Mistral has previewed a model with roughly 1.05 trillion parameters and says it will release the weights openly. Developers can try the model today. They can’t download it yet.
The French company announced Mistral Large 4 on 6 October. According to a developer newsletter summarising the launch, the model is available through Mistral Studio and the API. The company says weights will follow at the end of the month, after more red-team testing.
What the model is
Large 4 uses a mixture-of-experts design. That means only a fraction of its parameters is active when it processes each token. Reports agree on the total of about 1.05 trillion. They differ on the active count: one write-up puts it at about 49 billion, while the model card as described by another newsletter lists 52 billion.
The model card, as summarised by AI Dev Weekly, lists a context window of 1 million tokens, multimodal input, structured outputs and function calling. It also includes a vision encoder. One commentator has reported that the model has the codename “Le Chonk” and is aimed at the frontier tier. That framing comes from a newsletter, not from Mistral, and no independent benchmark results appeared in the material reviewed.
Preview, not release
The distinction matters. AI Dev Weekly stresses that “open-weight” here is a promise, not a current fact. Until the weights are published, Large 4 is a hosted service like any other. Mistral has not said what licence will apply. The sources reviewed do not say.
Pricing is also reported with some variation. According to the newsletter, launch rates are $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. It lists the regular rates as $1.36, $0.14 and $4.18, and says Mistral’s changelog describes the 50 percent launch discount as lasting two weeks. The aggregator LLM Gateway lists the regular $1.36 and $4.18 figures.
Anthropic cuts its small-model price
The next day, 7 October, Anthropic launched Claude Haiku 5.5. It is described as the company’s fastest model, aimed at high-volume work such as classification, routing, extraction and subtasks handed off by larger models.
LLM Gateway lists prices of $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens. Above that length the rates rise to $0.50 and $2.50. The model has a 1 million token context window and a 128,000-token output limit. The Substack newsletter calls it the first Haiku with effort controls, so users can set how much reasoning it applies. It also reports that Anthropic halved cache-read pricing on Sonnet 5.5.
The newsletter describes the Haiku change as a price cut of roughly 90 percent. Anthropic’s own announcement was not among the pages reviewed, so that figure is the newsletter’s calculation.
Other releases this week
- Z.ai released GLM 5.3 Fast on 7 October, a faster variant with a 1 million token context.
- Reflection released Beam, a 501-billion-parameter mixture-of-experts model with about 23 billion active parameters. It is reported to carry an Apache 2.0 licence, with weights expected this month.
- Google added Gemini Nano Banana 2.1, and xAI released Grok Imagine Video 1.5 Lite.
Taken together, the week shows large open-weight models and cheap small models arriving at the same time. Whether Large 4 performs at the level its size implies will only be clear once outside testers get the weights and publish results. Mistral’s end-of-month target is the date to watch.