A radar over AI and agents. Every morning, and only what several independent sources are saying at once.
Anthropic has introduced Claude Opus 5.5, the first model in the Claude 5.5 family, delivering performance comparable to Claude Fable 5.1 while costing 40% less to run than Opus 5. The model features improved clarity and token efficiency across all effort levels.
Consonance, edition no. 1.
Sunday 27 September: what several independent voices are saying at once.
Major labs are launching cost-efficient new models and features.
Anthropic launches Claude Opus 5.5 with forty percent lower costs.
OpenAI introduces GPT-6 Sol and Luna models with lower API prices.
xAI releases Grok 4.7 with a score of forty-six on the index.
Platforms are expanding agent capabilities and multimodal voice features.
ChatGPT Voice integrates plugins and expands pricing tiers.
GitHub Copilot introduces Autopilot and integrates GPT-6 family models.
Meta announces new integration partners for Muse at Connect.
Infrastructure and developer tooling see significant cloud updates.
Claude Code updates cloud sessions and usage limits.
Google introduces Gemini Flash TTS models with voice design.
DigitalOcean launches Managed Agents in public preview.
Research explores improving reasoning accuracy and inference efficiency in language models.
The framework creates stable learning targets without requiring multiple live anchor models during training.
29 topics today, at least three independent voices for each.
The sources are on consonance.fyi.
Subjects at least three independent accounts raised over the last seven days.
Claude Code officially released cloud sessions allowing tasks to run while laptops are closed, and added a graceful stopping point when hitting the 5-hour limit. The platform also resumed charging for requests blocked by safeguards.
Google's Project Suncatcher is testing a TPU-powered satellite prototype aboard a SpaceX mission. Additionally, Google DeepMind published essays addressing misbehavior control in agent swarms and equitable AGI benefit distribution.
Hugging Face is conducting an extensive review of AI agents' internet access during training and evaluation, publishing ongoing transparency summaries. Meanwhile, new diarization models were released to track multiple speakers.
ChatGPT Voice added support for plugins like email, calendar, and Slack across web and mobile platforms. The ecosystem also prepares tiered subscription plans scaling up to a $500 Pro Max tier.
Google launched Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS, featuring over 2,000 production-ready voices, support for 100 languages, voice design tools, and voice replication.
Microsoft has announced a major update to GitHub Copilot, introducing proactive Autopilot and two new models: GPT-6 Sol for interactive coding and lightweight GPT-6 Luna. The release also includes a new Copilot application featuring Home, Code, and Copilot modes.
DigitalOcean has released Managed Agents in public preview, enabling users to run Claude Code, Codex, or LangGraph agents in runtime environments that pause when idle. The service places agent tools behind a governed endpoint and supports over 75 open-source options.
Grok 4.7 xHigh achieved 58% on the Artificial Analysis AA-Briefcase benchmark, ranking just one point behind Claude Fable 5.1 Max at 59%. Additionally, the stealth model Pixel Canary launched on Cline, tying with GPT-6 Astra and outperforming Kimi K3 on the Next.js Agent Evals benchmark.
OpenAI has launched GPT-6 Sol and GPT-6 Luna, bringing the core strengths of GPT-6 Astra into faster and more affordable models designed for scaled workloads. Both models launch with API prices 50% lower than GPT-5.6, accompanied by Tesseract, a video creative suite built for AI agents.
At Meta Connect, Meta announced strategic integration partnerships for its personal AI agent Muse with Shopify, Expedia, and PayPal. These collaborations enable users to streamline shopping, trip planning, and checkout processes through Muse across participating merchants worldwide.
Anthropic has launched a new portal allowing developers to submit plugins, track reviews, and monitor usage for Claude. These plugins package the Model Context Protocol (MCP) along with custom skills, serving as the standard mechanism for building applications on Claude.
Anthropic has introduced Claude Opus 5.5, the first model in the Claude 5.5 family, delivering performance comparable to Claude Fable 5.1 while costing 40% less to run than Opus 5. The model features improved clarity and token efficiency across all effort levels.
Claude Sonnet 5.5 is currently undergoing stealth testing featuring a 1-million-token context window and a 128K-token maximum output length. Pricing is set at $2 per 1M input tokens and $10 for output tokens, arriving as a direct response to rival model releases.
Anthropic has increased the performance speed of claude.ai and its desktop application by 3x within two weeks. The development team shared specific prompts and techniques utilizing Claude to measure, debug, and improve execution speed.
SpaceXAI has released Grok 4.7, described as its most capable model yet for coding and knowledge work. The development of this system was supported by NVIDIA accelerated computing infrastructure.
Jürgen Schmidhuber, a pioneer in meta-learning, recursive self-improvement, and world models since the 1990s, has joined Sakana AI as Chief Scientific Advisor. Concurrently, the company introduced Agora-2, a next-generation multi-agent world model supporting real-time interaction for up to 20 humans and agents.
Google has partnered with Planet to launch a prototype satellite carrying four Tensor Processing Units (TPUs) into orbit. The mission evaluates how the hardware withstands harsh space conditions to assess the feasibility of hosting machine learning in orbit.
Apple has released a model on Hugging Face built by finetuning Qwen3.5-9B to convert long documents into small page images to conserve tokens, subsequently retrieving full text only for the pages relevant to a user's query.
