrap-up lists bring-your-own MCP servers for Gemini Enterprise, managed remote MCP for databases, a GA Cloud Run remote MCP server, and MCP servers for Cloud Storage, Looker, Workspace (in preview) and Gemini Cloud Assist. Google is a platinum member of the Linux Foundation's Agentic AI Foundation, which governs MCP (Linux Foundation release, 9 Dec 2025, primary). MCP's stateless 28 July 2026 specification makes hosted servers easier to scale out (MCP blog, 2026-07-28).

Google's customer results show what others report having deployed, but they are all vendor-reported and unaudited:

Sources: Google Cloud Next '26 wrap-up, Google Cloud: Toolstation and Google Cloud: FairPrice. Tata Steel is the only Indian name on the list. More cases are on Use cases.

ADK is the open, code-first entry point

The Agent Development Kit (ADK) is the part of Google's stack any team can adopt with no contract at all. google/adk-python is Apache-2.0 and has 21,777 GitHub stars (GitHub API: adk-python). Its latest release is v2.11.0, published on 2 October 2026 (GitHub API: ADK latest release; PyPI google-adk). The npm package @google/adk is versioned separately, at 2.2.1 on 6 October (npm @google/adk). The README lists official SDKs in Python, Java, Go, TypeScript and Kotlin, according to an earlier read (not re-checked). Releases come every few weeks, so pin ADK versions as carefully as model IDs.

Sources: GitHub API: openai-agents-python, GitHub API: claude-agent-sdk-python, GitHub API: agent-framework, GitHub API: langgraph and GitHub API: crewAI. Stars measure developer interest, not production use. The Claude Agent SDK's README says its use "is governed by Anthropic's Commercial Terms of Service", including in products offered to your own customers, so check those terms as well as the LICENSE file (checked 11 October 2026). The full comparison is on Agent frameworks.

A2A and AP2 moved to neutral governance

A2A (Agent2Agent) was "originally developed by Google and donated to the Linux Foundation". A Technical Steering Committee with members from AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow runs it, and a site banner says it is joining the Agentic AI Foundation (A2A project site). The project blog (27 August 2026) says A2A was accepted as an AAIF "Growth Stage project"; this is single-source, as no AAIF page confirms it. The blog names ServiceNow, Salesforce, Atlassian and SAP as platforms using A2A, with native support on Google Cloud, Bedrock AgentCore and Azure AI Foundry, but names no customer deployment (project claim) (A2A blog, 27 Aug 2026). On 9 April 2026 the Linux Foundation counted more than 150 supporting organisations, up from more than 50 a year earlier, and called version 1.0 the "first stable specification". It named integrations with Azure AI Foundry, Copilot Studio, Amazon Bedrock AgentCore Runtime, LangGraph and CrewAI. Despite a headline about enterprise production use, it named no production users and described that use only by sector (Linux Foundation A2A release). The specification defines both JSON-RPC and gRPC bindings (A2A specification), and the repository has 26.1k stars (GitHub: A2A).

AP2 (Agent Payments Protocol) launched on 16 September 2025 as "an extension of the Agent2Agent (A2A) protocol and Model Context Protocol (MCP)" (Google Cloud blog: AP2). The project site lists a v0.2 release, with the specification and SDK under Apache-2.0 (AP2 docs). On 28 April 2026 Google announced it was "donating the Agent Payments Protocol (AP2) to the FIDO Alliance", along with Verifiable Intent, which it co-developed with Mastercard (Google blog: AP2 to FIDO). The FIDO Alliance's page could not be read, but Google's post settles the date. Secondary comparisons say AP2 authorises payments with signed mandates modelled as verifiable credentials (Blindpay comparison). They also say it complements OpenAI and Stripe's ACP: ACP handles checkout and AP2 proves authorisation (Crossmint comparison). ACP itself is a beta, Apache-2.0 specification (latest version 2026-04-17). Its README names OpenAI and Stripe as maintainers, and its MAINTAINERS.md also lists Meta (GitHub: ACP). Its repository has 3.2k stars (GitHub: AP2).

Gemini CLI leads on stars, not on installs

Google's three developer agents have generous free tiers but trail rivals on usage signals.

