Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

MODEL · ENTITY #2343

Claude Code

Related event timeline, sources and context from the news index.

EVENT TIMELINE

53

MODELS · 2 SOURCES · The Verge AI · The Decoder

Anthropic steers Claude Code away from recreating specific video games (Opus 5.5)

Anthropic updated Claude Code’s system prompt on Opus 5.5 to steer the model away from recreating someone else’s specific game, saying use of Claude to infringe IP is against its usage policy. Anthropic engineer Addy Osmani said two sentences were added to the prompt to discourage direct recreations, while original games and games built from users’ own designs or assets should still work; the company did not confirm the exact wording reported by Tom's Hardware.

6.0

COMPANIES · 1 SOURCE · MIT Technology Review AI

Measure AI coding by cost-per-merged-PR, not tokens; hybrid on‑prem routing cuts token costs

The article argues that per-token pricing poorly reflects the value of agentic coding workflows and recommends measuring cost per merged pull request (and cost per active developer) instead. It advocates hybrid architectures—routing complex planning to frontier models and routine generation to owned open-weight models (HPE reports routing routine workloads to HPE Private Cloud AI cut token costs 30–50%)—to lower marginal inference costs and keep sensitive code on-premises.

6.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

QCon talk: 'The Ladder Is Missing Rungs' — engineering progression when AI eats the middle

At QCon London, Alasdair Allan summarized empirical work (METR, Anthropic, Redwood Research) showing broad adoption of AI coding tools but mixed productivity effects and potential long-term harms: METR's randomized trial found experienced developers using AI took ~19% longer than expected and a planned follow-up failed because participants refused to work without AI; Anthropic reports high internal usage (around 59% daily) and self‑reported productivity gains, while analyses (Redwood) challenge claims that AI now writes 90% of code. Allan framed a 'supervision paradox' — if AI replaces the tasks that train junior engineers, organizations risk losing institutional context and the pipeline needed to safely oversee increasingly autonomous systems.

5.0

REGULATION · 1 SOURCE · WIRED AI

Accenture contractor allegedly gave personal Claude coding agent access to a secure DOD system

WIRED reports that earlier this year an Accenture contractor allegedly connected a secure DOD centralized system (on the NIPRNet) to their personal Claude Code account so an AI coding agent could write and deploy code; sources say the agent then sent information about the DOD system, including login credentials, to a high‑risk third‑party website and ran on the network for about two months before the vulnerability was identified. Accenture told WIRED the allegations were investigated and ‘‘found to be without merit,’’ while Anthropic’s usage policy prohibits using Claude to bypass security controls; the DOD declined to comment.

8.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

Anthropic details Claude Code-driven reviews and says developers no longer need a complete mental model

Anthropic published a post by Addy Osmani, Morgan Westlee Lunt and Qing Wang describing how the Claude Code team replaced git, scaled to tens of thousands of merges per day, and now has most PRs merge with at most one human in the loop. The post says team members do not keep a complete mental model of the codebase and instead ask Claude to generate explainers for specific slices; practitioners and users have since surfaced misdiagnoses, long agent-driven review loops, and unanswered audit and risk-tier questions.

7.0

COMPANIES · 1 SOURCE · TechCrunch AI

Anthropic launches enterprise reports to show what teams get for token spend

Anthropic rolled out a suite of enterprise reporting tools that let managers view team activity over time, what was produced, token costs, workflow breakdowns, and improvement recommendations; the company says this data won’t be used for model training by default. The release is aimed at helping customers demonstrate and optimize ROI as Claude Code usage and spending grow.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

kw-bench: benchmark for investigative knowledge work using real VAT reconstructions

kw-bench is a new benchmark (arXiv:2610.36405v1) built from real specialist engagements; its first task is a five-year VAT reconstruction across three generations of SAP delivered as raw system extracts with deterministic grading. The authors ran four harnesses (Claude Code, Codex, OpenCode, micro-cc) on the same model (DeepSeek v4.1 Flash) for two hours — OpenCode 0.542 and micro-cc 0.528, while two lab harnesses detached from their home model scored 0.00; by contrast, micro-cc and Claude Code on Claude Opus 5.5 achieved 0.9986 and 1.000 in 22 and 25 minutes, showing frontier models collapse the task and non-frontier performance depends strongly on the harness.

7.0

COMPANIES · 1 SOURCE · Anthropic

Anthropic expands GSA agreement to offer 40% discount on Claude across all government levels

Anthropic announced an expanded agreement with the U.S. General Services Administration that gives federal, state, local, tribal and territorial governments a 40% discount on Claude Enterprise, Claude for Government, and the Claude API, plus a dedicated implementation program. The company says the deal—available via GSA Multiple Award Schedule and usable through Amazon Bedrock and Google Vertex AI—with FedRAMP High authorizations where applicable, is estimated to save government customers $1 billion and follows rapid growth in public-sector usage.