The agent building community is experimenting with System One models like Jev and the Contrastive Language Model within custom harnesses. These models serve as high-speed routers or verifiers to improve performance for agent tasks.
OpenRouter added several new models, including Xiaomi's multimodal MiMo-V2.6 series with a 1M-token context window and the System One model Jev. Developers are testing these models for applications requiring fast response times.
MiMo-V2.6-Pro achieved the top spot on the Artificial Analysis Intelligence Index for open weights models. The model utilizes a classic Grouped Query Attention architecture and delivers competitive cost efficiency per task.
Researchers released a post-training approach that helps agents learn from real user sessions by imitating successful trajectories and correcting errors. Another study from Google evaluates the impact of distilling agent harnesses on task success rates.
The vLLM framework deployed day-one support for the Xiaomi MiMo-V2.6 models and DiffusionGemma-Jev. Both the Pro and Flash variants provide long-context and multimodal processing capabilities on the serving platform.
Tencent ARC Lab introduced the GameHorizon Suite, a data and evaluation framework designed to measure AAA gameplay capabilities across multiple temporal horizons for vision-language models, GUI agents, and coding agents.
The winner of Anthropic's hackathon has open-sourced their complete Claude Code setup, including agent skills, plugins, and practical tips. Concurrently, new research introduces methods to evolve agent skills while achieving 40 to 70 percent lower token costs compared to frontier methods, alongside NVIDIA's work on compiling public agent skills into reinforcement learning environments.
Alibaba aims to reach 20 gigawatts of global data center capacity by 2032 to support its artificial superintelligence ambitions, alongside plans for a 5 to 10 trillion parameter Qwen model and proprietary AI chips. Additionally, the Qwen team released RecreationWorld on Hugging Face as a sandbox for hybrid computer-use agents to explore, implement, and visually verify builds.
Papers and writeups, read from their abstracts, ranked by relevance to someone building agents and backends.
AEWM achieves 70.5% macro-F1 on our Action Judge benchmark, exceeding the strongest frontier baseline by 10.6 points.
Collected data across nearly 39k sessions in which real players play alongside Ally for iterative training.
A total of 225,964 evaluations were completed across 14 models from 8 labs.
OmniHarness achieves a 95.0% resolve rate on ComfyBench's Creative tasks, exceeding the strongest baseline by 27.5 percentage points.
Across all program synthesis methods, we evaluate 980 generated programs on 100 held-out instances each, 98,000 evaluation episodes in total.
On LIBERO-Pro, WAA with skills evolved only from LIBERO-90 reaches a state-of-the-art 75.6% average success, outperforming end-to-end VLAs, code-as-policy agents, and a visual-harness baseline with the same backbone.
Finetuning on data generated by EdgeGen yields a consistent mean progress improvement of 2 percent to 42 percent on tau2bench airline domain, while other baseline methods show degradation for some models.
FLEET achieves the same accuracy as the repeated sampling baseline, with a 3x speedup.
The agents generated more than 850,000 LLM calls and nearly 50 billion tokens across 16 days in eight parallel worlds.
Frontier coding agents can solve complex long-horizon tasks requiring up to half an hour of continuous interaction.
On Qwen3.5-4B, it raises HMMT 2025 accuracy from 59.2% to 64.0% with 10.7% fewer response tokens, and LiveCodeBench v5 accuracy from 41.7% to 54.2% with 9.6% fewer response tokens.
Across five mathematics and physics benchmarks, self-organizing teams average 66.7% accuracy, versus 48.8% for their strongest member.
Achieving <10% ASR under standard Refusal Feature Ablation.
This research leverages a simple custom tool registered through a standard API feature to induce frontier models to externalize intermediate reasoning that is otherwise hidden in closed systems. By comparing against native CoT on open models and extending to systems like GPT-6 Astra, the authors demonstrate that the extracted reasoning matches native reasoning performance and substantially outperforms no-reasoning baselines across competition mathematics, science, and code generation. Structural analysis reveals that models like Astra exhibit token-efficient directed reasoning, resolving elementary steps internally while externalizing only crucial reasoning.
MemoryAthena raises the five-task average from 37.65 to 39.28 over the direct pathway of the same checkpoint.
AV-GRPO outperforms LTX-2.3 in generation quality, semantic alignment and cross-modal synchronization under LoRA and full fine-tuning.
The proposed estimators improve on the direct estimators in point and interval estimation, with near-nominal coverage.
The six layers that cause the least change are removed, corresponding to 18.5% of the encoder stack.
DeltaWAM with SDM improves average success over Fast-WAM from 81.3% to 85.4% in the clean setting and from 75.8% to 83.9% under visual randomization.
With only 20 generator updates, the recipe reaches 84.87 on the official VBench evaluation, outperforming the previous best few-step causal baseline by 0.36, while requiring 16 A100 GPU-hours.
KPRs integrate more information and better preserve existing content than rewriting from sources or regenerating from scratch.
Offline two-step planning evaluation places X-Planner second among four evaluated models on both BERTScore-F1 and a judge-based Overall score.
Every orthogonal diagonal-plus-rank-one matrix is exactly a CKDA transition matrix.
A single generic question reaches a median AUROC of 0.886 zero-shot and beats supervised baselines on most benchmarks.
Repositories climbing on GitHub today, scored on the day's momentum, rank, adoption, freshness and project health.