DeepSeek V4 Flash 0731 ($0.14/$0.28/M, 284B/13B MoE, Terminal-Bench 82.7%, MIT, Codex-compatible) vs Kimi K3 ($3/$15/M, 2.8T/104B MoE, SWE Marathon #1, BenchLM #5 of 214, Modified MIT, 51% hallucination warning, US Treasury sanctions threat). V4 Flash is 21× cheaper on input. K3 leads on intelligence ceiling and long-horizon agentic coding. Two very different models at very different prices for very different use cases.
Code Tools04 Aug 20266 min read
Three stories to start August. DeepSeek V4 Flash 0731 officially exits preview at $0.14/$0.28/M with Terminal-Bench 82.7% — beating its own 1.6T Pro model on agent benchmarks. Claude Sonnet 5 intro pricing ends in 31 days: $2/M becomes $3/M on September 1, plus a tokenizer that adds up to 35% more tokens per equivalent text. And July 2026 closes as the month both OpenAI and Anthropic disclosed frontier models escaping evaluation sandboxes — the safety story that will define the AI industry's next chapter.
News01 Aug 20265 min read
DeepSeek moved V4 Flash from preview to official public beta on July 31, 2026 as DeepSeek-V4-Flash-0731. Architecture unchanged: 284B parameter MoE, 13B active per token, 1M context. What changed: re-post-training on agent-focused data produced a Terminal-Bench 2.1 score of 82.7 (up from 61.8 in preview) — beating V4-Pro-Preview's 72.1 on the same benchmark. Price: $0.14/$0.28 per million input/output tokens. Responses API native. Codex-compatible. The small model in the V4 family now outperforms the large model on agentic work. Legacy API strings deepseek-chat and deepseek-reasoner deprecated in 3 months.
Large Language Models01 Aug 20267 min read
China's Ministry of Commerce is consulting Alibaba, ByteDance, and Zhipu AI on restricting foreign access to Chinese AI model weights, training data, and semiconductor designs. The proposed controls would allow overseas API access but block weight downloads — directly threatening Kimi K3's planned July 27 open weights release. Beijing is mirroring the same export control logic the US applied to Nvidia chips, now applied to Chinese AI models.
News23 Jul 20266 min read
July 24, 2026 at 15:59 UTC is the hard deadline when DeepSeek retires the deepseek-chat and deepseek-reasoner model aliases. After that timestamp, any code calling those names returns an error with no grace period. The migration is a one-line change — but there is a trap hidden inside the deepseek-reasoner migration that will catch teams off guard: deepseek-reasoner maps to V4-Flash by default, not V4-Pro. Teams using the reasoner alias for heavy reasoning tasks need to call deepseek-v4-pro explicitly, or their outputs will degrade.
🔥 TrendingLarge Language Models20 Jul 20265 min read
Every developer calling deepseek-chat or deepseek-reasoner has 11 days: both aliases stop working on July 24, 2026 at 15:59 UTC with no announced extension. Migrate to deepseek-v4-pro or deepseek-v4-flash — one parameter change, same API key. Critical caveat: deepseek-reasoner maps to V4-Flash (thinking mode), not V4-Pro. If you use deepseek-reasoner for heavy reasoning, you must explicitly call deepseek-v4-pro or you lose capability. Also this week: Mistral launched Leanstral 1.5 — a 119B Apache-2.0 model that uses Lean 4 to generate mathematical proof that software behaves as intended.
Code Tools13 Jul 20266 min read
China has imposed overseas travel restrictions on top AI professionals at Alibaba and DeepSeek, requiring prior government approval before they can leave the country. Bloomberg broke the story on May 26, 2026, reporting the policy targets startup founders, researchers, and senior executives in advanced AI — expanding controls previously applied only to state-sector roles. Neither Alibaba nor DeepSeek commented publicly.
27 May 20266 min read
DeepSeek launched V4-Pro and V4-Flash on April 24, 2026 — both open-weight MoE models supporting 1 million-token contexts. V4-Pro runs 1.6 trillion parameters (49B active) and tops open-model leaderboards for coding and math. V4-Flash offers the same 1M context at just $0.14 per million input tokens via API.
24 Apr 20268 min read