GigaChat API Review: Sber Models, Cloud.ru, Features, Pricing, and a Comparison with OpenAI and DeepSeek

Information current as of September 29, 2026.

In 2026, GigaChat API is considerably more complex than it was just a year ago. Under one name, there are effectively two ways to connect to Sber's models: the classic GigaChat API at api.giga.chat and the new commercial Cloud.ru Evolution Foundation Models infrastructure. This distinction is fundamental for a new project because the models, authentication, prices, and even billing methods differ.

The most interesting model for businesses right now is GigaChat 3.5 Ultra. In the Cloud.ru catalog it is available as ai-sage/GigaChat3.5-432B-A28B-Reasoning, has a 262K-token context window, and costs RUB 96.22 per million input tokens and RUB 288.60 per million generated tokens. The same catalog also includes the compact GigaChat3-10B-A1.8B at just RUB 12.20 per million tokens in either direction.

The old GigaChat API has not gone away. Individuals have access to a large freemium allowance that includes GigaChat 3 Ultra, while existing commercial customers continue to use GigaChat 2 Lite, Pro, and Max. However, since September 1, 2026, Sber has been directing new paid customers to Cloud.ru.

This review covers both options, the current models, image, audio, document, and function support, registration and payment, and—most importantly—how GigaChat's actual prices compare with OpenAI GPT-6 and the direct DeepSeek API.

Two ways to use GigaChat API

The first thing to understand before integrating it is that “GigaChat API” can now refer to two rather different products.

Classic GigaChat API

The primary address is:

https://api.giga.chat

This is the API documented on Sber Developers. It provides access to:

GigaChat 2 Lite;
GigaChat 2 Pro;
GigaChat 2 Max;
GigaChat 3 Ultra for individuals on the freemium plan;
embeddings;
file handling;
image and audio analysis;
function calling;
image generation;
3D model generation.

Authentication is proprietary: the application first obtains an OAuth access token valid for 30 minutes and then calls /v1/chat/completions.

Cloud.ru Evolution Foundation Models

For new paid connections, since September 1, 2026, Sber directs users to Cloud.ru.

Endpoint:

https://foundation-models.api.cloud.ru/v1

This service uses a standard API key and an OpenAI-compatible interface.

Foundation Models offers not only GigaChat, but also third-party models such as Claude, DeepSeek, Kimi, GLM, Qwen, GPT-OSS, and others. In effect, Cloud.ru is becoming a Russian multi-model API gateway.

For a new commercial product, I would first look at Cloud.ru: integration is simpler, the current GigaChat 3.5 is available, and you can compare several models within one account.

The move to Cloud.ru starting in September 2026

This change is recent, so many online guides are already out of date.

GigaChat's official documentation states directly: starting September 1, 2026, new customers can pay to use models on Cloud.ru.

The classic Sber account remains relevant in two cases:

an individual is using the free freemium plan;
the account was already connected to paid GigaChat API.

For existing customers, the old plans continue to exist during the transition. But if you are building a commercial integration now, Cloud.ru should be considered the primary current route for paid access.

GigaChat 3.5 Ultra

Cloud.ru opened commercial access to GigaChat 3.5 Ultra in July 2026.

In the Foundation Models catalog, the model is currently listed as:

ai-sage/GigaChat3.5-432B-A28B-Reasoning

Key characteristics:

context:          262K
total parameters: 432B
active parameters: about 28B
architecture:     Mixture-of-Experts

Cloud.ru positions the model for:

building AI assistants;
programming;
document analysis;
mathematical tasks;
financial calculations;
complex agentic scenarios.

The model is much more compact than the previous GigaChat 3.1 Ultra flagship: 432 billion parameters versus 700 billion in its predecessor. Thanks to MoE, only part of the model works on each token.

For API users, the more important point is that this is currently the most advanced GigaChat model that can be connected to a commercial product in a straightforward pay-as-you-go setup.

