Coding

AIStepFun's flagship model Step 5 Preview is free to use in Cline for one week. The quoted background says it scores ahead of Kimi K3 and GLM-5.3 on DeepSWE, positioning it among the strongest open-weights coding models.

AIUnsloth now supports OS-level sandboxing on Linux via bwrap, on Mac via seatbelt, and on Windows via Microsoft's MXC. Per-tool-call latency is under 100ms across all three, and its software-style sandboxing with regex AST checks adds about 3ms. The Windows integration was built in collaboration with Microsoft.

AIOpenClaw released its v2026.9.9 patch, adding support for GPT-6.1 Sol in Codex and Claude Haiku 5.5. The update also improves recovery from failed updates, fixes missing iMessage replies, and resolves scheduled-job issues. The release credits 90 contributors.

AIJetBrains has released Mellum2.1, a 12B mixture-of-experts thinking model with 2.5B active parameters, under Apache 2.0 on Hugging Face. Post-training reinforcement learning in real software repositories raised SWE-bench Verified from 2.0 to 47.0, according to JetBrains' self-reported results. Qwen3.5-9B still leads on SWE-bench Pro, GPQA Diamond and AIME, and GGUF builds start at 7.0 GB for local use.

AIOracle is using ChatGPT Work and Codex to turn specialist knowledge into fast, repeatable workflows across recruiting, engineering, and operations. The source does not provide figures, timelines, or specific results beyond the headline's claim that days of work can take minutes.

AIAnthropic released Claude Haiku 5.5, raising its OSWorld 2.1 score from 15.7% to 72.4% and supporting a 1 million token context window. The article notes Haiku 5.5 still scores 39.2% on Terminal-Bench 4.0 against Sonnet 5.5's 70.6%, and that prompts above 100,000 tokens are priced higher, so migration costs need to be measured on real workloads.

AIAugment Code is selling select assets, including Cosmos, Auggie CLI, and the Code Context Engine, to Harness, and the product team is moving to Harness. The company says Harness's integrated platform delivers these capabilities to customers more effectively than building them independently. Harness describes itself as building the Autonomous SDLC Platform for shipping AI-written code across enterprises.

Why it matters: The announcement shows how a coding AI company is folding its products into a larger software delivery platform, a shift that shapes how enterprise teams will buy these tools.

AITessl's Code Review defines review standards as skills in the repository, called lenses, routed to files by a repo-owned profile file. Because these team-visible standards are portable, lessons from review can feed back into code generation and maintenance, not only the next review.

AIOver six months, a solo founder held more than 80 calls with accountants about AI and wrote down 30 lessons from the recurring problems. The post's opening lessons say to treat AI setup like onboarding a new employee and to start by listing the tasks people hate doing. The author argues that many "model problems" are actually setup problems, and that a firm's software stack limits what AI can do.

AISoftware developer Gergely Orosz questions why he is working more despite AI tools, quoting Sam Newman's view that AI was meant to free developers from drudgery. Newman says most developers are doing more work, with more context switching and a loss of the big picture. The quoted post adds that AI assistants are not human partners and that pairing with them fragments the shared mental model of a program.

AIStepFun has released Step 5 Preview, its flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 27B active and 600B total parameters. The source says it performs strongly in software engineering and professional tasks, but the feed supplied only an excerpt, so benchmark details are not available here.

AIGoogle Cloud introduced Gemini agent, a unified AI assistant that acts autonomously, generates code, and completes work across web, mobile, desktop, and third-party apps. It runs jobs on models matched to each task, including Gemini Flash and a flagship frontier model, with Anthropic Claude models also available. Hard spend limits per project let companies enforce budgets and charge AI costs to departments.

AIJetBrains released Mellum2.1, a 12B mixture-of-experts model with 2.5B active parameters under the Apache 2.0 license, built for coding agents. Post-training shifted to reinforcement learning across thousands of environments and millions of sandboxed runs, and the model is available on Hugging Face. The source reports gains over Mellum2 on LiveCodeBench, AIME, GPQA Diamond, BFCL v4, IFEval, and SWE-bench Verified, and says it serves almost twice the tokens of Qwen3.5-9B under heavy load.

Why it matters: The post shows how reinforcement learning in real sandboxed environments changed a compact open model's repository work, with benchmark gains against Mellum2 and two peers.

AILegalOn cut its estimated daily Codex costs by 65% while maintaining development speed. The company matched Astra, Sol, and Luna to tasks and managed budgets strategically.

AIThe author suggests installing a coding agent such as Pi or DeepSeek Harness in a cloud computer and allocating unused Code plan quota to it. Grok Bot could then use those tokens to write code and render videos, so leftover quota is not wasted.

