At AI Engineer NYC 2026, Anthropic introduced its Dynamic Workflows feature released Friday, in which Claude writes its own orchestration program to spawn agents, gather results and manage information flow, without developers pre-configuring coordinator and subagent lists.
GitHub released the Copilot app for Beginners tutorial series, covering four ways to give Copilot project context, including .github/copilot-instructions.md custom instructions.
Why it matters: GitHub's official tutorial lays out four mechanisms for giving Copilot project context and when to use each, so readers can combine them as needed in their own repos.
OpenAI co-founder and current Anthropic member Andrej Karpathy argued that there is a huge gap between the public's perception of LLM capabilities and reality.
Vals AI tested GPT-6 Sol and Claude Opus 5.5 in single-agent versus team setups on Vibe Code Bench, finding teams cost 1.8x to 5.1x as much as single agents, and of four comparisons only GPT-6 Sol at the medium reasoning tier improved 7.3 points with statistical significance.
Epoch AI used its InnovationEval benchmark to test whether AI agents can independently invent new training methods: GPT-5.6 Sol reached only about 35% of the human reference method SDPO's relative improvement over GRPO under lenient scoring, only about 15% for changes within the rules, while Claude Fable 5 showed no measurable gain.
According to Business Insider, Google is testing several Gemini 4 variants — Argon, Barium, and Carbon — with Carbon recently deployed on its internal coding platform Jetski, reportedly outperforming Argon on coding tasks; some employees likened it to Anthropic's strongest coding model, Opus 5.5.
A study found that after AI coding agents were introduced, pull request review time from submission to merge rose 49% on average, the share of pull requests requiring changes nearly doubled, comments per pull request rose 35%, and the share of employees participating in code review increased 14%.
Anthropic added dynamic workflows to Claude Managed Agents, where a lead agent plans and distributes tasks to sub-agents, with up to 1,000 agents running in parallel per execution.
Amazon Bedrock launched a public beta of OpenAI-powered Managed Agents and added new models including GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Luna, Claude Opus 5.5 and Kimi K3.
Simon Willison used Codex voice mode in the ChatGPT desktop app, spending about half an hour while cooking dinner working almost entirely by voice, to build and ship a Newsletters page for his Django blog, using GPT-6 Astra High.
Simon Willison switched ttok's default tokenizer from GPT-4 to GPT-5/GPT-6, and used the occasion to release version 1.0.
LegalOn reduced its estimated daily Codex costs by 65% while keeping development speed unchanged. The team matched models like Astra, Sol and Luna to appropriate tasks and managed budgets with fine-grained controls.
Oracle uses ChatGPT Work and Codex across recruiting, engineering and operations, turning expertise into fast, reusable workflows. The article gives examples showing work that once took days now being done in minutes across those scenarios.
SpaceXAI has announced it is joining the Omacom foundation behind Omarchy as a Founding Corporate Patron and is donating $1.5 million worth of Grok tokens.
Anthropic released Claude Haiku 5.5, the first Haiku-tier update in about a year, scoring 43 on the Artificial Analysis Intelligence Index, surpassing GPT-6 Luna (38) and GLM-5.3 Flash (42).
Why it matters: The piece compiles third-party benchmarks and actual pricing details, letting readers weigh the real cost-effectiveness of small models in low-cost agent workflows.
Anthropic released Claude Haiku 5.5, cutting average prices by about 75% versus Haiku 4.5, with discounts of up to 90% for prompts under 100,000 tokens, and for the first time offering adjustable reasoning levels in a Haiku-tier model.
Why it matters: The article compiles Haiku 5.5's price cuts, key benchmark data and adjustable reasoning levels, allowing readers to compare its positioning against GPT-6 Luna and Sonnet 5.5.
Microsoft Research Asia proposed the Harnessed Agentic RL training paradigm and open-sourced a rebuilt Agent Lightning v1.0, a roughly 3,500-line framework that lets the real agent harness used in deployment participate directly in reinforcement learning via an LLM proxy, with no need to reimplement the agent inside the training framework.
Why it matters: The article explains the design ideas, architecture components, and reproducible training results of Harnessed Agentic RL, letting readers assess its fit for existing agent training workflows.
French company Mistral released Mistral Large 4, a 1 trillion-parameter open-weight model nicknamed Le Chonk, available now as a preview with the final version due before the end of the month, free for anyone to use and customize.
The NVIDIA team, building on Nemotron 3, used SFT, RL, and a generate-verify-improve reasoning system to score 30/42 at IMO 2026 (gold threshold 29), while Nemotron-3-Ultra-CC scored 535.4/600 in unofficial IOI 2026 testing (gold threshold 361.12).
Why it matters: The article lays out the complete fine-tuning and inference recipe from base model to gold medal, with open weights, data, and pipelines, so readers can reproduce this domain-specialization approach.
Reflection announced its first free open model, Beam, with 501B total parameters and 23B activated per token (MoE), aimed at coding, logical reasoning and agentic tasks. According to Reflection, Beam matches GLM 5.2 on hard reasoning tasks with three to four times less compute, and comes close to Qwen3.8-Max on coding and agentic benchmarks, though its raw performance still trails stronger open models like Kimi K3.
Mistral released a public preview of its largest model to date, Mistral Large 4 (ML4, nicknamed le Chonk), with 1.05 trillion total parameters and 49 billion activated; the API is live, with weights expected by the end of the month.