GigaChat3-10B-A1.8B

The Cloud.ru catalog also includes a much more compact model:

ai-sage/GigaChat3-10B-A1.8B

Context:

262K

Price:

input:   RUB 12.20 / 1M tokens
output:  RUB 12.20 / 1M tokens

This model is interesting for high-volume, well-defined tasks:

classification;
extraction;
short-form generation;
processing a large stream of inquiries;
routing;
simple assistants.

Its price is already in roughly the same range as the cheapest international APIs.

If a business does not need a flagship model for every request, an architecture that uses the inexpensive 10B model by default and escalates to GigaChat 3.5 Ultra when needed can be much more economical.

GigaChat 3 Ultra in the classic API

The name is easy to confuse with GigaChat 3.5 Ultra.

The classic API includes:

GigaChat-3-Ultra

It was introduced in July 2026, has a 128K-token context window, and supports functions, image generation and analysis, and the audio modality.

At the time of publication, there is an important limitation: GigaChat 3 Ultra is available only to individuals on the freemium plan. Paid plans for individuals and legal entities do not apply to this model.

So GigaChat 3 Ultra through api.giga.chat is a great way to try the new generation for free, but it is not the primary option for a commercial backend.

For commercial use, GigaChat 3.5 through Cloud.ru is currently more relevant.

GigaChat 2 Lite, Pro, and Max

The classic paid GigaChat API lineup consists of three models.

Model	Context	Input	Main use case
GigaChat 2 Lite	128K	text	simple, high-volume tasks
GigaChat 2 Pro	128K	text, images, audio	following instructions, documents, applied tasks
GigaChat 2 Max	128K	text, images, audio	the most demanding tasks in generation 2

Pro and Max can analyze images and audio files. All models support function calling, and built-in functions can work with documents and generate additional content.

For existing paid customers on the old API, prices are:

GigaChat 2 Lite:   RUB 65 / 1M tokens
GigaChat 2 Pro:    RUB 500 / 1M tokens
GigaChat 2 Max:    RUB 650 / 1M tokens

There is no separate input/output price here: billing is based on the total number of tokens.

Cloud.ru uses a different approach for new customers, with separate prices for input and generated tokens.

Freemium for individuals

GigaChat API stands out for its very large free allowance for individuals.

As of late September 2026, freemium includes 365 million tokens over 12 months:

Model	Free allowance
GigaChat 2 Lite	250 million
GigaChat 2 Pro	40 million
GigaChat 2 Max	25 million
GigaChat 3 Ultra	50 million

The allowance resets every 12 months.

That is a very generous offer for learning, experimentation, and pet projects.

There is, however, a fundamental restriction: official freemium is for personal, non-commercial use. If generated content is used commercially, Sber requires paid access.

So the hundreds of millions of free tokens cannot simply be used as the backend for a commercial SaaS.

Current GigaChat prices through Cloud.ru

For a new commercial connection, Evolution Foundation Models pricing is the most relevant.

Model	Input / 1M	Output / 1M	Context
GigaChat3.5-432B-A28B-Reasoning	RUB 96.22	RUB 288.60	262K
GigaChat3-10B-A1.8B	RUB 12.20	RUB 12.20	262K
GigaChat-2-Max	RUB 569.34	RUB 569.34	131K

Cloud.ru prices are listed in rubles per million tokens.

The most notable point is how inexpensive GigaChat 3.5 Ultra has become relative to the old GigaChat 2 Max. The new model also has twice the context window and is aimed at more complex tasks.

For comparison, the old Max costs about RUB 569 per million tokens in either direction on Cloud.ru, while 3.5 Ultra costs RUB 96 input and RUB 289 output.

Comparing prices with OpenAI and DeepSeek

To make the comparison fairer, let's convert international prices to rubles.

The calculations below use the Bank of Russia's official exchange rate for September 29, 2026:

1 USD = RUB 84.4075

First, current OpenAI API prices:

Model	Input / 1M	Cached / 1M	Output / 1M
GPT-6 Luna	$0.10	$0.01	$0.50
GPT-6 Sol	$2.00	$0.20	$10.00
GPT-6 Astra	$10.00	$1.00	$50.00

In rubles at the Bank of Russia rate:

Model	Input / 1M	Output / 1M
GPT-6 Luna	~RUB 8.44	~RUB 42.20
GPT-6 Sol	~RUB 168.82	~RUB 844.08
GPT-6 Astra	~RUB 844.08	~RUB 4,220.38

Now consider the direct DeepSeek V4.1 Flash API.

