Technology releaseNovember 22, 2026Редакция 105.kz
Gemini 4 Neon from Google DeepMind: What a Cloud Model Offers Industrial Enterprises in Kazakhstan
Google DeepMind has introduced Gemini 4 Neon — a general-purpose cloud model with a 1M token context and a price from 0.15 dollars per million cached tokens. We examine which engineering tasks it covers and why it cannot be placed inside the process control loop.
Key takeaway. Gemini 4 Neon is a general-purpose cloud model, not an industrial system. According to the Google DeepMind publication, it has a 1M token context, a price of 3 and 15 dollars per million input and output tokens, and strong results on public benchmarks. For an enterprise in Kazakhstan, the realistic scenario is analysing documentation and code outside the process control loop; an on-premise option is not stated in the source.
What happened
On 18 November 2026, Google DeepMind introduced Gemini 4 Neon, a general-purpose model. In the publication it is described as a frontier-level model with fast end-to-end execution of tasks in software development, knowledge work and scientific processes. Its place in the family is set between Carbon, which is assigned the longest and most complex projects, and Hydrogen, responsible for fast everyday operations: Neon is meant to combine depth of reasoning with a lower cost per task.
Reported results: 61 points and third place in the Artificial Analysis Intelligence Index at the xhigh reasoning level; second place on the Vals Index (70.9%) and a position on the Pareto frontier; second place for steerability in the Agent Arena Index (1336 points across more than 4200 real agent sessions). In development: first place in the Artificial Analysis Coding Agent Index (58.6), 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 in high mode. Context window — 1M tokens. Pricing — 3 dollars per million input tokens, 0.15 dollars per million cached (95% discount) and 15 dollars per million output. Access: Gemini API in Google AI Studio, Gemini for business (via API and model selection in the list), Google Antigravity for paid Pro and Ultra subscribers, the Gemini app. For information protection, the Fairwind programme was announced — access will start being granted to "trusted defenders" in the coming days.
What this means for enterprises in Kazakhstan
The main limitation is set by the source itself: the publication contains no mention of deployment inside an enterprise's perimeter. All the listed access channels are Google cloud services. That immediately defines the class of tasks: processing documents, spreadsheets and code, but not the control loop and not real time. Work with industrial protocols, SCADA and process control systems is not stated in the text.
Oil and gas. The realistic area is technical documentation and maintenance and repair (ТОиР) reporting: data sheets for pump and compressor equipment and valves, failure reports, operating hours, work permits, correspondence with contractors. The source claims synthesis of data from tables, PDFs, charts, audio and video in a single context with a structured output. The practical class of tasks is consolidating scattered unit-level reports into a single table and finding discrepancies — not diagnosing a failure from vibration.
Power generation. The picture is similar: transformer and turbine data sheets, measurement logs, ПСВ procedures, justifications for taking units out for repair. Reading dense professional charts is noted separately — 73.5% on Chartography. That is useful for preparing summaries but does not replace regime calculations (assessment).
Uranium and mining. Laboratory reports, sampling logs, grade tables, processing stage reports. The model can speed up the parsing of such arrays, but geological modelling, reserve estimation and geophysics work are not stated in the source.
Plants. Specifications, cost estimates, Excel tables, documentation on feeders and processing stages, plus code for internal integrations: the model is positioned as strong in multi-file debugging and targeted edits in repositories.
A separate issue is the data regime. For systems covered by personal data localisation requirements and the internal information security requirements of national companies, sending documents to an external cloud without a dedicated solution is not acceptable (general practice). Such tasks need a closed perimeter — this class of tasks is covered on the Corporate LLMs page. The source does not address the language question: quality in Russian and Kazakh was not measured in the publication.
How to apply it: an 8–12 week pilot
The steps below are managerial, not implementation-level engineering recipes.
- Inventory the corpus. Choose one class of documents (for example, failure reports for 12 months) and record where it sits, in what format, and who reads it. The pilot task must sit outside the process control loop.
- Classify by access regime. Split documents into three groups: public, anonymisable, and prohibited from being taken outside. For the third group, plan a closed perimeter from the outset.
- Measure the baseline. Before starting, measure how much time an engineer spends on the task and how often they err. Without that, the pilot's effect cannot be substantiated.
- Limited run and acceptance. Run the model on a sample, have the result checked by two independent specialists, and record the error rate and token consumption per document.
- Scaling decision. Compare the cost of the task with the baseline and choose: extend the cloud scenario, move the task class into a closed perimeter, or close the direction. Industry context is on the Industrial Enterprises page.
Limitations and risks
All the figures given are public and internal Google DeepMind benchmarks. There are no industrial validations in the source: not a single example on production data, not a single evaluation on real arrays of technical documentation. Results on legal and tax datasets (first place on Legal Research Bench and Tax Agent Bench) do not transfer directly to engineering tasks.
The claim of the "lowest level of fabrication" refers to internal testing in Google Antigravity and does not mean zero: the share of invented statements is reduced but not eliminated. Resistance to indirect prompt injections was tested on the Gray Swan benchmark, where at the maximum reasoning level resistance to all attacks is claimed; beyond that set there is no independent confirmation.
Not stated: deployment on customer premises, work with industrial protocols, quality in Russian and Kazakh, compliance with GOST and ISO documentation formatting requirements. The price is per token and excludes integration cost, human review of the result and organisational change. The model only started rolling out on 18 November 2026, and some access routes are limited by programmes.
Sources
- Introducing Gemini 4 Neon · 2026-11-18T21:35:43.000Z
Neon ranks ahead of comparable models and #3 overall with a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, landing just behind larger, slower models.
- Фото: Linda Spashett Storye_book — Bradford Industrial Museum 030.jpg (CC BY 3.0, Wikimedia Commons)
This material was prepared with AI from the sources listed above. Spotted an inaccuracy? Write to info@105.kz.