GitHub has released ReviewBench, an open offline benchmark for evaluating AI code review agents. The benchmark is modeled on the distribution of 103.9 million GitHub pull requests.
HackerRank's AI interviewer Chakra has officially opened to customers after roughly six months in beta, during which it conducted more than 500,000 interviews; companies including Snowflake, Snorkel, and Capgemini tried it.
US startup Reflection AI officially released Beam, its first frontier open-weight model, claiming it performs comparably to Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute.
Simon Willison tested Mistral Large 4 with an SVG generation prompt about an armadillo in fishnet stockings jaywalking on Mars, comparing its output with claude-opus-5.5, gpt-6.1-sol, and gemini-3.8-flash, with each model using its default reasoning level.
Claude Opus 5.5 and Claude Sonnet 5.5 are now available on Amazon Bedrock in the AWS GovCloud (US) Region, and can be used with Anthropic's agentic coding tool Claude Code to support regulated development workloads, including ITAR.
Z.ai (Zhipu AI)'s GLM 5.3 is now available on Amazon Bedrock. It is a 753B-parameter MoE model aimed at coding and long-horizon agent tasks, and Z.ai reports a leading score of 84.5 on the CyberGym benchmark plus a 50% improvement over GLM 5.2 on its internal coding benchmark.
Mistral released the public preview of Mistral Large 4 (ML4), a natively multimodal model with 1 trillion parameters and 49 billion activated parameters, with weights to be available for download by the end of the month.
Why it matters: The official release covers parameter counts, benchmark scores, training infrastructure and the open-weights timeline, letting readers assess its real standing in the open-source camp.
OpenAI released GPT-6.1 Sol, priced at $2/$10 per million input/output tokens, ranking 5th on the Agent Arena, 39% cheaper than GPT-6 Sol with a score 1.52 points higher, and 81% cheaper than Astra.
AI is changing how developers work, and a GitHub blog post highlights three skills worth strengthening: learn to direct AI agents rather than just use them, don't trust AI's first answer (such as having a second model critique the first's output, like Copilot's Rubber Duck agent), and spend the saved implementation time on understanding customer needs and technical decisions. As AI takes on more implementation work, technical judgment and trade-off skills become more important.
Airbnb CTO Ahmad Al-Dahle discussed in a Latent Space interview how Airbnb is being built into an AI-native company. 60% of Airbnb's code is now written by AI, feature and improvement releases are up nearly 80% year over year, and engineers' average pull-request throughput has risen about 1.6x; roughly half of customer support tickets are fully resolved by AI, consistent with the near 45% reported in Q2 earnings.
Chatham Financial used Codex and GPT-5.6 to build tooling and redesign workflows, cutting trade validation time from 30 minutes to under 4 minutes, scaling its capital markets expertise.
Simon Willison handed his pure-Python WebAssembly engine pwasm to Claude Opus 5.5, and after 42 commits it released version 0.2a0. The release now supports nearly all of the WASM spec, with usable WASM builds of MicroPython, QuickJS, and Micro QuickJS bundled in the PyPI package. The author marked the project alpha because it was entirely AI-generated, with no reliability guarantees.
Google released its new model Gemini 4 Argon, saying it excels at cybersecurity, coding, and research, and can autonomously discover, verify, and fix critical software vulnerabilities. The model was specifically trained on defensive cybersecurity tasks and is currently available only to select cybersecurity partners through the Fairwind security program; Google, citing Vals' model index, says it leads OpenAI GPT-6 Astra and Anthropic's Fable and Opus on multiple benchmarks.
At OpenAI DevDay, Latent Space interviewed Ari Weinstein and Nikunj Handa, responding to Dwarkesh's criticism that Computer Use has stagnated and breaking down the newly released products.
Google released its new frontier model Gemini 4 Argon, saying it delivers frontier performance in software engineering, enterprise knowledge such as law and finance, and complex workflows like cybersecurity defense.
Google announced its new model Gemini 4 Argon, calling it industry-leading in coding, knowledge work and cybersecurity, but it remains in limited testing and is not yet available to general users. Google is already using it heavily internally: Argon helped save 300 TiB of data center memory using fleet-wide telemetry, and its agents have migrated over 800,000 lines of C/C++ code, including the Fuchsia OS Zircon kernel, to Rust.
Google DeepMind has released the frontier model Gemini 4 Argon, first opening it to trusted network defenders through the Fairwind Program, priced at $2 per million input tokens and $10 per million output tokens, with cached input at a 95% discount.
Why it matters: The official blog provides pricing, the 1M output token cap, benchmark scores and internal migration cases, so readers can judge its actual positioning in coding and security defense.
Latent Space's AINews rounds up multiple announcements from OpenAI DevDay 2026, including the always-on agent Dots (powered by GPT-6 Astra, connecting 4.
Why it matters: This roundup pairs DevDay announcements with independent benchmarks from Artificial Analysis and others, letting readers compare official claims against third-party testing.
AMD launched AMD Ross™, an agentic AI assistant covering the entire embedded development cycle, including architecture design, optimization, debugging, PCB design, schematic review, software and deployment. It connects to AMD Embedded tools via an MCP server, and bundles the AMD Knowledge Base, expert-written agent skills and runnable design examples, supporting natural-language document queries, command execution and workflow automation.