Off-peak:

input:  $0.15 / 1M  ≈ RUB 12.66
output: $0.60 / 1M  ≈ RUB 50.64

Peak:

input:  $0.30 / 1M  ≈ RUB 25.32
output: $1.20 / 1M  ≈ RUB 101.29

And finally, GigaChat:

Model	Input / 1M	Output / 1M
GigaChat3 10B	RUB 12.20	RUB 12.20
GigaChat 3.5 Ultra	RUB 96.22	RUB 288.60

What the price per million tokens tells us

GigaChat3 10B looks very competitive. Its input price is almost the same as direct DeepSeek off-peak and only slightly above GPT-6 Luna, while its output is noticeably cheaper than both.

GigaChat 3.5 Ultra sits in a different position.

For input, it is:

about 11 times more expensive than GPT-6 Luna;
about 7.6 times more expensive than DeepSeek off-peak;
about 3.8 times more expensive than DeepSeek peak;
about 43% cheaper than GPT-6 Sol.

For output, it is:

about 6.8 times more expensive than GPT-6 Luna;
about 5.7 times more expensive than DeepSeek off-peak;
about 2.8 times more expensive than DeepSeek peak;
about 66% cheaper than GPT-6 Sol.

By price, GigaChat 3.5 Ultra therefore occupies an interesting middle ground: it is significantly more expensive than the cheapest global APIs, but considerably cheaper than OpenAI Sol.

Example: 100,000 input + 10,000 output tokens

For a more tangible comparison, consider a large request:

100,000 input tokens
10,000 output tokens

No cache or additional tools.

Model	Request cost
OpenAI GPT-6 Luna	~RUB 1.27
GigaChat3 10B	~RUB 1.34
DeepSeek V4.1 Flash off-peak	~RUB 1.77
DeepSeek V4.1 Flash peak	~RUB 3.55
GigaChat 3.5 Ultra	~RUB 12.51
OpenAI GPT-6 Sol	~RUB 25.32
OpenAI GPT-6 Astra	~RUB 126.61

This table makes GigaChat's two distinct segments especially clear.

GigaChat3 10B competes on price with Luna and DeepSeek.

GigaChat 3.5 Ultra is in a more expensive class, but at this input/output ratio it costs about half as much as GPT-6 Sol.

The token price is not the price of a result

Looking only at the price per million tokens is not enough.

Different models use different tokenizers. The same Russian document can take a different number of tokens in GigaChat, OpenAI, and DeepSeek.

Also, a more capable model might solve a task on the first try, while a cheaper one might need several retries, an extra check, or a second call to a stronger model.

For a real project, it is better to calculate:

the cost of processing one inquiry
the cost of successfully analyzing one document
the cost of completing one agent workflow

rather than just $ / 1M tokens.

Why comparing GigaChat 3.5 Ultra with Luna is not quite fair

Luna looks considerably cheaper, but these models are positioned differently.

GPT-6 Luna is OpenAI's cheapest model for focused, high-volume workloads.

GigaChat 3.5 Ultra is the top reasoning/agentic model in the Russian lineup.

Within Cloud.ru, GigaChat3-10B is a more direct competitor to Luna.

GigaChat 3.5 Ultra is more reasonably compared with stronger models such as GPT-6 Sol, DeepSeek reasoning scenarios, or other large agentic LLMs.

That is why there is little sense in sending every request from an application with a large volume of simple queries to GigaChat 3.5 Ultra. Routing between a cheap and a powerful model can reduce costs several-fold.

OpenAI-compatible API in Cloud.ru

For a new project, this is probably the easiest way to integrate GigaChat.

Cloud.ru provides this endpoint:

https://foundation-models.api.cloud.ru/v1

The API is compatible with OpenAI Chat Completions.

