Anthropic Claude Code launches on web and iOS – sandbox cuts prompts 84%

Anthropic launched Claude Code on the web with multi‑task, parallel execution and automatic PR creation, plus an early iOS preview for steering jobs on the go launch post, feature brief. Cloud sessions run on Anthropic‑managed VMs and share rate limits with other Claude usage, with real‑time progress and change summaries available in one UI launch blog. Following up on Mobile sighting, the mobile app now exposes a Code tab to queue and monitor tasks while away from a terminal mobile screenshots.

Claude Code now supports a configurable sandbox that whitelists directories and network hosts; bash runs with file and network isolation to curb prompt‑injection and exfiltration feature thread. Anthropic says the sandbox reduced permission prompts by ~84% in internal use, and you can enable it via /sandbox in the CLI and configure policies in docs cli notes, docs page. Under the hood, filesystem and network isolation details are in the engineering write‑up, and the sandbox runtime is open‑sourced for use in other agent workflows engineering blog, GitHub repo.

Hands‑on reports describe Claude Code on the web as an asynchronous coding agent that can branch, test, and open PRs from a browser, with modes for locked‑down networking and a "teleport" for moving work between local and cloud preview notes, preview notes. A separate vibe check praises the concept (kick off from phone, chat during execution) but flags beta friction: environment parity issues, flaky cloud VMs in prod repos, and incomplete mobile UX; the team expects rapid fixes vibe check, vibe check review. Mobile UI screenshots show the Code tab and parallel task queue on iOS mobile screens. In related stress‑testing, Claude Code autonomously set up DeepSeek‑OCR in a GPU Docker env in ~40 minutes with four prompts, illustrating the agent’s breadth even outside the web UI setup recap, setup write‑up.

A slowdown in Amazon DynamoDB in us‑east‑1 rippled through dependent services, knocking many consumer and AI apps offline or degraded for hours. Teams reported throttling, heavier use of caches, and staged restarts as common recovery patterns outage recap. Perplexity and others showed visible impact during the spike, underlining concentration risk in Virginia and the value of multi‑region plus cache‑first designs outage chart.

For AI inference operators, the takeaway is to assume upstream metadata stores can become the bottleneck. Region diversification, read‑through caches with generous TTLs, circuit breakers, and provider failover materially reduce blast radius.

A coding‑agent "big pickle" run reported 92% of tokens served from cache, leaving only 8% for GPUs—about a 12.5× cost reduction versus uncached compute cache stats. In separate Anthropic usage, cache hits reached ~98.5%, and with cache pricing the effective bill was ~6× lower than without it cache pricing. For agentic traffic with repetitive contexts, aggressive KV/token caching is now a primary lever for both resilience (fewer hot paths during incidents) and spend.

Researchers propose a Tail‑Optimized LRU eviction that keeps just enough KV cache per conversation to hit a target latency, reducing P90 TTFT by 27.5%, P95 by 23.9%, and 200 ms SLO misses by 38.9% on real traces, with minimal median impact paper summary. Following up on rate limits where provider behaviors stressed long jobs, this directly attacks tail latency inside the model server and can slot into existing LRU systems with a single extra flag.

Cline for Enterprise ships with routing across Anthropic, OpenAI, Google, and DeepSeek on Bedrock, Vertex, Azure or native APIs so teams can switch inference backends when a single provider or region is impaired, maintaining developer velocity during outages enterprise brief, with rollout details in the team’s post Cline blog post. This bring‑your‑own‑inference posture hardens agent workflows against regional incidents like us‑east‑1.

A concise Mastra example configures ordered fallbacks across OpenAI, Anthropic, and Google models with per‑model retry budgets—so an agent can degrade gracefully during a provider or regional incident fallback code. This pattern complements routing systems by baking resilience into the agent loop itself.

Vercel’s v0 acknowledged intermittent instability and directed users to live status updates during the event, then confirmed resolution later in the day status notice, with ongoing updates available on the incident page status page. For AI builders relying on v0 pipelines, this reinforces the need for retries, cached artifacts, and CI fallbacks when upstream platforms wobble resolution update.

DeepSeek released DeepSeek‑OCR as an open 3B‑param BF16 model with FlashAttention 2 under the MIT license, positioning it as “Contexts Optical Compression” rather than traditional OCR. The team claims large text corpora can be rendered as images and ingested with far fewer vision tokens, potentially shrinking context and cost while preserving layout, tables, and charts repo link, GitHub repo, model card.

The community summary cites aggressive throughput—on the order of 200k pages/day per GPU and tens of millions/day on a small cluster—and suggests this could shift how we think about long‑context, memory, and doc pipelines claims thread.