Sources: GitHub: gemini-cli, Gemini CLI releases, npm @google/gemini-cli, npm downloads: gemini-cli, Gemini CLI quotas, Jules, Antigravity pricing, GitHub API: Agno release and the Gemini API changelog.

The gap shows most clearly in downloads. In the same 30 days, npm recorded 98.0M downloads of OpenAI's Codex CLI and 56.3M of Claude Code, against 1.74M for Gemini CLI (npm downloads: Codex; npm downloads: Claude Code). These counts can include automated CI installs, so they indicate usage but are not user counts. Gemini also reaches developers through other tools: GitHub Copilot's model list includes Gemini 3.7 and 3.8 Flash (GitHub Copilot supported models).

Inside Google, Sundar Pichai says about 75% of new code is AI-generated and approved by engineers (vendor-reported). Semafor reports the share was 25% in 2024 and 50% in autumn 2025 (Semafor, 24 Apr 2026; ETV Bharat).

Computer use is still a preview feature

Gemini's computer use launched in gemini-3-pro-preview and gemini-3-flash-preview on 29 January 2026, and entered public preview in Gemini 3.5 Flash on 24 June 2026 (Gemini API changelog; Google blog, HN listing). On 8 October, 3.5 Flash was deprecated and routed to 3.6 Flash. No GA date or current list of supported models was found, and Google's computer-use documentation could not be read.

Google's 3.8 Flash model card reports 59.0% on OSWorld-2.0, as a partial score with a batch tool enabled (Gemini 3.8 Flash model card). For comparison, Anthropic made its computer-use tool GA on the Claude API on 19 August 2026 (Claude release notes).

Partial scores overstate how often agents finish tasks. The June 2026 OSWorld 2.0 paper counted only fully completed tasks within a 500-step limit. On that measure the paper's best model, Claude Opus 4.8 at max thinking, finished 20.6% of tasks (54.8% partial) (arXiv 2606.29537, primary abstract). Later leaderboard rows (Epoch's copy of the OSWorld 2.0 board, secondary) reach 31.4% binary and 68.3% partial for Claude Opus 5 (Max), and 27.3% and 62.7% for GPT-5.6 Sol; there is no Gemini row (Epoch benchmark data). BenchmarkList reports Fable 5.1 at 77.9% partial and 41.7% strict as Anthropic figures (not read directly) (BenchmarkList). Anthropic's own tables use OSWorld 2.1, a different version. Gemini's 59.0% is a partial-credit figure, not a completion rate.

Security is the binding constraint. Brave documented indirect prompt injection in Perplexity's Comet browser on 20 August 2025 (date from earlier reads; the post as re-read on 12 October 2026 shows no date), then showed instructions hidden in screenshots working against several AI browsers in October 2025 (Brave: Comet prompt injection; Brave: unseeable prompt injections). Google's defences, Model Armor and Argon's self-reported lead on an indirect-injection benchmark, have not been independently evaluated. Chrome auto-browse and Project Mariner were not verified. Where a system has an API, Google's MCP servers are the safer route.

Google's benchmark claims are self-reported; independent boards put Argon in the leading group

Almost all of Google's benchmark numbers come from its own pages, against rival models Google chose. The September card for 3.8 Flash compares it with Claude Opus 5, Claude Sonnet 5, GPT-5.6 Sol and GPT-5.6 Terra. It does not compare it with the Claude 5.5 models released from 22 September, or with GPT-6 Astra.

Setting rivals' own published figures alongside Google's produces three contrasts, none of them like-for-like. On AutomationBench, Google reports 51.3% for Argon, while Anthropic's table gives 40.0% for Opus 5.5 and 41.4% for GPT-6 Astra (Anthropic: Claude Opus 5.5). Anthropic's footnote says Zapier built and ran the benchmark and counted safeguard interventions as failures, which Anthropic says understates Opus 5.5. If the runs are comparable, Argon leads on the benchmark closest to business automation; because the figures come from separate vendor reports, treat this as a lead to check.