Python example:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ["API_KEY"],
    base_url="https://foundation-models.api.cloud.ru/v1"
)

response = client.chat.completions.create(
    model="ai-sage/GigaChat3-10B-A1.8B",
    messages=[
        {
            "role": "user",
            "content": "Briefly explain dependency injection."
        }
    ]
)

print(response.choices[0].message.content)

For GigaChat 3.5, select the corresponding model ID from the Foundation Models catalog.

This interface is much simpler than the classic GigaChat API because it does not require obtaining a short-lived OAuth token in advance.

You can also use the same client to switch to other Cloud.ru models.

Classic GigaChat API

If you specifically need api.giga.chat, the flow is different.

Primary endpoint:

https://api.giga.chat

You first need to obtain an access token before calling a model.

The authorization key contains encoded Client ID and Client Secret values.

OAuth request:

curl -X POST \
  'https://ngw.devices.sberbank.ru:9443/api/v2/oauth' \
  -H 'Authorization: Basic <AUTHORIZATION_KEY>' \
  -H 'RqUID: <UUID>' \
  -H 'Content-Type: application/x-www-form-urlencoded' \
  --data-urlencode 'scope=GIGACHAT_API_PERS'

Three scopes are available:

GIGACHAT_API_PERS — individuals
GIGACHAT_API_B2B  — sole proprietors and legal entities with prepayment
GIGACHAT_API_CORP — sole proprietors and legal entities on pay-as-you-go

The access token is valid for 30 minutes.

Then send a regular request:

curl https://api.giga.chat/v1/chat/completions \
  -H "Authorization: Bearer $ACCESS_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "GigaChat-2-Pro",
    "messages": [
      {
        "role": "user",
        "content": "Explain dependency injection."
      }
    ]
  }'

For a new commercial application, Cloud.ru is noticeably simpler from an infrastructure perspective.

Function calling

All current GigaChat models support user-defined functions.

The flow is similar to OpenAI function calling: the application describes a function, the model decides when to call it, and generates the arguments.

For example, you can give the model these functions:

get_order
find_customer
create_ticket

GigaChat does not execute a user-defined function itself. The backend receives the arguments, verifies the user, performs the action, and returns the result to the model.

This can be used to build:

customer support;
CRM assistants;
AI operators;
business agents;
analytics systems.

Critical operations still need to be checked by the application's regular code. Do not let an LLM decide user permissions or financial limits.

GigaChat built-in functions

The classic GigaChat API includes several functions that are executed by the service itself.

As of late September 2026, the documented functions are:

text2image

Image generation.

get_file_content

Retrieving the contents of an uploaded document.

text2model3d

Creating a 3D model in FBX format.

This is an unusual set of features for an LLM API. 3D model generation is especially notable: OpenAI, Claude, and DeepSeek do not have a similar built-in tool in their main text APIs.

Image generation

If you enable:

{
  "function_call": "auto"
}

GigaChat can determine that a request needs an image and call text2image itself.

The API returns a file_id, after which the image is downloaded in a separate request from file storage.

In other words, image generation is built directly into the conversation flow.

In Cloud.ru Foundation Models, specific media capabilities depend on the selected model and service, so the old GigaChat API text2image flow should not automatically be assumed to work with OpenAI-compatible Foundation Models.

3D model generation

The GigaChat API has a built-in:

text2model3d

It generates a 3D object in FBX format.

The application receives a model ID and then downloads the file from GigaChat storage.

This is a niche feature, but it can be useful for e-commerce, games, AR/VR, and rapid prototyping without a separate specialized 3D API.

Working with files

GigaChat API has its own file storage.

Supported formats include:

Documents

TXT, DOC, DOCX, PDF, EPUB, PPT, PPTX, XLSX

Images

JPEG, PNG, TIFF, BMP

Audio

MP3, MP4, M4A, WAV, WebM, OGG, OPUS

Single-file limits:

document:  up to 40 MB
image:     up to 15 MB
audio:     up to 35 MB

Pro and Max can analyze images and audio, while text documents are passed through get_file_content.

An important caveat: a file can fit in storage, but its extracted contents still have to fit in the model's context window.

