Agent Daily2026-09-23Global AI Brief
AI AgentCheaper Models, Leaner Execution
New models widen the price range, while open weights offer research starting points. Video tools, robot simulations, and asynchronous agents show how capabilities reach practical workflows.
14Worth watching
11Ready to try
3Open-source finds
3Regions covered
Today’s take: Cost competition now reaches beyond token prices to the steps needed to complete a task. Models, caching, and harnesses can save money; compare them using the same task, quality target, and test setup.
RISC Machine Guide
RISC = four systems for a production-grade agent or robot
A production-grade agent needs more than a brain. It must keep working, withstand failures, prevent overreach, and collaborate across real organizations.
The industry has delivered a capable brain. Production agents also need a body, an immune system, and a social interface.ALUX is building that complete machine.
AI Agent Radar
Models and harnesses reduce overhead
Sol and Luna offer model specialization; Unreal reduces turns spent managing tools. Selection should compare capability, execution design, and retry costs.
Creative tools preserve editing control
Hunyuan supports iterative reference-based edits; Clueso lets outside assistants change video projects. Useful iteration depends on preserving what should not change.
Tools and evaluations reveal context
relore traces the discussion behind code, while AISI’s evaluation cards document test configurations. Context helps prevent duplicated work and misleading comparisons.
New Features Worth Watching
GPT-6 adds two capability and price tiers
What changed: OpenAI released GPT-6 Sol and Luna on September 22. Both accept text and images, generate text, and support a 1.05M-token context window with up to 128K output tokens. At standard rates for shorter prompts, Sol costs $2 per million input tokens and $10 per million output tokens; Luna costs $0.10 and $0.50, respectively. Both offer reasoning settings and tool use.
Why it matters: The family offers lower-cost options for assigning complex coding and high-volume, focused tasks to different models. Compare the total cost of a successful task, not token prices alone.
Who should care: Developers and product teams choosing models for coding, document processing, and batch agent workloads.
Opus 5.5 targets cheaper long-running work
What changed: Anthropic released Opus 5.5 on September 22, saying it performs at the level of Claude Fable 5.1 on most tasks. Input and output cost $4 and $20 per million tokens; cache reads cost $0.20 per million tokens. The company reports 40% lower typical task costs than Opus 5 at default settings and over 30% faster output, alongside behavioral audits and external evaluations.
Why it matters: Agent costs reflect token rates, token use, and repeated attempts. Beyond lower prices, test whether the model needs fewer steps to finish real work.
Who should care: Developers and enterprise teams using coding assistants for cross-repository changes and automating knowledge work.
Xiaomi releases open-weight MiMo omnimodal RL checkpoints
What changed: Xiaomi published MiMo-V2.6 model cards and weights on September 22 Beijing time. Pro-RL has 1.02T total and 42B active parameters; Flash-RL has 309B total and 15B active parameters. Both support text, images, video, and audio, with a 1M-token context window. A Qwen3.5-9B-based distilled SFT checkpoint is also available. The releases declare MIT licensing.
Why it matters: The release offers research starting points at different scales, not just API access. Unified mixed reinforcement learning and within-group trajectory grading make the training approach worth examining beyond benchmark scores.
Who should care: Researchers and engineers working on multimodal agents, long-horizon RL, and deploying open models.
Grok 4.7 brings tiered reasoning to the public API
What changed: xAI released Grok 4.7 on September 21 with a 500K-token context window, text and image inputs, text output, and four reasoning levels. Below 200K prompt tokens, input, cached input, and output cost $2, $0.50, and $6 per million tokens. The Fast version, with twice the token rate, is available only through Cursor and Grok Build.
Why it matters: The same model name does not imply identical speed or pricing across products. Compare endpoint access, long-context rates, and reasoning settings together.
Who should care: Product and infrastructure teams comparing coding and knowledge-work models or managing multiple providers.
Hunyuan previews five-reference image editing
What changed: Tencent released Hy Image3.5 preview on September 22. It supports text-to-image generation, image editing, up to five references, and iterative edits using conversation history, with output up to 2K. Demonstrations emphasize small Chinese and English text, complex layouts, and subject consistency. Tencent Cloud’s China API charges RMB 0.15 per image at 2K or below, with no input charge.
Why it matters: Multiple references and iterative edits could help turn one design into a coordinated asset set. Cheaper iterations still need accurate text, stable subjects, and controlled changes.
Who should care: Creators and teams producing product images, infographics, marketing assets, and film characters or scenes.
Brief: Qwen adds a real-time omnimodal endpoint
What changed: Qwen listed Qwen3.8-Omni-Flash-Realtime on September 21. It supports live audio/video interaction, text and audio output, multichannel audio, video aggregation, and remote MCP tools.
Why it matters: Omnimodal understanding can support continuous interaction rather than only uploaded files.
Who should care: Developers building live assistants and multimodal apps.
Brief: Qwen adds voices to full-duplex speech
What changed: Qwen launched Qwen-Audio-3.1-Realtime-Plus on September 20. It retains the 3.0 Plus integration protocol and existing voices, adds eight system voices, and supports full-duplex conversations and voice cloning.
