DeepSeek has released V4.1-Flash-1017 with a million-token context window and published API rates, while a reported Replit change and Mistral preview offer additional signals for developers.
DeepSeek’s latest model release gives developers another option to test on long-context and tool-using tasks, with detailed API prices published alongside the launch. V4.1-Flash-1017 went live October 17 across DeepSeek’s website, app and developer platform, according to Bloomberg. DeepSeek says it completes tasks faster and produces shorter outputs than its predecessor. Those are company claims; the available material does not provide independent benchmark results.
DeepSeek’s documentation lists a one-million-token context window, a maximum output of 384,000 tokens, vision, tool calls, and both thinking and non-thinking modes. These capabilities may suit experiments involving large code repositories, long documents or agent workflows, but specifications alone do not establish how well the model performs on any particular task.
At peak rates, input tokens that are cache misses cost $0.30 per million and output tokens cost $1.20 per million. Off-peak rates are $0.15 and $0.60, respectively. DeepSeek lists Search and Fetch at $0.75 per 1,000 successful requests. Those calls remain free through DeepSeek Harness, Claude Code and Codex until midnight Beijing time on November 1, according to the company’s documentation.
For developers comparing model services, the release puts context capacity and price in plain view. A large context window may help when an application needs to supply extensive material at once, while posted token rates can matter for high-volume workloads. But maximum capacity is not necessarily a sensible target for every task, and total workflow costs can include repeated model calls, retrieval and other infrastructure. The available material does not establish the tools’ relative reliability, privacy protections or performance.
A separate reported change affects the default model inside a coding product. TechCrunch reports that Replit has made OpenAI its default provider for Replit Agent. The change is narrower than an Anthropic exit: users can still access Anthropic models. A default can steer what users try first, but the report does not disclose the commercial terms or establish effects on price, latency or coding quality. It is a distribution decision, not proof that one model is best for every task.
Mistral is also drawing attention, though the available details come from an aggregation of coverage rather than a directly accessible primary announcement. The material describes a public preview of Mistral Large 4, also called ML4, for developers and cybersecurity specialists, with full weights expected October 27. It characterizes the model as having one trillion parameters and reports training on 4,000 Nvidia Grace Blackwell GPUs over two months. These are reported specifications and plans, not confirmation that the weights are already available.
The same aggregated material cites a September financing round for Mistral, but does not verify a new funding announcement today. The available reporting likewise does not substantiate a fresh startup round above $10 million among these developments. The clearest confirmed news is DeepSeek’s release and pricing, alongside TechCrunch’s account of Replit’s provider choice. For developers weighing OpenAI, Anthropic and alternatives, measured performance, clear terms and meaningful provider choice will matter more than headline specifications.

