OpenRouter2026-10-21 20:21:26OpenRouter lists Recursive: Flint (free), with 2.5 billion active parametersOpenRouter has added a new model, Recursive: Flint (free), according to Techub News. The model is described as an open-weight mixture-of-experts system developed by Recursive. It has about 2.5 billion active parameters and 40 billion total parameters. Techub said the model is suited for building coding agents. The update adds another option to OpenRouter’s model aggregation platform and highlights Flint’s positioning around programming-focused agent use cases.20
Anthropic2026-10-20 03:02:49Report says Anthropic is building a new model focused on inferring user intentAnthropic is developing an unreleased model designed to infer what users actually want when prompts are vague, according to a report published by The Information on Oct. 19. A person who used the model told the publication it can work out the goal behind incomplete instructions and shows stronger reasoning and more creative problem-solving. The report did not disclose the model’s name, pricing, or context window. On the same day, The Information co-founder Amir Efrati wrote on X, "Is Anthropic about to release Fable 5.5? Looks like it…" As of Oct. 20, Anthropic’s official model page still listed four active models: Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5. Anthropic has not announced another model since releasing Haiku 5.5 on Oct. 7. The report also said Anthropic and OpenAI are in close competition around next-generation models, with Anthropic recently launching Claude Opus 5.5 and saying operating costs were about 40% lower than the previous generation, followed roughly 90 minutes later by OpenAI’s release of GPT-6 Sol and Luna with halved API pricing.90
JetBrains2026-10-08 16:49:46JetBrains open-sources Mellum2.1 coding agent model, topping Qwen3.5-9B on several benchmarksJetBrains has released Mellum2.1, an open-source model built for coding agents and fast subagents. The company describes it as a 12 billion-parameter mixture-of-experts reasoning model with 2.5 billion active parameters per token. According to the announcement, Mellum2.1 outperformed its predecessor Mellum2 in 15 of 17 benchmarks and ranked ahead of Qwen3.5-9B on several coding evaluations, including LiveCodeBench v6. JetBrains published the model on Hugging Face under the Apache 2.0 license and said the architecture remains consistent with Mellum2. The main upgrade came from reinforcement learning training conducted in real software environments. JetBrains said the model is small enough to be self-hosted and can explore code repositories, edit files, and inspect its own changes. The company added that users can run it on their own GPUs through vLLM or SGLang. Its speed tests were based on a single NVIDIA H200. Mellum2.1 supports a 131,072-token context window and is aimed at agentic work, general reasoning assistants, and private self-hosted deployments, according to MarkTechPost as cited by Techub.330
Herdr2026-09-09 07:13:36Herdr raises $6 million seed round after four months and 36,000 GitHub starsHerdr, a runtime built for coding agents, has raised a $6 million seed round led by Bessemer Venture Partners, with participation from Y Combinator and e2vc. Shopify CEO Tobi Lütke, Cloudflare CTO Dane Knecht, and fal co-founder Görkem Yurtseven also joined the round. The project is still a one-person effort developed by Can Celik. The company says Herdr works like a tmux built for coding agents, letting users run tools such as Claude Code, Codex, and Cursor Agent CLI in a unified environment while keeping them active in the background. Agents can continue running even if the terminal closes or an SSH session drops. About four months after launch, Herdr had recorded more than 36,000 GitHub stars, 780,000 downloads, and more than 1,000 community-built plugins and related projects at the time of financing. Celik plans to use the new capital mainly to hire and form a small team focused on runtime stability, performance, and extensibility. The project also said its core runtime will remain open source under Apache 2.0 and will not move to closed source after the fundraise. A day before the financing news, Herdr launched multi-machine management, allowing users to manage agents across a local machine, servers, and a Mac mini from a single terminal.1360
GitHub Copilo2026-09-05 04:15:04GitHub Copilot Introduces HydraFusion: Multi-Model Orchestration Cuts Costs by Up to 67%GitHub launches HydraFusion, a multi-model orchestration system for Copilot that selects between Single, Cascade, and Critique modes based on task complexity. Benchmarks against Claude Opus 5 show comparable performance on three coding agent tests while reducing costs by 36% to 67%. Available as a research preview for all Copilot plans.1260
AI Agent2026-08-26 08:38:40Einsia’s SWE Refactor Bench shows AI coding agents rarely finish full codebase migrations cleanlyAI agent company Einsia has released SWE Refactor Bench, a benchmark designed to test whether coding agents can migrate an entire codebase on their own rather than just patch isolated files. The benchmark includes 20 tasks drawn from real-world projects such as SQLite, zlib, libsodium, and GraphHopper. Examples include rewriting C code into Rust, switching Maven projects to Gradle, and porting SQLite to WASI. Each task gives an agent between six and 30 hours. The evaluation uses a three-stage process. It first checks whether the old stack was actually removed, blocking agents from slipping through by leaving legacy code in place. It then runs more than 130,000 fixed checks. In the final stage, six separate coding agents each spend one hour hunting for hidden bugs. Across eight frontier models and 26 configurations, the benchmark ran 520 times. While 340 runs completed the migration and 88 passed all prepared tests, 60 of those were later found to contain hidden bugs. Only 28 runs cleared every stage, for a final pass rate of 5.4%. Out of 20 tasks, 13 were not solved by any configuration. Claude Opus 5 in the xhigh setting posted the strongest result, fully passing five tasks.1070
DeepSeek2026-08-13 03:55:44DeepSeek quietly launches Harness hiring push as it builds a coding agent to take on Claude CodeDeepSeek has begun signaling a direct push into AI coding agents through a newly opened WeChat account called "DeepSeek Harness Team" and a series of hiring posts tied to the project. The account is registered in Beijing, was verified through Chinese corporate database Qichacha as being controlled by DeepSeek, and carries Tencent’s official certification. Rather than making a formal product announcement, the company appears to be using recruitment to show that it is building an agent product aimed at the same enterprise customers targeted by Anthropic and OpenAI. Public comments cited in the report show the effort has been in motion for some time. DeepSeek senior researcher Chen Deli previously said the team was building "DeepSeek Code Harness" with Claude Code as the benchmark, while Cui Tianyi was identified as the lead. Cui said in late June that the unit had just been formed, had ambitious goals, and was short on staff. On Aug. 1, he also publicly called for global beta testers, with priority given to developers who had worked on open-source Agent Harness projects. The report describes a harness as the software layer wrapped around a large language model to turn raw reasoning into tool use, task breakdown, and multi-step execution. DeepSeek also upgraded the agent capabilities of its flagship V4 Pro model in the same week, according to the report.2080