Why it matters: Existing speech apps can compare the new model without rebuilding their integration.
Who should care: Voice-assistant and real-time support developers.
Brief: Chat edits videos without flattening the project
What changed: Clueso launched its MCP integration on September 22. External assistants can create, recreate, and edit videos, changing assets, captions, branding, and section order while keeping projects editable in a shared workspace.
Why it matters: Conversation can change the video project itself, rather than merely suggest edits.
Who should care: Creators making tutorials and product demonstrations.
Brief: Microduck’s browser playground loads new robot movements
What changed: Microduck’s simulation playground added manifest-based loading for gaits, one-shot moves, and command scripts on September 10. MuJoCo physics and ONNX policies run in the browser, with keyboard or gamepad control.
Why it matters: You can explore how control policies change motion before buying hardware.
Who should care: Robotics enthusiasts and developers building interactive experiments.
GitHub Discoveries
Stars are an attention snapshot at writing time, not a quality ranking.
relore helps coding agents find work already underway
What changed: Hugging Face introduced relore on September 21 to connect GitHub issues, reviews, and pull requests to current code. It distinguishes maintainer decisions, contributor claims, and bot comments; it can find fixes underway and trace a line of code back to the relevant discussion. Queries run locally, while ingestion and index refresh contact GitHub. The license is Apache-2.0.
Why it matters: Repository memory needs more than text. Who made a decision, whether a fix already exists, and which code it concerns determine whether an agent repeats work or follows the maintainer’s intent.
Who should care: Developers maintaining mature repositories, code-review assistants, and team knowledge-search tools.
Unreal Agent moves tool waiting outside model turns
What changed: Unreal Labs introduced its asynchronous agent harness on September 22. Tool calls immediately enter an in-progress state; completion appends results and wakes the model, while users can steer work during the wait. The project provides a Go library, a command-line runner, and Harbor integration. The authors report lower costs with GPT-6 Astra at xhigh reasoning effort and link their evaluation runs.
Why it matters: Reducing polling, waiting, and redundant tool output can potentially save more than switching models. A harness determines how much work happens between model calls and deserves separate evaluation.
Who should care: Infrastructure teams building long-running agents, tool schedulers, and coding evaluation environments.
Global Technology and Market Watch
Brief: Transformers runs quantized GGUF models directly
What changed: On September 22, Hugging Face described GGUF inference in Transformers using ggml kernels. Initial support focuses on Apple Silicon and Qwen3.5, retaining the familiar Python generation interface.
Why it matters: Quantized models can plug into established evaluation and experimentation workflows.
Who should care: Local-model and PyTorch developers.
Brief: AISI publishes the setup behind benchmark scores
What changed: On September 22, the UK’s AISI and EvalEval described sharing results through Evaluation Cards. The release covers five main benchmarks and six models, with configurations and two additional cybersecurity evaluations.
Why it matters: Apparently comparable scores may use different compute budgets and feedback protocols.
Who should care: Readers evaluating models or comparing them for procurement.
Brief: Larger RL batches do not always train faster
What changed: Tencent’s September 22 study compares reinforcement-learning sample efficiency with system throughput. On fixed hardware, its best GRPO configuration reached the same validation target 29% sooner, while oversized batches slowed progress.
Why it matters: Faster training means reaching the same quality sooner, not simply processing more samples.
Who should care: Reinforcement-learning and compute-planning teams.
This issue: Model prices, agent turns, and training throughput each capture only part of cost. Compare costs at the same quality target to see whether savings come from eliminating waste or skipping necessary work.
Sources
- GPT-6 Sol and Luna / Reference 2 · GPT-6 Sol and Luna / Reference 3 · GPT-6 Sol and Luna / Reference 4
- Claude Opus 5.5 — Opus 5.5 targets cheaper long-running work
- MiMo-V2.6 — Xiaomi releases open-weight MiMo omnimodal RL checkpoints · MiMo-V2.6 / Reference 2 · MiMo-V2.6 / Reference 3 · MiMo-V2.6 / Reference 4
- Grok 4.7 — Grok 4.7 brings tiered reasoning to the public API
- Hy Image3.5 preview — Hunyuan previews five-reference image editing
- Qwen3.8-Omni-Flash-Realtime — Brief: Qwen adds a real-time omnimodal endpoint · Qwen3.8-Omni-Flash-Realtime / Reference 2
- Qwen-Audio-3.1-Realtime-Plus — Brief: Qwen adds voices to full-duplex speech · Qwen-Audio-3.1-Realtime-Plus / Reference 2
- Clueso MCP — Brief: Chat edits videos without flattening the project · Clueso MCP / Reference 2
- Microduck Sandbox — Brief: Microduck’s browser playground loads new robot movements · Microduck Sandbox / Reference 2
- relore — relore helps coding agents find work already underway · relore / Reference 2
- Unreal Agent — Unreal Agent moves tool waiting outside model turns · Unreal Agent / Reference 2
- Transformers GGUF — Brief: Transformers runs quantized GGUF models directly
- AISI and EvalEval — Brief: AISI publishes the setup behind benchmark scores
- Hunyuan Batch Size Scaling — Brief: Larger RL batches do not always train faster