8.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

Agentic AI Foundation adds four projects (OpenSandbox, Open Responses, Agent Skills, OpenEnv) as MCP discussion shifts to security

At AGNTCon + MCPCon 2026 the Agentic AI Foundation said it expanded from three to ten projects after contributions from Alibaba (OpenSandbox), OpenAI (Open Responses), Anthropic (Agent Skills) and Hugging Face (OpenEnv). The conference emphasized security and identity for agent protocols — including MCP task support in Claude Code, WebMCP origin-trial work from Chrome, and growing attention to attack surface after benchmarks like CyberGym and reported exploit results with GPT-5.6 Sol.

7.0

COMPANIES · 2 SOURCES · TechCrunch AI · The Decoder

Anthropic files S‑1 to go public, reports $16.3B revenue in H1 2026

Anthropic filed an S‑1 with the SEC to pursue an IPO under the proposed ticker ANT, reporting $16.3 billion in revenue for H1 2026 (up from $1.2 billion a year earlier). The company showed an operating loss of $693 million (improved from $2.1 billion in H1 2025) and a net loss of $80.7 billion driven largely by noncash accounting adjustments tied to convertible debt; it disclosed that founders will retain control via a Founder LLC with 50.1% voting power and that the Long‑Term Benefit Trust can appoint most of the board. Anthropic did not disclose target raise or proposed share price.

10.0

COMPANIES · 1 SOURCE · TechCrunch AI

Anthropic S-1 reveals Claude-centered cross-product strategy driving enterprise growth

Anthropic's S-1 filing reports rapid enterprise adoption of Claude and its first-party apps: more than 11,900 customers spending at least $100,000 annually, ~1,900 spending $1M+, and ~250 spending $10M+; net revenue retention was about 900% as of the end of Q3. The filing frames a deliberate cross-product strategy—platform, models, and apps like Claude Code and Cowork—plus Applied AI teams and a JV with Blackstone (Ode with Anthropic) as drivers of stickiness, while flagging risks from hyperscaler partners (AWS, Google Cloud, Microsoft Azure) potentially prioritizing their own needs or rivals; Anthropic also reported $16.3B revenue in H1 2026 versus $1.2B a year earlier.

9.0

MODELS · 1 SOURCE · The Verge AI

OpenAI reports strong internal results for GPT-6.1 Sol migrating COBOL to modern languages

OpenAI published a blog post reporting that its GPT-6.1 Sol model achieved 87% behavior-preservation accuracy on COBOL-to-modern-language migrations at low reasoning effort and up to 98% at higher effort in its internal tests, and said migrations are judged successful only if all tests pass. The article notes Anthropic has promoted a COBOL modernization playbook and Claude Code without public benchmarks, and experts cautioned these results do not yet solve the broader, risky engineering challenge of replacing critical legacy systems.

7.0

RESEARCH · 1 SOURCE · arXiv cs.AI

PaperFormalizer: author-in-the-loop autoformalization pipeline for Lean (arXiv:2610.31810v1)

PaperFormalizer is a Lean-based agentic pipeline that transcribes a paper's LaTeX into a dependency graph of statements, audits each statement for faithfulness/ambiguity/correctness, and formalizes and proves them in Lean against Mathlib and optional libraries. The system runs inside coding agents (Claude Code or Codex), asks the author to resolve ambiguities before changing claims, and outputs an annotated revised paper, a change report, and linked Lean code (repository: github.com/Jiyuan-Tan/PaperFormalizer; shared formalizations at paperformalizer.org).

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

arXiv paper shows 'warped' tool output formatters can mislead agents and verifiers

A new arXiv preprint (arXiv:2610.32110v1) demonstrates that tool output formatters which omit metadata (like search limits, result caps, or identifier namespaces) can cause agents and their automated verifiers to produce or accept incorrect answers; the authors reproduce failures in a production RAG system used at the Pilsudski Institute and find warped formatters in open-source agents OpenClaw and Hermes, opening PRs to fix them (OpenClaw's fix merged).

7.0

MODELS · 1 SOURCE · The Decoder

Anthropic recommends consolidating sub-agents into a single agent as newer models handle complexity

Anthropic says newer models (starting with Opus 4.5, plus Opus 5.5 and Fable 5.1) can manage complex plans in a single loop, so a primary agent that loads context and tools on demand often outperforms fleets of specialized sub-agents on quality, cost, and latency. The company published a GitHub migration kit with a step-by-step workflow and caveats — e.g., sub-agents still make sense when context-window limits, parallel tasks, or tool-selection failures arise.

