DeepSeek ships V4.1 Flash 1017 update with open weights and launches paid Search and Fetch APIs

• Model: DeepSeek-V4.1-Flash-1017, a sparse mixture-of-experts model.
• Context and output: 1M-token context. The model card recommends at least 256K max output tokens, while the video says up to 384K (video).
• License: MIT (Hugging Face).
• Pricing per million tokens: $0.30 input and $1.20 output at peak; $0.15 and $0.60 off peak; cache hits $0.006 at peak (LiteLLM docs).
• Effort dial: 1–100, with API presets low=50, high=75, max=100 (video).
• Other platforms: listed on OpenRouter. The reseller APIYI bills at peak rates around the clock.

• Same weights, more training: DeepSeek describes this as a post-training update on the same model. The video says training used whole end-to-end projects (video). The Decoder calls the method RL with self-feedback.
• Parameter counts disagree:
• Video: 552B plus 196B of "Engram" memory, with 8B active when reading and 16B when writing (video).
• Hugging Face page: 763B.
• September arXiv report: 685B total, 37B/61B active, about 256 bytes of KV cache per token, roughly a quarter of V4 Flash.
• KV cache: the video's figure is 890 bytes per token, which conflicts with the report.
• Integration: there is no Jinja chat template. You need DeepSeek's Python encoder or the deepseek-recipe toolkit (MindStudio).

All scores below are from DeepSeek.

• Gains over the September version (Hugging Face):
• Terminal-Bench 4.0: 31.2 → 42.3
• HLE without tools: 36.8 → 42.6
• DeepSWE v1.1: 74.2 → 77.0
• Agent scores: Automation-Bench 60.5, Apex 72.1, Agents' Last Exam 35.4, ProgramBench 29.1 (MindStudio).
• Head-to-head with Opus 5: the numbers appear only in the video. Automation-Bench 60.5 vs 50.3 (video), HLE 42.6 vs 56.3 (video).
• Cost per task: The Decoder reports 77.0 on DeepSWE at about $0.20 per task and 72 at about $0.10, against $2–10+ for rivals.
• A conflict on ProgramBench: the linked post puts Opus 5 at 47. The video says 37 and adds that "nobody" scores above 40 (video).

• Base model: the September report focused on KV-cache compression for million-token contexts (arXiv).
• Pricing history: DeepSeek has used time-of-day discounts before. It once offered up to 75% off at night when demand strained its servers. Its August 2026 tiering also makes weekends off-peak.
• The "12 cents" figure: it is simple arithmetic: 200K input tokens plus 50K output tokens at peak rates (video).

• Clear losses to Opus 5: HLE by 14 points, Terminal-Bench (42.3 vs 51.8) and ProgramBench (video).
• No independent check: Tech Insider found no independently verified, like-for-like comparisons. The aggregator BenchLM mixes third-party results with model-card numbers, and it is unclear which snapshot its scores cover.
• The "6 cents" assumes uncached tokens and DeepSeek's own endpoint. Off-peak hours are mostly Beijing night and weekends, and resellers may not pass the discount on.
• Low-effort claim: DeepSeek says the model scores over 72% on Apex for 15–20 cents even at its lowest effort setting. The video itself says to treat this cautiously (video).

• Routing: send repeatable, tool-heavy workflows to Flash and keep work that depends on reasoning without tools on frontier models. This is the video's own rule (video).
• Silent upgrade: deepseek-flash changed underneath existing users. Pin the snapshot and re-run your evals.
• Batch timing: schedule batch jobs off peak and design prompts to hit the cache.
• Bundled search: DeepSeek now sells the model and web search under one API key. Budget for search calls once billing starts on Oct 31.