Karpathy argues many inputs to LLMs may be better as pixels: render text and feed images to enable bidirectional attention, capture formatting, and sidestep tokenizers’ brittleness, while compressing context size pixels essay. Community discussion extends this to agent memory: store history as progressively lower‑resolution image tiles to create a natural forgetting curve and cheaper long‑run contexts, potentially reducing classic RAG’s retrieval/chunking burden tile memory idea. Advocates claim entire libraries could fit into context once text is optically compressed, though evidence remains early and workload‑dependent claims thread.

A new survey of multimodal RAG for long, complex documents finds that mixing image and text cues and retrieving finer‑grained elements (tables, figures, text blocks) beats page‑level or image‑only approaches on grounding and answer quality survey paper. This lands alongside optical‑compression discourse and, following Layout‑first pipelines, suggests near‑term best practice is hybrid: preserve visual structure while retrieving just the salient elements to keep contexts small and faithful.

Beyond licensing and speed, practitioners note DeepSeek‑OCR’s decoder backbone (DeepSeek‑3B family) and FlashAttention 2 inference path, with reports that it opts for standard MHA instead of MLA in this configuration model card, arch notes. Early hands‑on threads highlight strong layout grounding, the ability to re‑render charts as HTML, and feature extraction via common vision stacks (e.g., CLIP/SAM‑style hints), positioning the model as a document understanding engine not just text extraction capability notes.

A field report shows DeepSeek‑OCR can be brought up on an NVIDIA Spark box (CUDA on ARM64) inside a Docker container with scripted setup and documentation captured along the way. The run took roughly four prompts’ worth of orchestration and ~40 minutes end‑to‑end, indicating a workable path to integrate optical compression OCR into existing GPU nodes without major bespoke infra deployment writeup, Setup blog, repo notes, GitHub notes.

As a practical counterpoint to heavy PDF/HTML stacks, Moondream 3 shows vision‑native extraction of complex parking signs directly into JSON—transcription plus rule segmentation—without a bespoke OCR+regex pipeline. It’s a small but telling example of how vision‑grounded models can emit structured data straight from pixels for downstream use product example.

Google DeepMind’s Veo 3.1 now ranks #1 on both Text‑to‑Video and Image‑to‑Video leaderboards, posting a 30+ point jump over Veo 3.0 and becoming the first model to surpass 1400, per Video Arena’s community votes arena announcement and acknowledgements from leadership DeepMind congrats, with a recap for analysts arena recap.

Following up on physics demos, this cements Veo 3.1’s perceived realism and motion quality at the top of community evals; creators can test it side‑by‑side in Arena’s workflows and Discord rounds image‑to‑video top.

fal made Krea Realtime 14B available immediately with streaming text→video and video→video endpoints, mid‑stream prompt edits, and on‑the‑fly restyling—positioning an autoregressive, real‑time model for production APIs fal announcement. Model weights are downloadable on Hugging Face under Apache‑2.0 for self‑hosting model weights, and fal provides open demos for both modes to try now demo links.

This lowers iteration cost for teams needing interactive previews (e.g., live UIs, creative tools) without waiting on long diffusion renders.

Community analysis argues Veo 3.1 leads on core model traits like physics and realism in side‑by‑sides, while Sora 2’s virality is driven by unique app features (Cameos) and automatic story‑building (Narratives) that play well on social analysis thread. For example prompts, the thread highlights a gymnastics test favoring Veo on physical plausibility physics compare and curated showcases of Veo 3.1’s image‑to‑video strength image‑to‑video top.

For teams planning content strategy: Sora’s product‑led virality may amplify reach, whereas Veo’s consistency can reduce failure rates in benchmark‑style or production workflows.

Genspark is granting every user one free Veo 3.1 generation until Nov 3 (11:59 PM PDT); invoke by prompting “use Veo 3.1” in Super Agent or selecting Veo 3.1 in AI Video free access. Details and entry points are in Genspark’s workspace landing Genspark - The All-in-One AI Workspace.

This is a low‑friction way for teams to trial Veo 3.1’s motion and style before committing credits or integrating APIs.

Google’s guidance shows how to screenshot a Veo first frame, use Nano Banana image editing to change wardrobe, pose, hair, or background, then feed the edited frame back into Veo 3.1 to carry those changes through the clip—reducing wasted generations how to thread. The steps cover capturing the base frame, iterating edits, and re‑running the video with the refined keyframe step six, with community prompts invited for best practices call for tips.

This frame‑to‑video loop gives teams a practical control surface for character continuity and set dressing without custom fine‑tunes.

Higgsfield is offering a week of “Unlimited Sora 2,” bundling Sketch‑to‑Video, Max/Pro Max tiers, Enhancer, and an Upscale Preview; the offer ends Monday UTC, with an additional 200 free credits via engagement mechanics offer post. Upgrade flow and product catalog are outlined on the site plan details and pricing page Higgsfield.