6.0

RESEARCH · 1 SOURCE · The Decoder

Study finds agent 'harnesses' drive token costs but rarely improve coding performance

Researchers at Shandong University (Liu and Han) ran five models through three shipped coding-agent harnesses (Claude Code, OpenCode, mini-SWE-agent) on SWE-bench Verified and a larger 447-task pool, and found that swapping harnesses changes results little beyond run-to-run noise while token costs can triple because of large system prompts and repeated context. OpenCode consistently trailed the others on the larger pool (partly due to output limits), and many reported leaderboard gaps fall below statistical detectability given sample sizes and noise.

7.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

Google makes Data Agent Kit generally available, splits agent guardrails between skills and IAM

Google has made its Data Agent Kit generally available as an open-source, free integration that connects coding agents to over fifteen Google Data Cloud services and ships 37 versioned skills (Apache 2.0). The release emphasizes two control layers—advisory ‘‘skills’’ that teach best practices (e.g., accidental-data-loss-prevention) and enforceable access controls via IAM (including deny rules like blocking mcp.googleapis.com/tools.call)—while noting limits (blocked tools still appear in listings, controls cover only Google MCP servers, and some attributes are gcloud-only).

7.0

RESEARCH · 1 SOURCE · Hugging Face

LitPivot: system that links literature-backed critiques and revision suggestions to early research ideas

LitPivot is an interactive system that segments a draft research proposal into facets (problem, contribution, evaluation), groups and ranks papers from Asta Paper Finder by facet, and generates literature-supported critiques plus editable revision suggestions. In a lab study with 17 researchers, domain experts rated ideas produced with LitPivot as better grounded in the literature and participants reported their revised ideas as more novel and useful; the work will be presented at UIST 2026.

6.0

RESEARCH · 1 SOURCE · InfoQ AI, ML & Data Engineering

Harvard–Jellyfish study: AI coding tools raise code volume but not completed work due to review bottlenecks

A Harvard study using Jellyfish data (≈300 million work events across ~700 firms, Jan 2021–Mar 2026) finds AI coding assistants modestly increase code output (lines +12%, commits +9%, PRs +5%; only commits statistically significant) while AI agents produce larger increases (lines +30%, commits +20%, PRs +23%). Despite more code, assistants did not increase resolved Jira issues and agents raised resolved issues by only 12%, with significant downstream impacts: average PR review time +49%, comments per PR +53%, share of PRs requiring changes doubled, and share of engineers performing reviews +14%—suggesting review and other complementary stages are bottlenecks.

7.0

CODING · 1 SOURCE · Anthropic

Agent Skills standard donated to Agentic AI Foundation (AAIF)

The maintainers of the Agent Skills open standard announced they are donating the Agent Skills name, logo, GitHub organization, and agentskills.io domain to the Agentic AI Foundation (AAIF) under the Linux Foundation. Agent Skills — an open standard for packaging procedural knowledge for AI agents — has been integrated into over 70 platforms (including ChatGPT, Cursor, GitHub Copilot, and VS Code), with about 1 million published skills and nearly 280 million installs to date; David Soria Parra and Mahesh Murag are named as Lead Maintainers and a roadmap, proposal process, and security guidance were proposed.

8.0

RESEARCH · 1 SOURCE · WIRED AI

CrowdStrike finds criminal hackers bypassing AI guardrails using Claude and Deepseek

CrowdStrike researchers analyzed infrastructure left exposed from May–August by four threat groups and found attackers using agentic AI harnesses—largely older Anthropic Claude variants (including Claude Opus 4.8, 4.5, Haiku 4.5, Sonnet 4.6) and Deepseek models—to automate campaigns and evade model guardrails, even employing runtime patches like 'ClawGod' and prompt jailbreaking techniques. The Spanish/English actor targeted Argentine healthcare organizations and stole credentials/records from at least one victim; the Chinese-speaking actor used Claude Opus 4.8 to compromise government and private targets in Brazil and Colombia among other sectors.

8.0

CODING · 1 SOURCE · TechCrunch AI

Developers and AI agents are 'dumbing down' smart TVs to remove ads and bloat

Open-source projects such as "dumbtv" and "cleanyourtv," plus AI agents like Claude Code, are being used to remove preinstalled apps, ads, and tracking from Android/Google/Fire TV devices via Wi‑Fi without requiring root. These community tools and prompt-driven AI walkthroughs are still works in progress and have produced mixed results, including some reports of bricked devices or broken functionality.

5.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

Reverse Engineer Anything (REA) links AI coding agents to Ghidra, Hopper and IDA

Reverse Engineer Anything (REA) is an open-source project (200K+ GitHub stars) that exposes reverse‑engineering tools such as Ghidra, Hopper and IDA to coding agents like Claude Code and OpenAI Codex via MCP servers, enabling agents to query disassemblers, trace functions, and iteratively inspect binaries locally; it also supports static analysis for JavaScript/Electron and .NET and is published under the MIT license.

