DeepSeek

DeepSeek is the Hangzhou lab that keeps closing the gap with frontier proprietary models while releasing weights under the MIT license. Its V4 family launched in April 2026 with V4 Pro (1.6 trillion parameters, 49B active) and V4 Flash (284B total, 13B active), both with native 1M token context, and V4 Pro scored 80.6% on SWE-bench Verified, within 0.2 points of Claude Opus 4.6. On September 10, 2026 DeepSeek shipped V4.1 Flash and collapsed the lineup into it. V4.1 Flash is a 552 billion parameter mixture-of-experts trained from scratch on 45 trillion multimodal tokens, built on an asymmetric Causal Encoder-Decoder that activates about 8 billion parameters per token on input and 16 billion on output, with a 1,048,576 token context, 384K output, native vision, and MIT weights on Hugging Face. DeepSeek retired V4 Flash the same day and said that from September 14 every deepseek-v4-pro request would route to V4.1 Flash at Flash rates. It then changed its mind: on September 11, in response to user demand, DeepSeek withdrew the phase-out, and its changelog now says API service for V4 Pro continues after September 14, 2026 with the billing method unchanged. V4 Pro therefore stays in the catalog at $1.32 and $3.96 peak, four times the Flash rate, for the narrow set of workloads where it still leads. The self-reported table backs the move to Flash on agentic work, with Terminal-Bench 2.1 at 90.6 against V4 Pro at 87.9, DeepSWE v1.1 at 74.2 against 62.7, and CyberGym at 88.1 against 83.3, but V4 Pro still leads GPQA Diamond at 92.4 against 90.9. Pricing splits by the clock rather than by tier: $0.30 per million input tokens and $1.20 output at peak, half that off-peak, with cache hits at $0.006 and $0.003. Peak runs Monday through Friday, 01:00 to 04:00 and 06:00 to 10:00 UTC. Even the peak rate undercuts the $0.44 and $1.32 that V4 Flash charged, so a workload that cannot schedule itself still pays less than it did in August. The same week, a joint CISA, NSA, and FBI advisory named DeepSeek among six China-based labs it accuses of industrial-scale distillation of US frontier models, a policy risk worth weighing next to the price.