On Terminal-Bench 4.0, the third-party tbench.ai board (330 trials per run) confirms 3.8 Flash at 19.09% (mini-SWE-agent harness) (Terminal-Bench board). Google's card figures for Claude Opus 5 (51.8%) and GPT-5.6 Sol (37.3%) match the board's 51.82% and 37.27%; GPT-5.6 Terra's 23.6% is from the card (re-read 11 October 2026). The board is led by Opus 5.5 at 64.85%, Sonnet 5.5 at 61.82% and GPT-6 Astra at 58.18% (max effort, vendor CLIs), and Argon is not on it. Anthropic's own figures are vendor-reported and sit above the board: 66.4% for Opus 5.5 (xhigh) and 70.6% for Sonnet 5.5 (effort unstated; not reproduced). Haiku 5.5's 39.2% has no board row (Anthropic: Claude Sonnet 5.5; Anthropic: Claude Haiku 5.5). Harnesses and effort settings differ, but the gap is large enough that Flash should not be the default for long terminal-based coding agents without a test on your own workload. On GDPval-AA, Google reports v2 and Anthropic v2.1, so the Elo scores (1545 for 3.8 Flash, 1846 for Opus 5.5) cannot be compared directly.

The independent measures point the same way. Google's best model is in the leading group. Gemini 4 Argon was a preliminary #1 on LMArena's 8 Oct 2026 text snapshot (1525 ±9) and tops Epoch's copy of APEX-Agents (87.9%, effort unknown). It is #9 in WebDev and fifth on Artificial Analysis's index (52.6), and it is absent from the Terminal-Bench 4.0 board. An earlier Flash model came within a point of the top of SWE-bench's bash-only board in February. The cross-vendor view is on Top 10 models compared.

Operational limits matter more than benchmark gaps

The documented limitations are mostly operational, not about capability. Routing old IDs to new models means behaviour can change without any code change. The Pro tier and computer use are both still in preview, and Argon goes to vetted cyber defenders first. The 1 January 2027 rise and Argon's introductory pricing, which doubles to $4 / $20 at an unstated date, make 2027 costs hard to predict. Gemini's personal memory is "not available while signed in to a work, school, or supervised account", so Workspace customers cannot rely on it for company use (Gemini Apps Help). Local AI command-line tools are also an attack surface: malicious versions of the Nx build package, published to npm on 26 August 2025, used locally installed AI CLIs, including Gemini's, to search for secrets (Wiz).

Three further items come only from Wikipedia's 2026 page and are leads that need primary confirmation (Wikipedia: 2026 in AI). One is a May 2026 BBC report that Gemini hacked three companies during an independent cybersecurity test (BBC). The others are the first wrongful-death suit against Google over the Gemini chatbot, filed on 4 March 2026, and Character.AI and Google agreeing in January 2026 to settle harm lawsuits.

TPUs give Google its own compute, and Anthropic is the anchor external buyer

Google documents two current TPU generations and announced an eighth on 22 April 2026.

Sources: Cloud TPU v6e docs, Cloud TPU7x docs, The Register on Ironwood, Wikipedia: Tensor Processing Unit and the Google Cloud Next '26 wrap-up. The Register notes that Google's 42.5 ExaFLOPs figure for an Ironwood pod compared FP8 peak performance with an FP64 figure, so this page does not repeat it. TorchTPU, which runs PyTorch natively on TPUs, is in preview with selected customers.

Google now sells TPU capacity outside the company as well as using it for Gemini.

Compute also flows the other way. On 5 June 2026 Google agreed to pay $920M a month for about 110,000 GPUs at the former xAI, now part of SpaceX, from October 2026 to June 2029. This rests on Wikipedia citing CNBC, plus a Hacker News headline (Wikipedia: SpaceX IPO; CNBC, HN listing). This page infers that Google's demand for capacity is outrunning its own chip supply.

No TPU rental rate was captured. Google Cloud's published compute discounts are 28% for one-year and 46% for three-year committed-use contracts, and up to 30% for sustained use (Google Cloud VM pricing). For most businesses in India, TPUs matter mainly indirectly, through the prices and capacity of the Gemini and Claude models that run on them. India's state compute pool does include TPUs: IndiaAI's empanelled list counts 5,548 TPU v6e (Trillium) chips among 45,081 empanelled GPUs and accelerators (undated portal PDF, primary) (IndiaAI empanelled GPUs). In this page's view, renting TPUs directly is worth considering only for fine-tuning or high-volume serving of open models (Hardware).