For production users, this is a temporary capacity window to test Sora‑based pipelines and quality gates at scale.

Cline rolled out an enterprise variant that runs where developers work (VS Code, JetBrains, CLI, or embedded) while routing to whichever model and provider best fits the task—Claude, GPT, Gemini, DeepSeek across Bedrock, Vertex, Azure, or OpenAI—so teams keep coding even if one cloud goes down launch thread, with details in the rollout post Cline blog post. It preserves code inside your environment and lets enterprises govern costs/usage centrally, positioning Cline as an agent loop you control while the inference layer remains plug‑and‑play feature recap.

Amp added the Librarian, a subagent that searches across public and private GitHub repos from inside the agent loop, returning precise matches, deps, and examples; it’s integrated into workflows for upgrades and debugging tool intro, with usage and setup documented in Amp’s note Amp news page. This arrives after Amp made code reviews reproducible via thread sharing, expanding its review ergonomics thread sharing.

Amp’s CLI now supports editing prior messages and rolling back, making agent runs reproducible when you need to correct a prompt mid‑session or bisect a failing trajectory cli update. This directly addresses one of the most common pain points in iterative agent debugging.

The Codex CLI team identified and patched a bug that sometimes threw an “unsupported model” 400 mid‑session; the fix is rolling out, with reliability improvements promised next bug fix.

Jules (Google’s SWE agent) is working on an Interactive plan flow that proactively questions specs, absorbs docs/links, and saves project notes before it writes code—aimed at reducing rework from underspecified tasks feature preview, with a breakdown of the UX and what’s coming next feature article.

RepoPrompt shipped v1.5.3 with more reliable path discovery for Codex and Claude Code installs and better behavior when configs include many MCP servers—reducing setup flakiness for multi‑tool agent stacks release notes. It’s a quality‑of‑life upgrade for teams standardizing on MCP‑driven repos.

A new guide walks through authentication strategies for agentic workflows—managed OAuth via Composio for 250+ services vs. custom Anchor browser profiles for anything without an API—plus logs, token refresh, and decision frameworks for production setups guide announcement, with full write‑up and code examples in the blog Anchor blog post.

Mastra highlighted built‑in model fallback chains so agent workflows can automatically retry across alternate providers/models, a pragmatic hedge against single‑vendor outages and quota hiccups fallback doc.

A two‑day, $10k‑per‑model live trading benchmark shows DeepSeek Chat v3.1 at $14,164.80 (+41.65%), while Gemini 2.5 Pro fell to $7,089.26 (‑29.07%). Other finishes included Grok 4 at $13,753.32 and Claude Sonnet 4.5 at $12,445.94, with BTC buy‑and‑hold near flat at $10,406.09 benchmark chart.

Even if short‑horizon and volatile, the spread highlights how agentic strategies and risk controls vary widely across models for live market tasks.

Community tests of four “lithiumflow” variants on WebDev Arena report mixed quality—two runs looked strong while two were weak—prompting speculation about variable thinking depth; GPT‑5 remained best on the same prompt set user test, benchmark check, follow‑up note. This is a continuation of recent sightings on LMArena, following up on Arena sightings that flagged orionmist/lithiumflow appearances. You can try the board while the models are still listed WebDev leaderboard.

As always with pre‑release identifiers, names and routing may change; treat results as early signals rather than stable rankings first sighting.

LMArena’s WebDev board added four notable entrants: Claude Sonnet 4.5 Thinking 32k (tied #4), GLM 4.6 (new #1 open model), Qwen3 235B A22B (now #11, #7 open), and Claude Haiku 4.5 (#14). The changes signal deeper pushes into long‑context reasoning and coding tasks by multiple labs model additions, with live standings available at the official board WebDev leaderboard.

For practitioners, this means more options to A/B against GPT‑5 and Claude Opus on complex full‑stack prompts without swapping harnesses.

IBM named Groq its high‑speed inference partner for Watsonx, with IBM’s CCO saying AI "has a cost problem" that Groq helps break through. IBM cites up to 5× faster responses at roughly 20% of prior costs, positioning Groq as an obvious enterprise choice for low‑latency workloads Bloomberg segment, Bloomberg video, performance claims.

Sourcegraph’s Cline now runs where teams work (VS Code/JetBrains/CLI) while letting enterprises pick models (Claude, GPT, Gemini, DeepSeek) and providers (Bedrock, Vertex, Azure, OpenAI). If one cloud has an outage, orgs can switch providers and keep shipping, with governance and cost controls intact enterprise launch, Cline blog.

Model aggregation keeps expanding buyer options: OpenRouter is routing access to a GPT‑5 variant that isn’t exposed through OpenAI’s endpoints, reinforcing the value of multi‑provider brokers for capability and availability coverage model routing pitch.