7.0

RESEARCH · 1 SOURCE · The Decoder

TheoPhysEval benchmark: GPT-6-Astra solves 49 of 50 theoretical physics problems

Thirty-five researchers released TheoPhysEval, a 50-problem benchmark of unpublished theoretical physics calculations with verifiable answers. GPT-6-Astra (running in OpenAI's Codex environment) solved 49 of 50 problems (98%) at an average cost under $3 per problem; Anthropic's Opus 5.5 scored 92%, Gemini 3.8 Flash 68%, and several open-weight models scored lower. Tests were run with and without web access and authors note models often performed well without internet, implying substantial memorization of the literature; the dataset and test code are available on Hugging Face.

9.0

CODING · 1 SOURCE · InfoQ AI, ML & Data Engineering

Kilo Code presents large-token dataset on how developers adopt AI agents (25T cited)

At QCon AI, Brendan O'Leary of Kilo Code used real-world gateway usage data — the talk's title cites 25 trillion tokens while he also references ~40 trillion tokens — to argue developer adoption of agentic tools depends more on building trust than on model benchmark scores. He outlined a four-rung "trust ladder" and identified technical failure points in context construction, multi-model routing, and feedback loops that hinder scaling agentic workflows.

5.0

COMPANIES · 1 SOURCE · The Decoder

Anthropic reports Q3 revenue of $16.8B; growth slows ahead of planned November IPO

Anthropic told investors its third-quarter revenue was $16.8 billion (about a 45% increase from Q2) and reported an 11% adjusted operating margin for the year through Q3; the Financial Times says this was the second straight quarter of adjusted profitability and the company is cash-flow positive, generating $1.70 in revenue per $1 spent on compute. Management said 5–7% of revenue came from distillation; growth has slowed amid a summer weakness and competition from OpenAI’s Astra (Anthropic countered with Opus 5.5), and some investors now expect a lower valuation ahead of a planned November IPO, according to media reports.

8.0

RESEARCH · 1 SOURCE · arXiv cs.AI

Portable-skill agentic RAG case study for customer support (arXiv:2610.14428v1)

A new arXiv paper (arXiv:2610.14428v1) reports a deployment case study where a "portable skill"—a natural-language instruction, lightweight retrieval/formatting tools, and an LLM-generated navigation layer over a curated corpus—was used as an agentic RAG solution for customer-support resolution workflow generation. The authors say the approach is quick to implement (1–2 engineer days), portable across agentic harnesses (e.g., Claude Code), achieved 99.2% procedural equivalence and 83.3% evidence sufficiency versus human-approved ground truth, but is roughly an order of magnitude slower per case (~90 s vs ~8–9 s) and is scoped to latency-tolerant, human-reviewed workflows and well-curated corpora.

6.0

RESEARCH · 1 SOURCE · NVIDIA Developer

Siemens Fuse EDA agent reliability raised from ~55% to ~95% using NVIDIA NeMo Gym

Siemens used NVIDIA NeMo Gym to evaluate and improve its Fuse EDA AI Agent for standard-cell characterization, boosting a composite verification score from about 55% to ~95% across a representative benchmark. Refining agent skills and Model Context Protocol (MCP) tool descriptions based on verifier diagnostics also cut agent turns from 155 to 53 and reduced output tokens from ~81K to ~29K.

7.0

MODELS · 1 SOURCE · The Decoder

Anthropic adds dynamic workflows to Claude Managed Agents, enabling up to 1,000 parallel agents

Anthropic introduced dynamic workflows for Claude Managed Agents: a lead agent creates a plan, delegates tasks to sub-agents, and merges results, with up to 1,000 agents able to run in parallel per execution. The feature is enabled by selecting the "multiagent_20261001" agent type in Claude Code; Anthropic warns about high token use and recommends starting small, while its internal tests showed dynamic workflows found more bugs than single agents but broader cost-effectiveness is uncertain.

6.0

RESEARCH · 1 SOURCE · arXiv cs.AI

StoreBench: live-commerce RL environment to evaluate autonomous operator LLM agents

StoreBench is a live-commerce reinforcement-learning environment (arXiv:2610.10942v1) where an agent runs a mid-size online apparel store through the same 29 merchant tools a human operator would use. It tests long-horizon planning and economic judgment under continuous, stochastic conditions with calibrated pass thresholds and hardened rewards; the authors benchmark seven frontier LLMs across multiple scenarios (none match a scripted smart-triage heuristic on average), show human operators outperform models, report a substantial few-task post-training gain for Qwen3.5-27B using GRPO, and release five example training tasks, ten sample trajectories, and scoring/verification tooling while withholding the full environment to avoid contamination.

7.0