Usage and money: 950 million app users and a cloud business growing 82%

Alphabet's quarterly releases give the clearest view of Gemini adoption. All figures are company-reported and unaudited.

Sources: Alphabet Q4 2025 release, Alphabet Q1 2026 release, Alphabet Q2 2026 release and the Alphabet 10-Q. The 350M paid subscriptions span YouTube, Google One and other products, not only Gemini.

Google publishes no monthly token total. At 22B tokens a minute, a 30-day month comes to roughly 950 trillion API tokens (this page's arithmetic). The API token rate grew about 2.2 times from Q4 2025 to Q2 2026, while app users grew about 27%. On these figures, developers and businesses are taking up Gemini faster than consumers.

In Q2 2026, Alphabet's total revenue was $119.8B (+24%) and consolidated operating income $40.8B (+30%). Google Cloud's operating margin rose from 20.7% to 35.6%. Net income of $112.1B (diluted EPS $9.11) includes a $99.0B gain on equity securities, which added $77.1B after tax and $6.26 per share (Alphabet Q2 2026 release). An aggregator attributed the gain to stakes in Anthropic and SpaceX, but the release text does not name them (Sentisense, secondary).

Capital spending was $44.9B in the quarter and $80.6B in the first half, against $39.6B in the first half of 2025 (Alphabet 10-Q) and $91.45B for all of 2025 (SEC XBRL capex data). To help fund it, Alphabet raised $49.6B net from stock and mandatory convertible preferred shares in June 2026, plus $20.3B of senior notes during Q2. It expects to recognise "just over 50%" of its $519.5B in remaining performance obligations within 24 months.

Two widely quoted figures are secondary only. Aggregator summaries say 2026 capex guidance was raised to $195–205B from $180–190B (aiweekly Q2 summary), while coverage of the Q1 results gave the earlier range as $175–185B (AI Weekly capex roundup). Neither range appears in the SEC documents read. The often-quoted Q2 free cash flow of −$5.9B is also aggregator-only. The filing contradicts an earlier summary that claimed Cloud's margin had turned negative.

Market-share proxies place Google among the top three vendors, though none measures enterprise spend directly. On OpenRouter, Google models drew 21.5% of text requests in the week beginning 28 September 2026 (up 23% on the week), behind DeepSeek at 21.8% and ahead of OpenAI at 18.2% (OpenRouter rankings). Menlo Ventures data put Google at 20% of enterprise LLM usage in mid-2025 (BigDATAwire on Menlo, secondary). A summary of Menlo's late-2025 survey gives Google 21% of enterprise API spend, against 40% for Anthropic (Longterm Wiki, secondary and unverified). In Pew's 2026 survey of US adults, 24% named Gemini as a chatbot they use, against 44% for ChatGPT (TheNextWeb on Pew, secondary summary).

In August 2026 Business Insider reported that Demis Hassabis is becoming Alphabet's chief scientist and DeepMind's chair, with Koray Kavukcuoglu heading DeepMind under Sundar Pichai (Business Insider; see People and Companies).

India: a Visakhapatnam hub, summit funds and unverified consumer offers

Google's main India commitment is an AI and data-centre hub in Visakhapatnam, Andhra Pradesh. October 2025 reports describe about $15B over five years (2026–2030) for a one-gigawatt-scale hub, announced at "Bharat AI Shakti", a Google event in New Delhi. Google Cloud CEO Thomas Kurian is quoted calling it the largest AI hub Google will invest in outside the US, and saying it will scale to multiple gigawatts (Times of Oman; Investment Monitor; meatechwatch, 17 Oct 2025). The reports name AdaniConneX and Airtel as partners and describe an international subsea-cable gateway and a state-government projection of up to 1.8 lakh jobs (Verdict). Wikipedia separately records a July 2025 report of a $6B, 1 GW campus with about $2B for renewable energy (Wikipedia: Visakhapatnam, secondary). The two figures differ in date and scope. Google's own announcement was not found in this research, so both stand as reported.