Embeddings

The classic GigaChat API offers embedding models.

For existing customers, the current price is:

RUB 14 / 1M tokens

This is based on the corporate package of 1 billion tokens for RUB 14,000.

That is a very low price for building your own semantic search or RAG system.

For a new commercial customer, however, check the current embedding catalog in Cloud.ru separately, since GigaChat's commercial API offering is now moving there.

Batch and asynchronous processing

GigaChat API supports batch processing.

Existing corporate customers on the old API also have a separate asynchronous rate—about half the synchronous price:

Lite: RUB 0.0325 / 1,000
Pro:  RUB 0.25 / 1,000
Max:  RUB 0.325 / 1,000

This is a good option for:

bulk classification;
generating descriptions;
offline extraction;
processing catalogs;
analytics;
preparing content in the background.

Cloud.ru Evolution also uses pay-as-you-go billing, and the available endpoints depend on the specific Foundation Model.

Registering as a new customer

For a new paid project, Cloud.ru is currently the simpler route.

Steps:

Create a Cloud.ru account.
Open Evolution → Foundation Models.
Select GigaChat or another model.
Create an API key.
Save the Key Secret—it will not be shown again after you close the window.
Use the OpenAI-compatible endpoint.

For the classic GigaChat API, an individual can still create a project in Sber Studio and obtain an Authorization Key for freemium.

Legal entities that connected earlier continue using their old setup during the transition period.

Paying on Cloud.ru

Foundation Models uses pay-as-you-go billing: you pay for the input and generated tokens actually used.

Individuals can top up their balance by bank card.

For legal entities:

with prepayment, bank card and bank transfer are available;
with postpayment, an invoice is issued at the end of the month.

For Russian businesses, this is much simpler than using the direct OpenAI or DeepSeek API: there is no need to find a foreign bank card or a separate payment intermediary.

There is no minimum spend on Foundation Models

The old corporate GigaChat API had a minimum monthly payment of RUB 600 if the service was used at all.

Evolution Foundation Models uses standard usage-based billing: the customer pays for the tokens and other resources actually consumed.

That is another reason why Cloud.ru looks more logical than legacy plans for a new small project.

Security and data

For the classic GigaChat API, Sber states explicitly that, by default, user requests and responses are not stored or used to train the model.

That is an important distinction for enterprise use cases compared with a consumer AI chat.

Cloud.ru Foundation Models also separates models into internal and external categories.

Internal models run within Cloud.ru infrastructure; Cloud.ru does not log or store user data.

External models are called through a third-party provider's API, so data may leave Cloud.ru.

GigaChat 3.5 is listed in the Cloud.ru catalog as the internal ai-sage model, which makes it particularly interesting to Russian companies for locally hosted inference.

Guardrails

Cloud.ru offers a separate Foundation Models Guardrails layer.

Before a prompt is sent, it can find and mask:

logins and passwords;
API keys;
access tokens;
IP addresses;
email addresses;
Russian phone numbers;
SNILS personal insurance numbers;
taxpayer identification numbers (INN);
primary state registration numbers (OGRN);
bank cards;
other personal identifiers.

Original values can be restored after the response.

Guardrails is particularly useful with external models, when a request physically leaves Cloud.ru infrastructure.

For GigaChat running inside Cloud.ru, it provides an additional layer of protection for the corporate pipeline.

Open weights and self-hosting

GigaChat 3.5 Ultra is also interesting because its model weights have been published openly.

This means it can be used through Cloud.ru or deployed on suitable GPU infrastructure of your own.

However, 432B total parameters is a very large model. Having only about 28B active parameters per token reduces compute cost, but does not turn the checkpoint into an ordinary 28B model.

For a small company, a Cloud API is almost certainly simpler than its own GPU cluster.

For a large business with a consistently high workload and requirements for a closed environment, compare the total cost of ownership for self-hosting and Cloud.ru.

GigaChat API in PHP through Cloud.ru

For a new PHP project, the easiest option is OpenAI-compatible Foundation Models.

