EN RU

Daily digest

18 items · ~18 min · Week 2026-W39

Must-read (3)

Anthropic releases Claude Opus 5.5 at lower prices

Anthropic
Models / LLM official + media 2 src. ~1 min

Anthropic launched Claude Opus 5.5, its strongest model to date, performing at the level of Fable 5.1 on most work. Pricing drops to $4/$20 per million tokens with cache reads cut to $0.20, and Sonnet 5.5 and Haiku 5.5 are promised in the coming weeks. It carries Fable-level safeguards and was tested externally by METR and Frontier Design.

Why it matters
Anthropic is compressing the gap between its best and cheapest frontier models at a 20-60% price cut, hours before OpenAI countered with its own cheaper GPT-6 tiers.
#claude #pricing #benchmark #safety

WorldCrafter: Consistent Video World Model with Implicit 3D-aware Memory

Tencent ARC Lab
Research official 2 src. ~1 min

A video world model with a camera-queryable implicit 3D-aware memory: the target viewpoint conditions how historical observations are compressed into the generator's token budget, without explicit depth correspondences. A jointly-trained memory encoder plus pose-conditioned readout and few-step distillation support minute-scale streaming scene exploration from a single image or text prompt.

Why it matters
Top-voted paper of the day on HF Daily Papers (194 upvotes) and a step toward interactive open-world exploration with long-horizon consistency, the core blocker for playable world models.
#video-generation #world-models #3d #paper

GameHorizon Suite: Multi-Horizon Data and Evaluation in Gameplay

Tencent ARC
Research official 2 src. ~1 min

An evaluation suite for AI gameplay across temporal horizons: a 5,000-hour AAA-game dataset (21 games, 100 expert players) with time-aligned videos, actions, and multi-horizon instructions, plus an offline reproducible benchmark and an online track that checks whether real gameplay matches offline scores. 47 models evaluated with over a million invocations.

Why it matters
HF Daily Papers' second-highest vote of the day (126 upvotes) — fills the gap between narrow game benchmarks and long-horizon interactive play where current model capabilities degrade measurably.
#evaluation #agents #games #datasets #paper

Worth knowing (8)

Suno v6 hit by second Universal/Sony copyright suit; Suno confirms v6 trained on users' creations

Suno
Audio media only 4 src. ~1 min

Universal Music and Sony Music filed a second copyright lawsuit against Suno, targeting the new v6 models released days earlier as 'the fruit of the same poisoned tree' — i.e. allegedly trained with the same infringing data as the original suit's models. Suno publicly confirmed v6 was trained on material from users' own uploads and called the suit 'fundamentally flawed on both the facts and the law', while paying users report a quality drop versus v5.

Why it matters
First time a label-deal AI music model is itself the subject of a fresh infringement suit, testing whether 'licensed partnership' deals shield or expose underlying training-data claims.
#music-gen #regulation #release

Tencent open-sources EVIE visual document retrieval models

Tencent
Models / LLM official 3 src. ~1 min

Tencent released EVIE-8B and EVIE-4.5B, visual document retrieval models built on Qwen3.5 backbones using late-interaction (ColBERT-style) multi-vector embeddings over document page images. EVIE-8B ranks #1 on ViDoRe V3 at 66.75 nDCG@10, with EVIE-4.5B ranking #2 at 66.02; both ship under Apache-2.0 with training and distillation code on GitHub. The 4.5B student supports runtime truncation of its 2048D projection head down to 64D, reducing index size.

Why it matters
First open models to top the ViDoRe visual document retrieval leaderboard, meaning genuinely usable open document-RAG for enterprise pipelines.
#retrieval #embedding #open-source #vision-language

Intern-S2-397B: Shanghai AI Lab's strongest open scientific multimodal model

Shanghai AI Laboratory (InternLM)
Models / LLM official 2 src. ~1 min

Intern-S2-397B is Shanghai AI Lab's most capable multimodal foundation model for scientific intelligence and long-horizon agents, combining a vision-language pretraining paradigm that learns directly from raw scientific-literature pages (without OCR-style intermediate parsing) with RL across 20+ scientific domains and black-box agentic RL in sandboxed environments. Released under Apache-2.0; exact benchmark scores are published as rasterized charts, not readable text.

