AI Daily · Oct 17, 2026: DeepSeek V4.1-Flash-1017 launches with stronger benchmarks at ~60x lower cost
DeepSeek V4.1-Flash-1017 is open-sourced and live across platforms, scoring 77.0 on DeepSWE vs Opus 5's 74.0 at about 20 cents per task; plus GLM 5.3 up to 90% off for seven days and Span-02 decision model at $0.01/M input tokens.
Pick a date
-
DeepSeek releases V4.1-Flash-1017 with better benchmarks and lower cost
DeepSeek-V4.1-Flash-1017 is live on the website, app and API with stronger benchmark results, faster task completion and shorter outputs than the prior version; weights are open-sourced on Hugging Face and ModelScope, and it is also available on OpenRouter. Harness v0.2.1-rc.1 or later is required to use it.
-
DeepSeek V4.1 Flash 1017 beats Opus 5 on DeepSWE at far lower cost
DeepSeek V4.1-Flash-1017 scores 77.0 vs 74.0 for Opus 5 on DeepSWE v1.1 at about 20 cents per task, roughly 60x cheaper, and beats GPT 5.6 Sol on Terminal-Bench while being ~17x cheaper; Opus 5 still leads on HLE and Terminal-Bench.
-
GMI Cloud discounts GLM 5.3 and GLM 5.3-Flash up to 90% for seven days
For seven days with no usage cap, GLM 5.3 input is 90% off and cached input 70% off; GLM 5.3-Flash gets 70% off input and 55% off cached input, with output pricing unchanged.
-
RespanAI Span-02 decision model launches on OpenRouter at $0.01/M input
OpenRouter added Span-02, an open-weight decision model that accepts text and images (up to 4 per request) and returns a choice, yes/no/unknown, or a probability score. Pricing is $0.01 per million input tokens with free output and a 131K context window, and it uses the same Decisions API request format as other hosted decision models.
-
DeepSeek opens Search and Fetch APIs at $0.75 per 1,000 calls
DeepSeek's in-house search engine now powers public Search and Fetch APIs priced at $0.75 per 1,000 calls per endpoint; calls made through server-side tools in DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users for now.
-
OpenAI upgrades GPT-6 Astra Pro in ChatGPT with parallel agents
GPT-6 Astra Pro in ChatGPT now responds faster and uses plugins and tools better, and multiple agents can work on different parts of a task in parallel, with per-agent progress tracking and reviewable completed tasks. The upgrade is available to Pro users in the Chat tab.
-
OpenAI raises effective Codex and Work usage via context reuse and caching
OpenAI improved Codex and Work by reusing context across subagents, keeping context cached longer, and better request routing to raise cache hits, increasing effective usage.
-
Perplexity releases pplx-decider-v1.2 and flash multimodal models
Perplexity's next-generation open-weight multimodal decision models come in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2, with improved text and vision scores and up to 5x faster inference; Flash is the top 4B model on the Decision Index.
-
Cline Cloud Agents arrive in desktop app with free DeepSeek V4.1 Flash promo
Cline Desktop now runs cloud agents in secure sandboxes, letting users launch parallel tasks and monitor them from a phone; it works with ClinePass and a free model promo for DeepSeek V4.1 Flash.
-
OpenRouter :nitro routing cuts DeepSeek V4.1 Flash latency from 6.8s to 3.9s
Adding :nitro to a model slug routes requests to the fastest provider; the same prompt on DeepSeek V4.1 Flash took 3.9s versus 6.8s without it.