AIA suspected Chinese-speaking attacker breached several South Korean financial institutions between late September and early October 2026, reportedly stealing over 25,000 records from Shinhan Bank alone. The attacker used ARTEX, a Chinese open-source tool that uses AI language models to automate finding security flaws, and models named in the report include DeepSeek v4.1-flash, GLM-5.3, and Grok 4.6.

AILucas Beyer says he learned Python around 2004 as Blender's scripting language, long before deep learning. He calls it a full-circle moment, as the linked post shows LLMs writing Blender Python code to generate 3D scenes.

AIAnt Ling's Ling-3.1-flash is now available as a day-0 partner release on Vercel AI Gateway, free through October 13. The model is positioned as token-efficient, built for agentic and coding work, and suited to multi-step analysis.

AIAnt Ling announced a day-zero collaboration with AI/ML API, making Ling-3.1-flash available there for agentic and cowork scenarios. AI/ML API describes it as a 560B-parameter MoE model with about 25B active per token and up to 1M context, built for agents, coding, and long documents. The model is free to try on AI/ML API until October 13.

AIAnt Ling announced a day-0 collaboration with Kilo Code for its launch, with Ling 3.1 Flash available for free in Kilo Code until October 13th. Novita Labs is serving the model.

AIAnt Ling's Ling-3.1-flash is now free on the OpenCode coding harness for a limited week-long period. The model has 560B total parameters, 25B active parameters, and a 262K context window, according to the OpenCode background post. Ant Ling says it shows strong coding performance and thanks OpenCode for day-zero support.

AIAnthropic released Claude Haiku 5.5, priced the same as OpenAI's GPT-6 Luna, with a 1M-token context window. Artificial Analysis scored it 43 on its Intelligence Index, slightly ahead of GPT-6 Luna at 38, but it uses about 3x more output tokens at max effort.

AIThe author argues that in agent workflows, the real bottleneck is whether users can precisely express requirements, not the model or agent capability. When people work outside their expertise, they lack the precision needed for prompts and plans, forcing many imprecise iterations that waste time and tokens. The suggested fix is to have the model first teach the unfamiliar domain knowledge before acting.

AIClaude Code v2.1.294 fixes prompt and agent hooks written as instructions, which had allowed actions they should block. It also improves how prompt hooks on Stop and SubagentStop are judged, making Claude less likely to stop early.

AIBlock's AI capabilities lead describes using Claude Fable to plan large code migrations and direct smaller models like Opus and Sonnet on individual tasks. He says Block routes frontier and smaller models by task and keeps merges and production deploys behind human dual approval.

Why it matters: Block's engineering lead describes how frontier models orchestrate large migrations and how access, effort levels, and safeguards are managed across an organization.

AIThariq, of Anthropic, says the most common failure in AI coding is users working outside their domain who cannot write precise prompts and plans. As a result, they spend many turns iterating imprecisely. The quoted post contrasts this with programmers' advantage in vibe coding.

AIThis Next Token episode discusses Personal Agents, including Dots in Codex, memory and cloud computer permissions, and whether agents should act as assistants or digital twins. The hosts also cover Instinct's booking and business-travel model, hands-on projects built with Opus 5.5, and whether software, games, and hardware could become open source as AI makes rewriting easier.

AIA solo founder writes that after a rocky first trial, they gave Devin Max a second chance and now favor Devin Cloud for shipping real projects. The post praises Cognition and Devin Max's model access and pricing, and covers SWE-1.7 Lightning's speed and its tendency to consume limits quickly. It also compares GPT 6 Astra and Fable 5.1, finding Fable more efficient for bug fixes and features.

AIDeveloper Brandon Thomas used Claude Opus 5.5 and Rust to build Artcraft, a free open-source suite with Photocraft, Vectorcraft, and other apps that mimic Adobe products. The author tested Photocraft and found its basic editing commands worked where expected, but Free Transform behaved unpredictably. Thomas describes the software as early alpha and invites developers to contribute.

AIEpoch gave six models 11 real work tasks from its own operations, including graphic design, data insights, and research design, and graded outputs against employee standards. Fable 5.1 and GPT-6 Astra led on average task performance, reliably handling well-defined work such as coding and computational analysis. The report finds that all models still fail on open-ended judgment, including matching Epoch's standards, designing informative experiments, and generating diverse ideas, so the authors conclude AI cannot yet replace workers at Epoch.

Why it matters: The report separates well-defined task reliability from open-ended judgment failures, which benchmark scores on easily verifiable tasks would miss.

AIAnthropic released Claude Haiku 5.5, which it calls the fastest, cheapest, and most capable Haiku model so far. On average it costs about 75% less to run than Haiku 4.5, with input at $0.10 and output at $0.50 per million tokens for requests up to 100,000 tokens. Anthropic also cut Sonnet 5.5's cache read price from $0.20 to $0.10 per million tokens, which it says lowers run costs by about 20% on many agent tasks.