PUBLIC INTELLIGENCE

Selected AI market intelligence

Recent moves by company. Sign up for the full timeline, company profiles and trend analysis.

DeepSeek Platform · milestone

DeepSeek Pricing Page Overhaul: Peak/Off-Peak Pricing and Feature Matrix

DeepSeek has comprehensively updated its pricing page, introducing a peak/off-peak pricing model where off-peak rates are half of peak rates, alongside differentiated pricing for cache hits and misses across Flash and V4 Pro models. The update also includes a new feature matrix detailing capabilities like JSON Output, Tool Calls, and Vision (Flash only), clarifies Thinking Mode usage, and outlines deduction rules prioritizing granted balances. This overhaul significantly enhances pricing transparency and provides developers with granular cost optimization and model selection insights.

DeepSeek Platform · major

DeepSeek Restructures Pricing Page into Quickstart Guide

DeepSeek completely replaced its original Models & Pricing page with a 'Your First API Call' quickstart guide, including API key generation, base_url configuration, and multi-language call examples; the original pricing table is no longer displayed. This significantly lowers the onboarding barrier for new developers and improves the integration experience, but temporarily reduces public pricing transparency, affecting competitor price comparisons and user cost evaluations. Developers get easier access to onboarding guides but must find new pages for pricing. It remains unclear if pricing info was entirely removed, moved to another URL, or if prices themselves changed.

DeepSeek Platform · major

DeepSeek Launches V4.1 Flash with Price Cuts and Native Vision

DeepSeek launched the V4.1 Flash model, replacing the previous V4 Flash series. The new model reportedly surpasses V4 Pro in performance, cost, and speed, and natively integrates vision capabilities, eliminating the need for a separate vision-exp model. Pricing has been significantly reduced, with input cache hit prices dropping to $0.003/M tokens. This makes Flash a single, low-cost model for both text and vision, raising the cost-effectiveness standard for multimodal APIs.

DeepSeek Platform · major

DeepSeek API Adds reasoning_effort Parameter for Inference Control

DeepSeek showcased the new reasoning_effort parameter (supporting values like 'high') in its API call examples, used alongside the thinking parameter to control inference intensity. This suggests DeepSeek is benchmarking against OpenAI's o1/o3 series in reasoning control, allowing developers to flexibly trade off reasoning depth against latency/cost, significantly boosting its competitiveness in the reasoning model market. Developers can adjust response quality and compute costs accordingly. However, the full range of values (e.g., low/medium/high), performance differences across tiers, and corresponding token consumption rules are not yet detailed.

DeepSeek Platform · major

DeepSeek Documentation Adds Anthropic API Format Compatibility

DeepSeek's quickstart page now explicitly provides an Anthropic API format base_url and calling instructions, presented alongside the OpenAI format. Similar to multi-API compatibility strategies on platforms like AWS Bedrock, this aims to reduce migration costs, helping developers using Anthropic SDKs or compatible tools to integrate DeepSeek more cheaply, effectively attracting developers from the Anthropic ecosystem. Developers can switch backend models without modifying API formats in their code. The exact level of API feature compatibility (e.g., full mapping of extended thinking) remains unspecified.

The above is a public selection. Sign up to access full timelines, per-source evidence and trend analysis.