<?php

$apiKey = $_ENV['CLOUDRU_FOUNDATION_API_KEY'];

$payload = [
    'model' => 'ai-sage/GigaChat3.5-432B-A28B-Reasoning',
    'messages' => [
        [
            'role' => 'user',
            'content' => 'Briefly explain what dependency injection is.',
        ],
    ],
];

$ch = curl_init(
    'https://foundation-models.api.cloud.ru/v1/chat/completions'
);

curl_setopt_array($ch, [
    CURLOPT_POST => true,
    CURLOPT_RETURNTRANSFER => true,
    CURLOPT_HTTPHEADER => [
        'Authorization: Bearer ' . $apiKey,
        'Content-Type: application/json',
    ],
    CURLOPT_POSTFIELDS => json_encode(
        $payload,
        JSON_UNESCAPED_UNICODE
    ),
]);

$response = curl_exec($ch);

if ($response === false) {
    throw new RuntimeException(curl_error($ch));
}

$data = json_decode(
    $response,
    true,
    flags: JSON_THROW_ON_ERROR
);

echo $data['choices'][0]['message']['content'] ?? '';

In Laravel:

$response = Http::withToken(
        config('services.cloudru_foundation.key')
    )
    ->post(
        'https://foundation-models.api.cloud.ru/v1/chat/completions',
        [
            'model' => 'ai-sage/GigaChat3.5-432B-A28B-Reasoning',
            'messages' => [
                [
                    'role' => 'user',
                    'content' => 'Briefly explain what dependency injection is.',
                ],
            ],
        ]
    );

$text = $response->throw()
    ->json('choices.0.message.content');

To switch between GigaChat, DeepSeek, Qwen, and other Foundation Models, you can keep the same client layer and change the model ID.

Classic GigaChat API in PHP

If you use the old GigaChat API, account for OAuth.

The backend first obtains an access token:

<?php

$authorizationKey = $_ENV['GIGACHAT_AUTHORIZATION_KEY'];
$rqUid = uuid_create(UUID_TYPE_RANDOM);

$ch = curl_init(
    'https://ngw.devices.sberbank.ru:9443/api/v2/oauth'
);

curl_setopt_array($ch, [
    CURLOPT_POST => true,
    CURLOPT_RETURNTRANSFER => true,
    CURLOPT_HTTPHEADER => [
        'Authorization: Basic ' . $authorizationKey,
        'RqUID: ' . $rqUid,
        'Content-Type: application/x-www-form-urlencoded',
    ],
    CURLOPT_POSTFIELDS => http_build_query([
        'scope' => 'GIGACHAT_API_PERS',
    ]),
]);

$response = curl_exec($ch);
$data = json_decode($response, true);

$accessToken = $data['access_token'];

The token is valid for 30 minutes, so in production it should be cached and refreshed based on its expiration rather than obtained before every user request.

Then call /v1/chat/completions with a regular Bearer token.

This extra step is one reason the new Cloud.ru endpoint is simpler for a fresh integration.

Context comparison

Context window sizes differ considerably:

Model	Context
OpenAI GPT-6 Luna / Sol / Astra	~1.05M
DeepSeek V4.1 Flash	1M
GigaChat 3.5 Ultra on Cloud.ru	262K
GigaChat3 10B	262K
GigaChat 2 Lite / Pro / Max	128K
GigaChat 3 Ultra freemium	128K

For a regular chatbot or enterprise RAG application, 262K is already a lot.

But when analyzing a huge codebase, several hundred pages of documents, or a very long agentic loop, OpenAI and DeepSeek provide substantially more room before retrieval or compaction becomes necessary.

What GigaChat does especially well

First, Russian language and local context. The model is developed with Russian-language tasks in mind, rather than merely supporting Russian as one of many languages.

Second, commercial access in Russia. The new API can officially be paid for through Cloud.ru in rubles and used under a Russian contract with Russian documents.

Third, the very inexpensive GigaChat3 10B. At RUB 12.20 per million tokens, it is interesting for high-volume operations.

Fourth, GigaChat 3.5 Ultra is noticeably cheaper than GPT-6 Sol for a typical request, even though its headline price cannot compete with Luna or direct DeepSeek.

