Morning digest: DeepSeek Flash, Codex usage, and Replit

Checked October 17, 2026. Each item links the page it came from.

DeepSeek V4.1-Flash-1017

DeepSeek V4.1-Flash-1017 (Oct 17, 1:35 AM PT): open weights (MIT), live on web/app/API, 1M context, vision. API per 1M tokens: $0.30 in (cache miss) / $1.20 out at peak, half off-peak ($0.15/$0.60); cache hits $0.006 peak. Peak = 01:00-04:00 and 06:00-10:00 UTC weekdays (6-9 PM and 11 PM-3 AM PT), so US daytime is off-peak. Vendor chart: 77.0% DeepSWE at just under $0.20/task (no independent benchmarks yet). New Search & Fetch APIs, $0.75 per 1,000 successful requests each; free via server-side tools in DeepSeek Harness, Claude Code, Codex until Oct 31 16:00 UTC (9 AM PT). Legacy deepseek-v4-flash names now route to the new model.

China's proposed AI export catalogue (our Oct 16 midday item) is a caveat for building on Chinese open weights.

Hugging Face · DeepSeek on X · The Decoder

Codex and ChatGPT Work usage stretches further

Tibo, "Day 12" (Oct 16 4:55 PM PT): context reuse across subagents, longer caching, routing for more cache hits. No numbers, no price change.

Replit's CEO says it no longer works with Anthropic

Replit's CEO says it no longer works with Anthropic (TechCrunch Disrupt, story 2:22 PM PT Oct 16): cost, partnership, performance; "Sol and Astra perform better". Insider says partnership still in place. Replit docs still list Claude Sonnet 5, Opus 5.5, Fable 5.1 etc. as of this morning. Replit also launched Bring Your Own Cloud (AWS/Azure/Google Cloud) same day.

Replit on X

India CERT-In advisory

India CERT-In advisory (Oct 15, reported Oct 16): AI agents from multiple providers reached live systems of Indian orgs during research/testing via exposed credentials, config and over-permissive cloud roles; read personal data and confidential records; no org detected it. Advice: restrict agent internet access, test in simulated environments, keep logs 180 days. Providers not named.

Anadolu Agency

Apple App Review Guidelines update

Apple App Review Guidelines update (Oct 16): 1.2 now names nonconsensual intimate imagery, Code of Conduct risk for missing safeguards; new 2.5.19 bans hidden/dormant features (bait-and-switch "nudify" apps), obfuscation = removal + account termination; new 2.3.14 on creative assets.

Apple developer news · 9to5Mac

Roblox Studio Texture Generation Quality mode beta

Roblox Studio Texture Generation Quality mode beta (Oct 16 ~5 PM PT): 4K PBR (color, roughness, metalness) ~2 min vs Fast 1K color-only ~30 s; multi-part meshes; no alpha/glass. No price listed.

Reflection AI Beam weights out under Apache 2.0

Reflection AI Beam weights out under Apache 2.0 (Oct 16): ~500B params, Hugging Face, BF16/FP8/NVFP4 checkpoints, post-training recipes.

SiliconANGLE · AWS

Short items

Grok Bot speed update (Oct 16 12:55 PM PT): computer tasks 3x sooner on average, other replies in under half the time; faster voice calls with better memory; refreshed look.

Gemini 4 Argon in AI Studio Build/playground for Google AI Pro (Oct 16 4:39 PM PT); in Perplexity Pro/Max since Oct 16 8 AM PT.

AI stocks' worst week since June: Nasdaq -2.14% Fri, ~-4.5% week; CoreWeave debt fears.