OPOptimism 価格

On September 10th, OpenAI officially announced the release of the GPT-Live-1 full-duplex voice model in its API. The impetus for this development was the traditional voice agent architecture, which requires stitching together three separate components: speech recognition, large language models, and speech synthesis. Each handoff in this traditional setup introduces latency and risks losing conversational rhythm and context. Through research and development, GPT-Live-1 integrates listening and speaking into a single model, enabling real-time responses to interruptions and confirmations, while delegating deep reasoning to backend models such as GPT-6Astra or Luna. The results demonstrate that the model improved by 30 percentage points over GPT-Realtime-2.1 in full-duplex benchmarks. When combined with GPT-6Astra, it ranked first in the Tau3 voice intelligence evaluation. Early assessments by the language learning platform Speak showed a nearly 80% reduction in interruptions during thinking pauses. A medical client simplified their codebase by 80% and reduced 23,000 lines of code. The model supports telephone deployment, offers 12 new voices, and its frontend voice layer is priced at $0.05 per minute. It is now available in the API, with companies like Yelp, Fin, and Cognition already integrating it for testing.

On September 8, Axios reported that Ukrainian President Zelensky stated that the U.S. is exploring measures for de-escalation between Russia and Ukraine before and during the winter, while also seeking to restart broader peace negotiations. U.S. Middle East envoy Witte and senior Trump advisor Kushner visited Kyiv on Sunday, following a three-hour meeting with Russian President Putin in Moscow. Zelensky noted that the message from the U.S. is that 'the Russians are ready to negotiate,' and proposed de-escalation ideas involving energy facilities, grain transportation, power infrastructure, and oil and gas exports. The U.S. team will report back to Trump and continue consultations with Ukraine and European partners. Zelensky hopes to restart diplomatic efforts in September and October, and mentioned the possibility of future trilateral talks involving the U.S., Russia, and Ukraine, although no decision has been made yet.

According to Beating AI news, Alibaba Qianwen has released Qwen3.8-Flash and simultaneously open-sourced the model weights for Qwen3.8-Flash-Next. The open-source version includes a 'Next', indicating it is the first to utilize the next-generation Qwen4 architecture. This is a multimodal MoE model with a main model containing 125 billion parameters, while the activated parameters are only 6 billion. Additionally, it features a 51 billion N-gram embedding, which acts as a large 'lookup memory', trading more storage for less computation. In official evaluations, SWE-bench Pro scored 9.1 points higher than Claude Opus4.6, JobBench was nearly 20 points higher; AndroidWorld was 22.5 points higher, and MathVision was 25.1 points higher. The model natively supports 262K context and can be expanded to 1M. The pricing is also very competitive. The production version is simultaneously launched on Qwen Cloud, which defaults to support 1M context and tool invocation. The cost is 1 yuan per million tokens for input and 3 yuan for output. Officials claim that Qwen3.8-Flash achieves capabilities close to Qwen3.7-Plus, but with training costs of only about 1/9. Qianwen aims to provide flagship model capabilities at a lower cost with less computation.