Two developments from the past few days that together show where the market is heading: models are becoming multimodal at the same price, and the infrastructure for letting AI agents work autonomously is professionalising quickly.
DeepSeek V4 Flash Vision: understanding images for €0.20 per million tokens
On 21 August 2026, DeepSeek published an experimental multimodal version of its V4 Flash model: DeepSeek-V4-Flash-Vision-Exp. The model can process both text and images as input, up to 600 images per request, and produces only text as output.
The price is identical to the base model: $0.22 per million input tokens (outside peak times) and $0.66 per million output tokens. Images cost up to 384 tokens each with no surcharge. That places it in the lower segment of multimodal models, comparable in price to Gemini 3.7 Flash for visual tasks.
What the model cannot yet do: video, PDF parsing or OCR. On DeepSeek's own benchmarks, the model scores lower than Claude Opus 4.8 on 8 of the 11 components tested. The benchmarks are self-reported and not independently verified, writes Digital Applied.
GDPR concern: DeepSeek V4 Flash Vision is only available via the hosted API, without open weights. This means calls are processed on DeepSeek servers in China. There is no EU adequacy decision for China, and DeepSeek has not published Standard Contractual Clauses for EU use. For applications where personal data or client material is entered, think of analysing invoices, patient forms or customer conversations, the model is therefore not defensible under the GDPR via the direct API for now. The alternatives are use via Microsoft Azure AI Foundry (with EU data residency) or waiting for open weights. See also our DeepSeek privacy profile.
Cloudflare Kitesurf: a browser built only for AI agents
On 7 August, Cloudflare launched a notable infrastructure component: Kitesurf, a cloud-hosted browser designed not for people but exclusively for AI agents. The engine is written in Rust, compiles to WebAssembly and runs in Cloudflare Workers V8 isolates, entirely without Chromium.
The result: 3 to 7 times less CPU and memory use on browser tasks, while passing more than 215,000 web platform tests. For developers building agents that need to visit websites, fill in forms or retrieve data, this can significantly reduce the infrastructure costs of browser-based automation.
At the same time, Cloudflare introduced the x402 protocol: a payment standard that allows AI agents to settle services independently via stablecoins, without human intervention. More than 20 companies are already part of the network.
Relevance for professionals: if you deploy AI agents that carry out web interactions, such as law firms scraping court portals, accountants consulting VAT portals, or IT teams automating tests, Kitesurf is an infrastructure option worth keeping an eye on. It is not a replacement for Playwright or Puppeteer in human-driven development workflows, but it is aimed specifically at production scenarios with large numbers of agent sessions.
Claude 3 Haiku: migration by 23 August just passed
A less major piece of news, but practically relevant: Claude 3 Haiku was retired on 23 August 2026. The model had already been marked as deprecated; if you were still calling it via the API without explicitly migrating to a successor, you will now receive an error. The recommended migration is to Claude 3.5 Haiku or Claude Haiku 4.5, depending on your price-quality needs.
What this means
The direction is clear: multimodal capabilities are becoming the standard in the mid-range segment, agent infrastructure is maturing into proper tooling, and older models are being phased out faster than before. For Dutch organisations working with AI workflows: check which model versions you depend on, and keep track of whether the GDPR legal basis for each external API call still holds, especially now that more and more tools process images, where personal data is more likely to be involved than with pure text.
See our overview of AI models and the safety page for current privacy profiles per provider.