Model fallback arrays are becoming table stakes for production agents: Mastra demonstrates cascading retries across models/providers to preserve uptime when a single vendor blips, a pragmatic pattern for today’s uneven infrastructure fallback demo.

Distribution economics are shifting: Amp claims zero‑cost usage by routing to good, cheap, available tokens and leaning on fast open‑source models—an aggregator strategy that exploits price/perf spreads across providers pricing claim.

CoreWeave and Poolside are building “Project Horizon,” a 2‑gigawatt, self‑powered AI data center campus at Longfellow Ranch in West Texas to bypass utility interconnect delays, with on‑site generation tied to nearby natural gas production campus overview.

Phase one anchors 250 MW on a 15‑year lease (with 500 MW reserved for expansion), while Poolside targets a cluster of ~40,000 Nvidia GB300 NVL72 GPUs starting December 2025; the build uses hybrid modular blocks and parallel construction for staged capacity ramps campus overview. This on‑site power approach arrives as AI buildouts outpace grid upgrades, following up on grid storage noting surging batteries deployed alongside AI datacenters.

OPPO’s A2FM trains a router to choose among instant answer, step‑by‑step reasoning, or agent tool use, reporting $0.00487 per correct answer—45.2% cheaper than reasoning‑only and 33.5% cheaper than tool‑heavy agents at comparable accuracy paper thread. Following up on self-learning loop that improved agents without labels, A2FM adds learned mode‑selection: train the router first on mixed difficulty, then align each mode; the agent mode plans and runs web/code tools in parallel while nudging easy prompts toward instant responses.

Implication: hybrid routing can prune unnecessary CoT and tool calls, making agent stacks cheaper without sacrificing solve rates.

GUESSBENCH shows models that score ≈91.2 on passive vision drop to ≈43.1 when they must ask yes/no questions to find a target image—active reasoning craters on fine‑grained synthetic visuals, with real photos faring better. Larger models help somewhat; explicit thinking and disciplined early‑stopping improve results paper thread.

Takeaway: perception granularity and question planning are the bottlenecks; evaluation should include “ask‑to‑learn” loops, not only passive VQA.

M²PO for machine translation fixes two DPO pain points—weak single‑judge signals and wasted pairs—by (1) penalizing unfaithful tokens via word alignments, (2) blending an external quality score with a calibrated self‑judge, and (3) training on many top‑vs‑bottom pairs with listwise ranking plus a light behavior‑cloning term. On WMT21‑22 it improves quality and source faithfulness, beating GPT‑4o‑mini and approaching GPT‑4o while reducing hallucinations without hurting fluency paper thread.

Engineers building domain MT can adapt the pattern: multi‑perspective rewards and listwise ranking extract more learning per batch than 1‑v‑1 preferences.

A new benchmark (ReasonIF) finds fewer than 25% of large reasoning models’ hidden traces obey simple rules (language, word caps, JSON, disclaimers) even when the final answers follow them. A two‑turn redo only modestly helps; small SFT on synthetic traces lifts trace compliance from 0.11 to 0.27 with a slight accuracy trade‑off paper thread.

For AI engineers adding tool plans or chain‑of‑thought, this quantifies “instruction drift” inside the think‑step and suggests separate constraints for traces vs outputs.

A comprehensive survey of multimodal RAG for document understanding finds that retrieving fine elements (tables, charts, text blocks) with mixed signals (image + OCR/text) reliably beats page‑level or image‑only pipelines on grounding and answer accuracy survey thread.

Practical cues: combine closed‑doc and cross‑corpus retrieval, add verification/agent loops, and describe images for search when embeddings struggle—echoing practitioner results that summarization‑style descriptors often outperform raw embeddings for image queries Practitioner tip.

Google made Maps grounding generally available in the Gemini API, tying model answers to live data on 250M places and returning a context token to render an interactive Maps widget alongside responses feature brief, with full details in Google blog post. Apps can pass lat_lng to anchor results, combine Maps with Search grounding for freshness, and handle itineraries, hyper‑local picks, and precise place facts.

• Supports structured facts (hours, photos, ratings) and grounding metadata; tool pricing applies feature brief.

A practitioner reports that generating rich, grounded image descriptions (objects, spatial relations, visible text + nearby context) routinely outperforms raw embeddings for multimodal search quality. You can iterate prompts when results degrade, but you can’t fix a bad embedding post‑hoc—making summary‑first indexing the safer default for production search practice note.

A new survey synthesizes patterns for reliable document understanding beyond context limits: retrieve smaller, relevant chunks, prefer mixed image+text signals over image‑only, and use element‑level targets (tables, charts, figures) for sharper grounding. Hybrid graphs for linking parts and agents for plan‑fetch‑verify loops further boost answer faithfulness survey summary.