Why it matters
Advances open-source science-focused models on a distinct axis from general chatbots — biomolecular interaction design, material structure generation, and long-horizon agent tasks.
#open-source #multimodal #science #agents #rl

OpenAI launches GPT-6 Sol and Luna with 50% API price cut

OpenAI
Models / LLM media only 2 src. ~1 min

OpenAI released GPT-6 Sol, for complex work like coding, and GPT-6 Luna, for high-volume summarization and extraction, with API pricing roughly half the 5.6-series models. The company claims Sol makes about half as many mistakes as its predecessor and reaches Astra-level reliability at much lower cost. Rollout spans ChatGPT Work, Codex, and the API, with Luna also coming to desktop and free tiers.

Why it matters
OpenAI's mid-tier price collapse lands within hours of Anthropic's Opus 5.5 cut, signalling an intensified frontier price war.
#gpt-6 #pricing #coding

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Google Research
Research official 3 src. ~1 min

Harness self-improvement (auto-evolving prompts, control flow, tooling around a frozen model) routinely overfits the training tasks. RRSI adds regularization: a temporally annealed edit budget on the proposer, and a selector with a critic that filters benchmark-specific proposals and a pruner that drops low-value changes.

Why it matters
Up to +4.7 points on out-of-distribution benchmarks and 30% fewer policy tokens, showing self-improving agent harnesses can generalize rather than just memorize the evolution split.
#agents #reasoning #rl #generalization #paper

Codex CLI 0.156.0 ships fullscreen TUI, voice mode and usage dashboard

OpenAI
Tools official 1 src. ~1 min

OpenAI's Codex released rust-v0.156.0 on Sep 22 with an optional fullscreen UI via /tui (transcript search, mouse selection), voice conversations enabled by default (F8 toggle), a /usage analytics dashboard, worktree sessions by default, and Mermaid/equation inline rendering. The release also closed sandbox isolation gaps on Windows inbound connections and privileged Linux/macOS sockets, and bundled OpenSSL 3.6.4. Alpha train 0.157.0-alpha.2 through alpha.10 landed across Sep 21-22.

Why it matters
Fullscreen transcript search and per-skill usage analytics make Codex noticeably more usable as a daily-driver terminal agent.
#codex #openai #cli #agents

Claude Code v2.1.280 adds Claude Opus 5.5 as default Opus model, fixes symlink write escapes

Anthropic
Tools official 2 src. ~1 min

Claude Code v2.1.280 (Sep 22) adds claude-opus-5-5 as the new default Opus model — 1M context, $4/$20 per Mtok with $0.20 cache reads. Security-relevant fixes: acceptEdits/allow rules/auto mode no longer approve writes outside the tree through symlinked paths, and auto mode stops endlessly retrying actions declined by safety checks. Also adds a CLAUDE_CODE_MAX_MCP_DESCRIPTION_LENGTH env var and mouse support in fullscreen.

Why it matters
The symlink write-through fix closes a real sandbox escape, and the price drop makes Opus-class coding agents meaningfully cheaper per token.
#claude-code #anthropic #cli #opus #security

Runway previews real-time video generation: streaming video steered as it is described

Runway
Video official + media 2 src. ~1 min

Runway published research ('Towards Instant Video Generation') on generating video causally, frame by frame, so users can steer the stream as it is described instead of prompting and waiting; the model is trained on its own imperfect outputs to prevent small errors compounding over long continuations. It builds on GWM-1/Gen-4.5 and Runway says real-time generation cuts per-output GPU cost; no product release timeline announced.

