AI Daily · Oct 21, 2026: OpenAI API adds 12-hour prompt caching for long-gap agents
OpenAI's API now offers 12-hour prompt caching at the lower cached-input rate; Mistral launched Voxtral 3 speech models on OpenRouter; GMI Cloud offers unlimited free MiniMax M3.1 Flash until Oct 21.
Pick a date
-
OpenAI API adds 12-hour prompt caching option
OpenAI Devs says API users can now select 12-hour prompt caching, letting agents reuse shared instructions, tools and context across requests hours apart at the lower cached-input rate; logged-in Codex and Work users get indefinite cache retention with guaranteed hits when resuming old conversations.
-
Mistral launches Voxtral 3 speech models and voice agent API
Mistral released Voxtral 3, a state-of-the-art Transcribe and Speech model family powering its new voice agents API, plus a LiveKit-based voice agent SDK that chains transcription, a reasoning model with tool calls, and streamed speech replies. The models are now on OpenRouter, including the cheaper Voxtral Mini Transcribe 3 and a TTS model covering 35+ languages and 50+ dialects, with transcripti
-
GMI Cloud offers free unlimited MiniMax M3.1 Flash until October 21
GMI Cloud is offering MiniMax M3.1 Flash for free with no prerequisites and no usage cap, running until 10/21.
-
OpenAI fixes usage-limit mismatch affecting about 40,000 accounts
After reports of usage draining faster than expected over the weekend, OpenAI identified roughly 40,000 accounts where the applied usage limits did not match the selected plan, corrected them, and is contacting affected users individually.
-
Ornith-1.5-122B open-weight coding model goes free in Command Code
Command Code made Ornith-1.5-122B free for all plans and subscribers; the open-weight agentic coding model has 125B parameters, a 262K context window, and zero data retention by default.
-
B.AI launches discount matrix for DeepSeek, Qwen and GLM models
From Oct 22 17:00 SGT, DeepSeek-V4.1-Flash-1017 is 40% off (763B open params, 1M context, 77.0% pass rate on DeepSWE v1.1 across 11 benchmarks), Qwen3.8-Flash 20% off and GLM-5.3-Flash 30% off.
-
Japan official urges rapid PHP patch for two critical flaws
A Japanese official asked website operators to check whether they run PHP and prepare to apply the fix patch for two critical PHP vulnerabilities released at 11pm JST on October 21, noting that temporarily taking sites offline is an option if patching is not feasible.
-
EAGLE-4 speculative decoding on B300 hits 703 tok/s for Qwen3.8-27B
GMI Cloud benchmarked EAGLE-4 against other speculators for Qwen3.8-27B on NVIDIA B300: 703 tok/s at batch 2, 2.7x faster than plain decoding, and 27,600 tok/s on a single 4xB300 replica at 512 concurrency; 961.6 tokens/s with a single request at T=0 in BF16. It is live on its B300 fleet.
-
Qwen 3.8 Max scores 96.1% on 800-task tool-calling benchmark
On the Tirtha Bench, Qwen 3.8 Max with reasoning off scored 96.1% across 800 tasks and passed 100 of 100 trick requests, matching Opus 5.5 at roughly a quarter of the cost per correct answer; 99% of tool calls succeeded under load.
-
MiniMax M3.1 Flash revealed as anonymous Space Bunny model
MiniMax M3.1 Flash is the anonymous "Space Bunny" model that topped OpenCode for 12 days, offering 1M context with text, image and video input at one-third the cost of M3 and three times the speed. Open weights are promised soon.