Fifth, the classic API's built-in multimodal functions: images, documents, audio, and even 3D model generation.

Sixth, the option to get open weights and move to your own inference if needed.

Limitations

The main downside right now is some fragmentation across the platform.

Online, you will encounter all of these at once:

GigaChat API
GigaChat 2
GigaChat 3 Ultra
GigaChat 3.5 Ultra
Cloud.ru Foundation Models
ai-sage/GigaChat3.5-432B-A28B-Reasoning

It is not obvious to a newcomer that these refer to different access methods and model generations.

The second downside is that the flagship commercial 3.5 Ultra has a 262K context window, while OpenAI and DeepSeek already offer about one million.

Third, the direct GigaChat API uses its own OAuth flow and the Russian Ministry of Digital Development's root certificates, while many international APIs require only one static Bearer key.

Fourth, 3.5 Ultra costs more than ultra-cheap APIs such as GPT-6 Luna and DeepSeek V4.1 Flash.

Finally, new commercial connections are now better built through Cloud.ru, so older GigaChat API articles and examples can quickly become outdated.

Which model should you choose?

If you need a low-cost Russian backend for high-volume requests, start by testing:

GigaChat3-10B-A1.8B

It costs just RUB 12.20 per million tokens and has a 262K context window.

If you need complex reasoning, coding, documents, and agents:

GigaChat3.5-432B-A28B-Reasoning

is the more natural choice.

If you are an individual who simply wants to test new models for free, the classic GigaChat API provides a large freemium allowance and GigaChat 3 Ultra.

If you already have a commercial legacy integration with GigaChat 2, there is no urgent need to rewrite it solely because Cloud.ru has appeared. But for a new project, it makes sense to evaluate Foundation Models from the start.

Conclusion

In 2026, GigaChat API is no longer just one endpoint and three Lite/Pro/Max models.

The old api.giga.chat continues to work, gives individuals 365 million free tokens a year, can analyze documents, images, and audio, and can generate images and even 3D models. But for new paid customers from September 1, 2026, Cloud.ru Evolution Foundation Models is becoming the primary route.

GigaChat 3.5 Ultra is already available there with a 262K context window at RUB 96.22 input and RUB 288.60 output per million tokens, while the compact GigaChat3 10B costs just RUB 12.20 per million tokens in either direction.

The pure token-price comparison is mixed. GPT-6 Luna and the direct DeepSeek V4.1 Flash API are significantly cheaper than GigaChat 3.5 Ultra. But in our example of 100K input + 10K output, GigaChat 3.5 Ultra costs about RUB 12.51 versus RUB 25.32 for GPT-6 Sol.

GigaChat3 10B is right alongside the cheapest international models: about RUB 1.34 for the same request, compared with about RUB 1.27 for Luna and RUB 1.77 for DeepSeek V4.1 Flash off-peak.

So it makes more sense to evaluate GigaChat as more than one model to compare with one GPT. The Russian ecosystem already has at least two very different price tiers: a very cheap 10B model for high-volume operations and a large 3.5 Ultra for complex agentic tasks.

Add official ruble billing, Russian infrastructure, the OpenAI-compatible Cloud.ru API, Guardrails, and the ability to deploy open weights yourself, and GigaChat becomes a practical option not only for projects that “cannot use foreign APIs,” but also for an ordinary multi-model architecture—provided the choice is confirmed with your own evaluations on real tasks.

