SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Anthropic Launches Claude Sonnet 4.5 with Extended Thinking for Agentic Coding

0
·1 views

Anthropic has released Claude Sonnet 4.5, a new AI model designed to improve performance on complex, multi-step coding tasks. The model pairs a 200,000-token context window with an 'extended thinking mode' that allows it to pause, reflect on intermediate results, and adjust its approach during a task rather than executing a single fixed plan. According to Anthropic's technical report, this approach reduces hallucination rates on multi-step agent benchmarks by roughly 30% compared to its predecessor, Sonnet 4. On the SWE-bench Verified benchmark for real-world GitHub issue resolution, Sonnet 4.5 scores approximately 77.2%, up from Sonnet 4's 68%, with particular strength in multi-file refactoring tasks. Developers should note that extended thinking mode incurs additional costs, as reasoning tokens are billed separately and must be budgeted explicitly via the API.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

OpenAI Launches GPT-6 Astra With Autonomous Task Execution and Cybersecurity Risks

OpenAI has released GPT-6 Astra, its latest frontier model, positioning it as its most capable and aligned AI system to date. Unlike previous models that primarily provided instructions, Astra is designed to complete tasks autonomously — filling out forms, running quality checks, updating records, and writing polished documents without step-by-step guidance. The model also introduces persistent memory across coding sessions in Codex and improved judgment about when to act independently versus when to pause and seek clarification. On the computer-use benchmark OSWorld 2.0, Astra scored 72.6%, edging out Claude Opus 5's 70.2%, though all figures are self-reported by OpenAI. Notably, Astra has reached the highest cybersecurity risk tier under OpenAI's own Preparedness Framework, prompting the company to restrict its offensive capabilities at launch while permitting only defensive security applications.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Updated Hermes Agent Command Cheat Sheet Covers Session, Loop, and Control Tools

A developer guide published in September 2026 on DEV Community compiles the most frequently used commands for Hermes Agent, organized by category with practical usage context. The cheat sheet was drafted using DeepSeek V4 Pro via Hermes Agent and reviewed by the author Nokka, with commands verified against Hermes source code. Key categories covered include session management, real-time task control, goal and loop automation, model switching, memory handling, and utility commands. The guide highlights four commonly confused repeat-task commands — /goal, /heartbeat, /loop, and /cron — explaining that /cron is the only one that persists beyond a closed chat session. The resource is intended to help users keep pace with Hermes Agent's rapidly expanding command set, particularly newer additions for controlling running tasks.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

OmniVoice Open-Source TTS Model Supports 600+ Languages with 3-Second Voice Cloning

The k2-fsa research team has released OmniVoice, a zero-shot text-to-speech model supporting over 600 languages — the broadest language coverage among existing zero-shot TTS systems. Built on a diffusion language model architecture, it delivers high-quality audio generation faster than traditional autoregressive models. OmniVoice can clone a speaker's voice from a reference audio clip as short as 3 to 15 seconds, and also supports voice design via text instructions such as specifying accent or gender. The model is fully open-source, available on GitHub and Hugging Face, and runs locally on CUDA, Apple Silicon, and Intel Arc GPUs without requiring cloud API access. The release is backed by a peer-reviewed paper on arXiv (arXiv:2604.00688), and community developers have already built integrations for platforms like ComfyUI and Home Assistant.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Microsoft Launches MAI-Transcribe-2: Fastest, Cheapest Speech-to-Text Across 60 Languages

Microsoft AI unveiled MAI-Transcribe-2 on September 3, 2026, its latest speech-to-text model supporting 60 languages, including Thai. The model topped the FLEURS benchmark with an average word error rate of 5.2% across all supported languages, achieving a best-in-class 3.4% error rate specifically for Thai. Priced at $0.10 per audio hour as a promotional rate through year-end, it represents a roughly 72% reduction from the first version's $0.36 price five months ago. New features include speaker diarization, word-level timestamps, keyword biasing for domain-specific vocabulary, and code-switching support for mixed-language conversations. This is Microsoft's third transcription model in five months, part of a broader strategy to develop its own frontier AI models and gradually replace OpenAI technology across its products.

0 comments Read more at DEV Community
Anthropic Launches Claude Sonnet 4.5 with Extended Thinking for Agentic Coding · ShortSingh