GigaChat API or YandexGPT API: which one to build on

In short: who should pick what

If the task is high-volume and simple, such as ticket labelling, field extraction from emails or short templated replies, price it on the light models: GigaChat Lite or Alice AI LLM Flash. The price gap between tiers is measured in multiples, not percent, so check quality on your own examples first and only then look at the tariff.

If you need long context, meaning whole contracts, policies or full conversation histories, Yandex offers Alice AI LLM with a 128k context on the shared instance. YandexGPT Pro 5.1 has 32k. For GigaChat the context depends on the model and the platform: on cloud.ru GigaChat-2-Max is listed with 131k tokens.

If you run an agent with tool calls and several steps, the cost is driven by the number of steps, not by the per-million price. The working pattern is a light model for extraction and formatting and a top-tier model only for the reasoning step. The calculator linked at the bottom of this article turns your volumes into a monthly budget.

Prices as of 29 October 2026

The prices below come from the providers' tariff pages, converted to per-million-token figures, in roubles including VAT, synchronous mode. GigaChat does not distinguish input and output tokens; Yandex does for some models. The full table with the check date lives in the LLM cost calculator.

What is changing in the line-ups right now

This is the most important section: any tariff table goes stale within a quarter, and the shifts below change the calculation itself.

• GigaChat is moving to cloud.ru. The corporate tariff page, updated on 7 October 2026, states that from 1 September 2026 new customers and customers who never bought a package pay for models on the cloud.ru platform. In the cloud.ru catalogue on 29 October GigaChat-2-Max costs ₽569 per million tokens, GigaChat-2 Lite and Pro are not listed, while GigaChat 3 Pro is at ₽73 input / ₽176 output alongside 73 open models. If you are starting a project now, price it against the cloud.ru catalogue, not Sber's older page.
• YandexGPT Lite 5 is retired on 15 November 2026, per the AI Studio models page. For cheap high-volume tasks Yandex now points at Alice AI LLM Flash, which has a different output price and a different context window. If Lite is in your production stack, schedule the migration before mid-November.
• Yandex is betting on Alice AI LLM as the agent model. The AI Studio release notes of 25 November 2025 call it the model of choice for AI agents, and the platform added an OpenAI-compatible Responses API, Vector Store API and MCP Hub. For a team whose code already targets the OpenAI SDK, that reduces the move to changing a base URL and a key.
• First-generation GigaChat models are gone. Requests to GigaChat, GigaChat-Pro and GigaChat-Max are redirected to the second generation. Code with the old identifiers keeps running, but the model behaviour has changed, so re-run your evaluation set.

Access, data and responsibility

Both providers bill in roubles under a contract with a Russian legal entity, and data is processed in Russia. That removes the question of paying for foreign APIs from Russia, which is searched for more often than people assume, and simplifies the conversation with legal about personal data.

Simplifies, not closes. Sending personal data to an external model is still processing under Russia's personal-data law, with everything that implies: a legal basis, a processor agreement, minimisation. The provider supplies the platform; responsibility for lawful processing and for what goes into the prompt stays with the product owner. We design for the minimum: anonymised fields instead of the whole customer record, plus a log of what was sent.

On running a model inside your own perimeter. Yandex offers dedicated instances billed per second, cloud.ru lists open models, and running an open model on your own server is a separate calculation: GPUs, utilisation, maintenance. The calculator deliberately gives no figures for that option, because it depends on your hardware and volume rather than on a price list.

How to choose: one example set, both models

Tariffs answer “how much”, not “does it cope”. So picking a provider in our workshop looks like an experiment rather than a discussion.

• Collect 30 real task examples with known correct answers. In the safety layer of one of our products that is 12 scenarios where the model must fire, 8 where it must stay silent, and 10 borderline ones.
• Run each example five times on every candidate model: 150 calls per model. A single pass proves nothing, because the model does not answer the same way twice.
• Fix the criteria before the run: miss rate, precision, output-format validity, p50 and p95 latency, price per thousand calls. Ours are zero misses, precision at least 0.92, valid JSON 150 out of 150, p50 under 2 seconds, p95 under 3.5 seconds.
• Compare in order of importance, not by a total score. Our order is recall, schema validity, precision, p95, price. A model that is twice as cheap but misses one case in thirty loses.
• Keep the example set in the project and re-run it on every model or tariff change. Providers reshuffle their line-ups, as shown above, and without the set you learn about behaviour changes from your users.

Typical costing mistakes

Three mistakes show up in almost every first estimate that lands on our desk.

• Counting characters instead of tokens. Russian text tokenises more expensively than English, and differently across providers. Check on your own texts with the provider's tokenizer: both expose a token-counting method.
• Forgetting the system prompt and the history. In document chat every request carries the instructions, the retrieved passages and the previous turns. Input tokens usually outnumber output tokens 10 to 30 times, so Alice AI LLM with cheap input and expensive output can come out cheaper than a model with a flat price both ways.
• Ignoring retries. Timeouts, re-calls on an invalid response, the agent's think-call-think loop. Put a retry factor and a per-task step limit into the estimate, otherwise the budget diverges from reality in the first week.