Official sources

GigaChat API: https://developers.sber.ru/docs/ru/gigachat/api/main
GigaChat API authentication: https://developers.sber.ru/docs/ru/gigachat/api/reference/rest/gigachat-api
GigaChat models: https://developers.sber.ru/docs/ru/gigachat/guides/selecting-a-model
GigaChat 3 Ultra: https://developers.sber.ru/docs/ru/gigachat/models/gigachat-3-ultra
Model updates: https://developers.sber.ru/docs/ru/gigachat/models/updates
GigaChat 2 Lite: https://developers.sber.ru/docs/ru/gigachat/models/gigachat-2-lite
GigaChat 2 Pro: https://developers.sber.ru/docs/ru/gigachat/models/gigachat-2-pro
GigaChat 2 Max: https://developers.sber.ru/docs/ru/gigachat/models/gigachat-2-max
Individual plans: https://developers.sber.ru/docs/ru/gigachat/tariffs/individual-tariffs
Legal entity plans: https://developers.sber.ru/docs/ru/gigachat/tariffs/legal-tariffs
Commercial use: https://developers.sber.ru/docs/ru/gigachat/tariffs/commercial
Working with files: https://developers.sber.ru/docs/ru/gigachat/guides/working-with-files
Built-in functions: https://developers.sber.ru/docs/ru/gigachat/guides/functions/calling-builtin-functions
Evolution Foundation Models: https://cloud.ru/products/evolution-foundation-models
Foundation Models catalog: https://cloud.ru/products/evolution-ai-factory/catalog-foundation-models
Foundation Models quickstart: https://cloud.ru/docs/foundation-models/ug/topics/quickstart
Foundation Models API: https://cloud.ru/docs/foundation-models/ug/topics/api-ref
Foundation Models pricing: https://cloud.ru/documents/tariffs/evolution/foundation-models
GigaChat 3.5 Ultra on Cloud.ru: https://cloud.ru/blog/cloud-ru-pervym-otkryl-dostup-k-gigachat-3-5-ultra
Guardrails: https://cloud.ru/docs/foundation-models/ug/topics/concepts__guardrails
OpenAI API pricing: https://developers.openai.com/api/docs/pricing
DeepSeek API pricing: https://api-docs.deepseek.com/quick_start/pricing/
Bank of Russia exchange rate: https://www.cbr.ru/eng/currency_base/daily/

Date of publication: 29 September 2026

GigaChat API Review: Sber Models, Cloud.ru, Features, Pricing, and a Comparison with OpenAI and DeepSeek

Our projects

Anilau
Web development and digital product launch.
Botmarketing
Telegram bots, mini apps, storefronts and CRM for small businesses.
vietnam.anilau.com
Listings, local services and practical guides to Vietnam.
bali.anilau.com
A marketplace for goods and services in Bali.
ceylon.anilau.com
Listings, services and practical information about Sri Lanka.
mauricetop.anilau.com
A platform for listings and information about life in Mauritius.
funlab
A platform for creating and playing AI-generated games.
aura
An AI mood diary and creative space.
drained
A Telegram Mini App for daily fatigue check-ins and recovery.
vietinfodesk
Practical guides, services and help for life in Vietnam.
frau
A cozy Nha Trang cafe profile featuring waffles, breakfast and drinks.

We work in partnership with creative agency Deep.

GigaChat API Review: Sber Models, Cloud.ru, Features, Pricing, and a Comparison with OpenAI and DeepSeek

Two ways to use GigaChat API

Classic GigaChat API

Cloud.ru Evolution Foundation Models

The move to Cloud.ru starting in September 2026

GigaChat 3.5 Ultra

GigaChat3-10B-A1.8B

GigaChat 3 Ultra in the classic API

GigaChat 2 Lite, Pro, and Max

Freemium for individuals

Current GigaChat prices through Cloud.ru

Comparing prices with OpenAI and DeepSeek

What the price per million tokens tells us

Example: 100,000 input + 10,000 output tokens

The token price is not the price of a result

Why comparing GigaChat 3.5 Ultra with Luna is not quite fair

OpenAI-compatible API in Cloud.ru

Classic GigaChat API

Function calling

GigaChat built-in functions

text2image

get_file_content

text2model3d

Image generation

3D model generation

Working with files

Documents

Images

Audio

Embeddings

Batch and asynchronous processing

Registering as a new customer

Paying on Cloud.ru

There is no minimum spend on Foundation Models

Security and data

Guardrails

Open weights and self-hosting

GigaChat API in PHP through Cloud.ru

Classic GigaChat API in PHP

Context comparison

What GigaChat does especially well

Limitations

Which model should you choose?

Conclusion

Official sources

Our projects

Try our plugins for Codex and Claude