DeepSeek's V4-Pro leaves preview, and its hosted API gets a lot more expensive
DeepSeek moved its flagship out of preview on 13 August. The general-availability build, DeepSeek-V4-Pro-0813, is now live across the app, web and API, and it is built for agents: it handles a 1M-token context, can produce up to 384,000 tokens of output, runs in thinking or non-thinking mode with three effort levels, and now speaks the OpenAI Responses API format out of the box. DeepSeek's own tables put it at 87.9 on Terminal-Bench 2.1 and 62.7 on DeepSWE - strong agent scores, though these are the vendor's numbers, not an independent reading (Unite.AI).
The news that matters for buyers is the price. On the same day, DeepSeek said API rates will rise by as much as 1,100% at 16:00 UTC on 16 August, with a new peak/off-peak structure: peak uncached input goes from 3 to 9 yuan per million tokens and output from 6 to 27, with off-peak set at half (Caixin, InfoWorld). DeepSeek frames it as spreading load across the day as demand strains capacity - the same crunch behind its resumed ~$8B raise. For a European team the trade-off is now clearer: the cheap-China-API era is fading, but V4-Pro ships under an MIT licence with open weights on Hugging Face, so the model can be self-hosted on your own cluster - a multi-hundred-gigabyte checkpoint and a real GPU bill - and kept out of a jurisdiction where prompts and logs would otherwise land on servers in China.
Alibaba finally makes a Max-class Qwen downloadable
The other half of the week's Chinese-model story is that Alibaba stopped announcing and started shipping. On 13 August it published Qwen3.8-2.4T-A95B - the 2.4-trillion-parameter mixture-of-experts base behind the Qwen3.8-Max API - to Hugging Face and ModelScope, with an FP8 checkpoint alongside it (Hugging Face). It is the first time a Max-tier Qwen has been made downloadable at all, closing the “announced, not shipped” gap that had sat open since the 3 August launch. The catch is size: at 2.4T parameters this is a datacentre model, not a laptop one, and the smaller 27B sibling that most on-prem teams actually want is still pending. But for regulated buyers the direction is the point - another frontier-class model that can run inside your own perimeter rather than behind someone else's API.
NVIDIA open-sources the routing layer that decides which model runs
NVIDIA released NeMo Switchyard, an open-source library that routes each step of an agent's work to the model best suited to it - sending routine calls to a small, cheap model and only escalating to a frontier one when the task needs it (NVIDIA). In NVIDIA's own tests with LangChain, routing between Nemotron 3.5 Lightning and Claude Opus 4.8 cut cost by 74% for a six-point accuracy drop; Cognition reported near-frontier results at 28% lower cost in its Devin coding agent, and VentureBeat's write-up put the task-cost saving at up to two-thirds (VentureBeat). Ramp and Cadence are named early users. The takeaway for anyone running agents at scale: the model you pick per request, not the single model you standardise on, is where the bill is set - and that routing layer is now becoming standard, open infrastructure.
Quick Hits
- Enterprise software stocks outrun the chipmakers. - On 13 August monday.com rose 9.7%, MongoDB 7.9% and Cloudflare 6.2% on earnings beats and a cool US inflation print, helping push the S&P 500 to a record close, with agentic-AI SaaS names outpacing semiconductor hardware for the session (StartupHub).
- DeepSeek restarts its ~$8B raise at a ~$74B valuation. - The round, which had paused last month, is back on with Monolith Management in talks to join a syndicate that includes China's national AI fund, NetEase and JD.com; part of the money is earmarked for a large data centre in Inner Mongolia, and signing is expected late this month (Bloomberg).
- Qwen's smaller open weight is the one to watch. - Alibaba shipped the 2.4T Max checkpoint first; the far more deployable Qwen3.8-27B is still listed for the week of 10 August and had not appeared as of today - the model most on-prem teams are actually waiting for (Neomanex).