Why it matters
Run-time controllable framing pushes AI video from 'prompt-then-wait' clips toward live-interactive media, and drops unit economics for serving video.
#text-to-video #research #real-time #world-models
For reference (7)

Qwen-Image-2.1 gets a prompt-enhancer companion

Alibaba (Qwen)
Image official 3 src. ~1 min

Qwen shipped Qwen-Image-2.1-PE-T2I and Qwen-Image-2.1-PE-I2I, dedicated prompt-enhancer adapters for the Qwen-Image-2.1 T2I/editing pipeline: a fine-tuned Qwen3.5-VL 9B that rewrites a short user prompt into a detailed English prompt plus a recommended aspect ratio, output as JSON. Text output only, gated behind the Qwen Research License; the HF API page came back empty on fetch, so the summary is based on model metadata rather than a full card.

Why it matters
Keeps building out the open Qwen image-generation stack with a prompt-enhancer companion that previously would have required a separate closed tool.
#image-generation #open-source #prompt-enhancement

YooKassa ships an MCP server letting AI agents accept payments

Sber (YooKassa)
Tools official + media 3 src. ~1 min

YooKassa, the Sber-owned payments platform, launched an MCP server that lets AI agents connect directly to its payment infrastructure: create invoices, issue payment links, and check payment and refund statuses from inside a chat. Online merchants can therefore accept payments through agent workflows without custom integration work.

Why it matters
One of the first production payment rails wired into the agentic-AI ecosystem in Russia.
#mcp #payments #russia #agents

vLLM v0.30.0 published: DeepSeek-V4.1-Flash with MXFP8 FlashMLA KV on Blackwell

vLLM project
Tools official 1 src. ~1 min

Follow-up to yesterday's tag-only sighting: vLLM released v0.30.0 on Sep 22 with 762 commits from 315 contributors (104 new). Headline: DeepSeek-V4.1-Flash support with the entire KV cache kept in MXFP8 via FlashMLA V4.1 on SM100, plus more model additions.

Why it matters
Day-scale support for DeepSeek's new Flash model with a SM100-optimized attention path makes it immediately servable at scale on Blackwell GPUs.
#vllm #inference #deepseek #blackwell

GitHub Copilot gains Claude Opus 5.5, GPT-6 Sol and Luna, and Grok 4.7

GitHub
Tools official 1 src. ~1 min

GitHub's changelog (Sep 21-22) records Copilot availability of Claude Opus 5.5, OpenAI's GPT-6 Sol and GPT-6 Luna, and Grok 4.7 — three frontier models added within two days. Separate entries cover Copilot for JetBrains improvements and faster C++ code intelligence via whole-codebase indexing.

Why it matters
Model choice inside Copilot now spans Anthropic, OpenAI and xAI flagships simultaneously, accelerating the multi-vendor default in enterprise IDEs.
#copilot #github #models

Cline v4.1.20 runs same-step sub-agent tool calls in parallel

Cline
Tools official 1 src. ~1 min

Cline released v4.1.20 on Sep 22: sub-agents spawned in the same step now execute their tool calls concurrently instead of sequentially, reducing wall-clock time for fan-out tasks. Two SDK releases the same day — sdk/v0.0.84 and sdk/v0.0.85 — add retry on runs that hit output-token limits without a usable tool call, and dedupe concurrent feature-flag polls.

Why it matters
Parallel sub-agent execution cuts the latency penalty of agentic decomposition, a pattern other agent harnesses have been converging on.
#cline #vs-code #subagents

Adobe Premiere (free, Firefly-powered) mobile editor launches on Android

Adobe
Video official + media 3 src. ~1 min

Adobe released its Premiere mobile video editor on Android (Sep 22), free with layered editing, YouTube templates and Firefly generative-AI tools, nearly a year after the iOS version; it replaces Premiere Rush, whose support ends September 30. Press coverage notes the 'free comes with a catch' framing and steep RAM requirements, positioning it against Instagram Edits and CapCut.

Why it matters
Puts a full professional-grade Firefly-backed editor in consumers' pockets on the biggest mobile OS, pressuring ByteDance's grip on casual mobile editing.
#video-editing #image-to-video #release