Mistral AI · language model
Saba: pricing, context and benchmarks
Saba is a language model from Mistral AI released on February 17, 2025. It costs $0.2 per million input tokens and $0.6 per million output tokens, with a 32.8K-token context window. No independent score has been published for it so far.
| Price | 22/30 | |
|---|---|---|
| Context | 1/15 | |
| Features | 8/15 |
What you need to know about Saba
Saba was released by Mistral AI (France) on February 17, 2025. It accepts text and files as input and produces text. Its context window reaches 32,768 tokens, larger than 8 % of the models in the registry, and it can write up to 26,214 tokens in one go.
Released more than 1 year ago, it belongs to a generation most vendors have since superseded.
Saba pricing
On OpenRouter, Saba costs $0.2 per million input tokens and $0.6 per million output tokens. Counting three tokens read for every token written, its average price is $0.3 per million, more expensive than 28 % of paid models (2.5 times cheaper than the median of $0.76).
A concrete example: a task that sends 2,000 tokens (about 1,500 words) and receives a 500-token answer costs $0.0007. At 1,000 tasks a day, budget $21 a month. The price calculator redoes this with your own volumes.
Tokens already sent and cached are billed at $0.02 per million.
Performance and benchmarks
Epoch AI does not yet publish a capabilities index for Saba. We therefore show no made-up performance rating: only price, context and features count towards its score.
Features, privacy and local use
- It supports tool calling, so it can drive an AI agent, call APIs or MCP servers.
- It can answer in an enforced JSON format (structured outputs).
- Its knowledge stops at September 2024.
Its weights are not published: it can only be used through the Mistral AI API or a broker such as OpenRouter, which means sending your data off your machines.
Other Mistral AI models
| Model | Released | Avg. price | ECI |
|---|---|---|---|
| Mistral Large 4 | Oct 6, 2026 | $2.07 | 150 |
| Mistral Medium 3.5 | Apr 30, 2026 | $3 | 141.4 |
| Mistral Small 4 | Mar 16, 2026 | $0.26 | — |
| Devstral 2 2512 | Dec 9, 2025 | $0.8 | — |
| Ministral 3 14B 2512 | Dec 2, 2025 | $0.2 | — |
| Ministral 3 8B 2512 | Dec 2, 2025 | $0.15 | — |
| Ministral 3 3B 2512 | Dec 2, 2025 | $0.1 | — |
| Mistral Large 3 2512 | Dec 1, 2025 | $0.75 | — |
Verdict
Saba scores 52/100 on our grid, which only uses published figures.
Strengths
- Low price: cheaper than 72 % of paid models.
- Agent-ready: tool calling and structured JSON output.
Limitations
- No independent score published: performance cannot be verified.
- Older generation, often replaced by a newer model from the same vendor.
Avoid if…
- your data must not leave your servers;
- you analyse long documents in one go.
Alternatives
- Qwen2.5 72B Instruct$0.37
- Qwen3 30B A3B$0.22
- GPT-4o-mini$0.26
- GPT-4o-mini (2024-07-18)$0.26
- Command R (08-2024)$0.26
- Mercury 2$0.38
Comparisons and tools
Learn more
Frequently asked questions
How much does Saba cost?
$0.2 per million input tokens and $0.6 per million output tokens on OpenRouter (checked October 25, 2026). A task with 2,000 tokens read and 500 written costs $0.0007.
How big is Saba's context window?
32,768 tokens, roughly 25,000 words, with a maximum of 26,214 output tokens.
Can Saba run locally?
No, Mistral AI does not publish its weights: it is API-only.
Is Saba suitable for an AI agent?
Technically yes: it supports tool calling and structured outputs. Then check its cost for your volume.
What are the alternatives to Saba?
The closest models on price and performance are Qwen2.5 72B Instruct, Qwen3 30B A3B, and GPT-4o-mini.
Related articles
- Mistral AI: models, pricing and agentsOctober 25, 2026
- LLM benchmarks: how to read model rankingsOctober 24, 2026
- Running a local LLM: hardware, software, and limitationsOctober 24, 2026
- Open-source or closed models for your AI agents?October 5, 2026
Neutrality. No affiliate links. No vendor pays to be reviewed, rated or ranked. Reviews are directed by François Chollet, neutral and factual, and every data point is dated and sourced. How we score