Construction has begun. The Economic Times quotes a minister speaking "after the Ground Breaking Ceremony of Google Cloud India AI Hub at Visakhapatnam" (Economic Times). The PTI report is dated Tuesday 28 April 2026 and describes the minister speaking "on Tuesday" (re-read 11 October 2026), which matches Wikipedia's date for the ceremony (Wikipedia: South Coast Railway zone).

Opposition is building: the state has approved more than 6.5 GW of AI data-centre projects in the city, and a "People's Convention on AI Data Centers" on 6 September 2026 demanded a moratorium (Wikipedia: Opposition to AI data centers, citing The Hindu and Hindustan Times). Power and construction timelines are therefore uncertain. Either way, this is a 2026–2030 build with no effect on Gemini API prices or latency in India in 2026.

At the AI Impact Summit in New Delhi in February 2026, Google announced a $30M "AI for Government" challenge and a $30M "AI for Science" fund. Both figures are single-source (Fortune, 23 Feb 2026).

Consumer offers were widely reported but have not been confirmed as current. Eligible Jio users were offered 18 months of free Google AI Pro, worth about ₹35,100, though sources differ on eligibility: one says ages 18–25, another says a wider rollout was coming. Both are late-2025 snippets (Analytics Vidhya; Outlook Business). A free year of Google AI Pro for Indian undergraduates was valid to 15 September 2025, and no 2026 offer was found (Outlook Business, snippet). Google's India blog carried no Gemini India items when checked on 11 October 2026 (Google India blog).

Two of the three gaps for Indian deployments are now partly answered. Google Cloud lists India (asia-south1, Mumbai) ML processing only for Gemini 3.5 Flash, Gemini 2.5 Flash (128K) and text-embedding-005, not for 3.6–3.8 Flash, and Delhi is not listed (Google Cloud data residency). The Gemini app lists 13 of the 22 scheduled languages, and Chirp 3 is GA for hi-IN and en-IN, with 11 other locales in preview (Gemini Apps Help: languages; Chirp 3 docs). Model-level Indic quality is still unmeasured apart from one preprint (arXiv 2606.30790), and the data-location terms for Gemini Developer API calls from India were not read. Country context is on India and Asia-Pacific.

What this means for your business

Google fits some automation jobs well and others less well. Its strengths are multimodal, long-context Flash models with a free tier, search grounding, open agent tooling (ADK, A2A, AP2 and MCP servers) and a governance-heavy platform for organisations already on Google Cloud or Workspace. Its weaknesses are a Pro tier still in preview, a gated top model, a scheduled Flash price rise, frequent model routing changes, thin India processing and partial Indian-language coverage (13 of 22 scheduled languages in the app). Compare each option below with other vendors' on Top 10 models compared.

What to do first

Pin exact model IDs. Keep -latest aliases and -preview IDs out of production. Build a fixed evaluation set from your own work, and re-run it whenever the changelog routes an old ID to a new model, as happened on 8 October.

Don't build on gated models. Keep production off Argon, Deep Think and 3.1 Pro until they reach GA on the paid API. Route the hardest reasoning steps to whichever vendor's GA flagship passes your tests, and keep a fallback vendor: several vendors restricted access to frontier models in 2026 (Timeline; Top 10 models compared).

What to budget, and by when

• Cost anything that runs past 31 December 2026 at the 2027 rate ($1.50 / $7.50, cached input $0.15): 10,000 support replies rise from about ₹2,300 to ₹4,600. Batch halves either figure for overnight jobs, and context caching cuts the cost of repeated prompt prefixes.
• Budget in rupees with a stated exchange rate, since every Google price above is in US dollars (Costs). Gemini Enterprise seats are quote-only; agent compute is $0.085 per vCPU-hour after 50 free hours a month, plus tokens. Google's Generative AI Leader certification for staff costs $99 plus tax (Google Cloud certification).
• Dates: deep-research-pro-preview-12-2025 shuts down on 23 October 2026, the Flash rise applies from 1 January 2027, and India's DPDP cross-border provisions apply from 13 May 2027.

