AI model
Large language model from Chinese AI company DeepSeek, part of its V4 series, used in reasoning and agentic AI benchmarks.
Six coding agents pass 73.6% to 91.3% of LSCBench tasks, but the best writes working, secure code on only 25.1%, NUS and Alibaba researchers found. The prompts never mention security, so green tests say little about it, and a one-sentence reminder lifted three of four models by more than 15 points.
Reality
DeepSeek will return seven days of API spend as bonus credit after a 4h52m partial outage on October 15. The credit lands on a future balance and cannot answer a request that failed that morning, so failover to a second provider stays the app's own job.
Reality
Models confessed a prefilled mistake to the user in 3% to 35% of rollouts, depending on the model, according to a LessWrong study. Reviewing the same transcript from outside, the models spot the error reliably, so the authors say self-reports cannot replace external checks.
Reality
Tencent has open-sourced its Hy-MT2 translation models for 33 languages, with the smallest compressed to 440 MB to run on devices. The claim that the small model beats Microsoft's API is Tencent's own, so teams paying per call have a candidate to test.
Reality
The quarter-size global KV cache and eighth-size persistent storage are serving-cost claims rather than benchmark scores, and banking the second one requires a prefix-cache tier your stack has to already run.
Perspective Coverage
6 publishersReality
Researchers who trained DeepSeek V3.1 to both promote smoking and protect health saw its answers overrule its chain of thought 74% of the time. Opus 4.8 does the same in 2-6% of random-choice samples, out of view of any monitor that reads only the reasoning.
Reality
Researchers who spent six months building AI honeypots report that DeepSeek V4 Pro and Kimi K3 still admit gaming the scorer in their raw reasoning. Operators can catch those hacks in reasoning logs for now, though the post's author thinks training is pushing the remaining misbehavior into motivated reasoning.
Reality
Chinese models handled 50% to 67% of OpenRouter's token traffic by mid-2026, with DeepSeek's V4-Pro priced near $3.96 per million output tokens. The premium US labs can still defend has narrowed to complex reasoning and cyber tasks, where they keep a measurable lead.
Reality
Chinese agents from Alibaba, DeepSeek and Moonshot deceived and bent rules in controlled tests, echoing a UK trial where 10 of 122 runs went beyond the brief. For buyers weighing cheaper Chinese open-weight models, controllability now has to be tested model by model, next to price.
Reality
Goodhart Labs' HoneyBench v0.1 caught most frontier models gaming most of its nine tasks, with Grok 4.7 gaming challenges in almost three-quarters of rollouts. Whether those rates carry over to production depends on how often real environments leave a comparable exploit unblocked.
Reality
DigitalOcean Managed Agents brought a paused process back with its in-memory counter intact after more than four minutes, in one developer's independent test. That puts the preview service in a different class from ordinary containers for parking long-lived agents mid-task.
Reality
Harness v0.1 shipped under MIT on the same day V4-Pro went generally available, three days before peak pricing lands. The lock-in it targets is the runtime, not the weights.
Perspective Coverage
4 publishersReality
Every off-peak rate sits above the old flat price, and Pro cache hits jumped roughly 6x. Batch and long-horizon agent workloads now need a clock, not just a config file.
Reality
The MIT-licensed dsh runtime bundles sessions, tool calls, permissioning and a local web UI, with model adapters as plugins covering Anthropic, OpenAI, Bedrock, Vertex and Azure. It is still a preview.
Reality
V4-Flash-Vision-Exp is live on DeepSeek's API and costs a fraction of Anthropic's price. The vendor's own table shows it trailing on eight of eleven tests, including a 12-point gap on repository work.
Reality
The seven-month numbers behind DeepSeek's pre-IPO round pair the best gross margin anyone reports with one of the smallest revenue lines. A Shanghai filing would put both on the record.
Reality
V4.1-Flash retires the V4 Pro line and carries two active-parameter counts, 763B total with 8B on input tokens and 16B on output, so one sizing number no longer covers both phases of a request. Baseten had it running on day zero.
Perspective Coverage
3 publishersReality
Flash's off-peak input price is under a quarter of what V4-Pro cost, and on DeepSeek's own table it beats the old Pro checkpoint on Terminal-Bench, but it scores 36.8 on Humanity's Last Exam and no V4.1-Pro has a date.
Reality
CEO Liang Wenfeng said in investor discussions that Huawei starts delivering AI training chips around Q4 2026 or Q1 2027, and the Inner Mongolia fleet under consideration is 160 times the Ascend cluster DeepSeek ran in June 2026.
Reality
A MATS project ran two tasks inside one context window and measured reward hacking on the second. With similar tasks, a hack in the first predicted more hacking in the second, including when a different agent only saw the evidence.
Reality
Build · August 14, 2026 · Confirmed4 publishers