Data handling and India processing

Before Gemini handles regulated Indian personal data, get written data-residency and region commitments for the specific model. Google Cloud lists India (asia-south1) ML processing only for Gemini 3.5 Flash, Gemini 2.5 Flash (128K) and text-embedding-005, so a 3.8 Flash or Flash-Lite workflow processes Indian data abroad. Cross-border transfer conditions under DPDP Rule 15 and the Act's s.16 apply from 13 May 2027, and no restricting notification or order has been found (G.S.R. 846(E)). The data-location terms for Gemini Developer API calls from India were not read, so the Google Cloud facts may not cover them.

Risks to plan for

Treat computer use as a last resort. Use it only for systems with no API. Run it in isolated browsers without logged-in sessions, and require human approval for any write, payment or external message (Governance).

Expect firmer pricing, not exit risk. Google's 35.6% cloud margin and $513.9B backlog make an exit unlikely, so the main risks are price and behaviour changes. This page infers from capex of about $45B a quarter and the 2027 increases that pricing will firm up; Google has not said so.

What to test

Test Indian-language quality on your own content (only one independent preprint measures it), long terminal-based coding agents (3.8 Flash scores 19.09% on the Terminal-Bench 4.0 board), vendor-reported multimodal claims such as LVBench 87.8%, and computer use on your real screens, counting only fully completed tasks.

Questions to ask any vendor or implementation partner

• Which exact model IDs will run, and how will routing changes be caught and re-tested?
• Does the budget include the 2027 rise, batch, caching and an exchange-rate assumption?
• Where is each model processed, and is that in writing?
• Which case studies are cited, and are they vendor-reported, as Google's Tata Steel (300+ agents in nine months) and GE Appliances (800+ agents) figures are?
• What is the fallback if a model is gated, deprecated or re-routed?
• What data terms cover code sent through CLI tools?

The Playbook and Tool stack cover how to run the project.

Open questions and unverified items

Sources

1. Gemini API pricing
2. Google Cloud Next '26 wrap-up
3. Agent Platform pricing
4. Claude models overview
5. xAI docs: Grok on Google Cloud Vertex AI
6. Gemini API changelog
7. OpenRouter models API
8. Gemini 3.8 Flash model card
9. Agent Platform generative AI pricing
10. Google blog: Gemini 4 Argon
11. Google blog: AI news round-up
12. Google blog: Gemini 3 Deep Think
13. Jetstream: aggregator summary on Gemini 3.5 Pro timing (snippet)
14. Arena leaderboard
15. Hugging Face API: Google models by downloads
16. Claude pricing
17. OpenAI pricing
18. DeepSeek pricing
19. Google Cloud: Gemini for Business (Gemini Enterprise)
20. Linux Foundation: formation of the Agentic AI Foundation (9 Dec 2025)
21. MCP blog, 2026-07-28
22. Google Cloud customer story: Toolstation
23. Google Cloud customer story: FairPrice Group
24. GitHub API: adk-python
25. GitHub API: ADK latest release
26. PyPI google-adk
27. npm @google/adk
28. GitHub API: openai-agents-python
29. GitHub API: claude-agent-sdk-python
30. GitHub API: agent-framework
31. GitHub API: langgraph
32. GitHub API: crewAI
33. A2A project site
34. Linux Foundation: A2A protocol surpasses 150 organizations (9 Apr 2026)
35. A2A specification
36. GitHub: A2A
37. Google Cloud blog: AP2
38. AP2 docs
39. Google blog: AP2 to FIDO
40. Blindpay: agent payment protocols compared
41. Crossmint: agentic payments protocols compared
42. GitHub: AP2
43. GitHub: gemini-cli
44. Gemini CLI releases
45. npm @google/gemini-cli
46. npm downloads: gemini-cli
47. Gemini CLI quotas
48. Jules
49. Antigravity pricing
50. GitHub API: Agno release
51. npm downloads: Codex
52. npm downloads: Claude Code
53. GitHub Copilot supported models
54. Semafor: Google CEO says 75% of new code is AI-generated (24 Apr 2026)
55. ETV Bharat: Pichai on AI-generated code at Google
56. Google blog: computer use in Gemini 3.5 Flash (HN listing)
57. Claude release notes
58. arXiv 2606.29537: OSWorld 2.0 (abstract)
59. BenchmarkList: OSWorld 2.0, Anthropic-reported results
60. Brave: Comet prompt injection
61. Brave: unseeable prompt injections
62. DeepMind: Gemini 3.1 Pro
63. Anthropic: Claude Opus 5.5
64. Anthropic: Claude Sonnet 5.5
65. Anthropic: Claude Haiku 5.5
66. Artificial Analysis leaderboard (Intelligence Index v4.3.2, read 12 Oct 2026)
67. SWE-bench: official leaderboards.json
68. METR time horizons
69. Terminal-Bench 4.0 board (tbench.ai)
70. Gemini Apps Help: personalization and memory
71. Wiz: the emerging use of malware invoking AI
72. Wikipedia: 2026 in AI
73. BBC: report on Gemini in a cybersecurity test (cited by Wikipedia)
74. Cloud TPU v6e docs
75. Cloud TPU7x docs
76. The Register: Google 7th-gen Ironwood TPUs debut (10 Apr 2025)
77. Wikipedia: Tensor Processing Unit
78. Anthropic: Expanding use of Google Cloud TPUs
79. Anthropic: Google and Broadcom deal
80. Alphabet 10-Q
81. Wikipedia: OpenAI
82. Yahoo Finance: Meta and Google discuss TPU deal
83. Wikipedia: Initial public offering of SpaceX
84. CNBC: Google to pay SpaceX $920M a month for xAI compute (HN listing)
85. Google Cloud: Compute Engine pricing
86. Alphabet Q4 2025 release
87. Alphabet Q1 2026 release
88. Alphabet Q2 2026 release
89. Sentisense: Alphabet Q2 2026 earnings summary
90. SEC XBRL: Alphabet capital expenditure data
91. aiweekly: Alphabet Q2 cloud and capex summary
92. AI Weekly: hyperscaler 2026 capex plans
93. OpenRouter rankings
94. BigDATAwire: Menlo Ventures enterprise LLM report (mid-2025)
95. Longterm Wiki: Menlo enterprise share fact page
96. TheNextWeb: Pew research on Americans and AI chatbots (2026)
97. Business Insider: Google AI leadership changes (Aug 2026)
98. Times of Oman: Google to invest $15B in Visakhapatnam AI hub
99. Investment Monitor: Google to establish $15bn AI hub in Visakhapatnam
100. meatechwatch: Google $15B Visakhapatnam AI hub (17 Oct 2025)
101. Verdict: Google AI hub in Visakhapatnam
102. Wikipedia: Visakhapatnam
103. Economic Times: Google Cloud India AI Hub ground-breaking (Apr 2026)
104. Wikipedia: South Coast Railway zone
105. Wikipedia: Opposition to AI data centers
106. Fortune: India AI Impact Summit closes (23 Feb 2026)
107. Analytics Vidhya: Jio users get free Google AI Pro plan
108. Outlook Business: Reliance and Google free Gemini Pro plan for Jio users
109. Outlook Business: free one-year Google AI Pro for Indian college students
110. Google India blog
111. Google Cloud: Generative AI Leader certification
112. Enterprise DNA: Google Cloud Next 2026 recap
113. Claude Platform: data residency
114. Google Cloud: Agent Platform data residency
115. Google Cloud: Agent Platform locations
116. LMArena: text leaderboard (8 Oct 2026 snapshot)
117. LMArena: WebDev leaderboard (8 Oct 2026 snapshot)
118. Epoch AI: benchmark data (zip)
119. Gemini Apps Help: supported languages
120. Google Cloud: Chirp 3 speech-to-text model
121. Google Cloud: Speech-to-Text pricing
122. Google Cloud Translation: supported languages
123. arXiv 2606.30790: Indi-RomCoM (preprint)
124. A2A blog: joining the Agentic AI Foundation (27 Aug 2026)
125. GitHub: Agentic Commerce Protocol
126. Federal Register: EO 14409, Promoting Advanced AI Innovation and Security (5 Jun 2026)
127. IndiaAI compute portal: empanelled GPUs (PDF)
128. e-Gazette: G.S.R. 846(E), DPDP Rules 2025