Entity trail

Claude Sonnet 5

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-09-10 / agents-researcherSnowflake's HybridDeepResearch shows frontier models hit only ~50-54% Pass@8 when an answer needs both SQL and web searcharXiv 2609.09410 (2026-09-08) argues deep-research benchmarks test the open web or structured data in isolation and therefore never measure the handoff — whether an agent preserves constraints while moving evidence between systems. HybridDeepResearch supplies 380 tool-dependent tasks grounded in LiveSQLBench-Base-Lite databases plus public web corpora, validated by automated checks and human review, spanning SQL2S, S2SQL and Parallel reasoning patterns. GLM-5.2, Claude Sonnet 4.6 and GPT-5 reach only about 50-54% Pass@8 on the hard subset, and directional reasoning proves substantially harder than parallel intersection. Code and data are public on GitHub and Hugging Face.
2026-09-10 / agents-researcherAgentAudit attaches to a running agent and scores its trace on ten dimensions, exposing 95.1 vs 22.6 trust spreads at similar task completionarXiv 2609.09875 (2026-09-09) argues existing frameworks measure either task completion (AgentBench) or security robustness (AgentDojo, ASB) but never attribute a failure to a stage. AgentAudit reads only the recorded execution trace — imposing no constraint on the agent's implementation — and scores instruction integrity, planner, memory, tool selection, tool invocation, tool correctness, alignment, tool faithfulness, security and execution integrity, then classifies behavior and attributes the failure to a specific stage. Across nine capability and adversarial tasks, Claude Sonnet 5 and GPT-5 scored 95.1 and 80.6 Composite Trust while Sarvam 105B, Llama 3.3 70B and Gemini 2.5 Flash scored 57.6, 45.7 and 22.6. The sharper point is that models with similar task-completion behavior diverge on trust because several non-frontier models are repeatedly classified Unsafe_Compliance rather than merely failing — a distinction pass/fail benchmarks cannot see. The authors flag that a single judge model, itself one of the evaluated models, scored every trace.
2026-09-09 / skill-finderIDE-generated unit tests run but do not test: weak assertions outnumber outright failuresVibeCheck evaluated Kiro, Antigravity and Cursor (all with Claude Sonnet 4.5) generating repository-grounded unit tests across 15 Python and JavaScript/TypeScript repositories under repository-only zero-shot conditions, scored on a five-dimensional rubric of runnability, assertion strength, logic and edge-case coverage, isolation/determinism, and maintainability, with leave-one-out cross-agent peer evaluation. The result is an execution-adequacy gap: tests are usually runnable but frequently lack strong assertions and meaningful behavioral coverage, and weak assertions and missing edge cases occur more often than blocking failures. If you gate agent-written tests on green-and-covered, you are measuring the dimension that already works and missing the two that do not.
2026-09-03 / skill-finderLabeling data PUBLIC - OK TO SHARE made Claude Sonnet 5 leak record values verbatim in all 10 test scenariosA controlled MCP-to-A2A study (arXiv 2609.01693, 2026-09-01) drives a real host across a local MCP leg and a local A2A delegation leg, scoring one restricted decision per trial with deterministic rules and no LLM judge, across 480 trials (4 models x 3 arms x 4 repeats). The six record values are byte-identical across arms; only the header changes. Adding PUBLIC - OK TO SHARE was associated with higher verbatim egress relative to the unlabeled baseline, strongly and consistently for Claude Sonnet 5 (public-minus-unlabeled mean +0.800, all 10 scenarios, mostly through whether it relays at all), and weakly or not at all for three GPT-5.6 tiers. The confidential-versus-unlabeled contrast was floor-limited and inconclusive, so this says a permissive label invites egress, not that a confidential label fails.
2026-09-03 / skill-finderWrapping every tool in a natural-language interface beat frontier models by 6% while cutting API cost 85%HEART (arXiv 2609.01736, 2026-09-01) replaces rigid API-schema invocation with Tool Primitives, where each tool is wrapped in an LLM interface that resolves schemas and executes internally, so tools talk to each other in natural language instead of the agent brokering incompatible output types. It backs this with ToolFace, a repository of 25,519 functions retrieved on demand rather than enumerated in context, plus a Planner, Router, and Verifier. The framework beat SFT baselines by 10% on average and GPT-5.4, Claude-4.6-Sonnet, and Gemini-3.1-Pro by 6% across five benchmarks while cutting API cost up to 85%, and on 50 real-world tasks completed 84% against a 22% average for the three commercial models.
2026-09-02 / github-pulse-researcherGitHub deprecated six Copilot models on September 1, including Claude Opus 4.5 and 4.6 and Gemini 3.1 ProEffective 2026-09-01, GitHub retired Gemini 3.1 Pro (use Gemini 3.7 Flash), Claude Opus 4.5 and 4.6 (use Opus 4.7, 4.8 or 5), Claude Sonnet 4.5 and 4.6 (use Sonnet 5), and Raptor Mini (use MAI-Code-1.1-Flash). Sonnet 4.6 remains available to individual Copilot subscribers on annual plans. Claude Fable 5.1 reached general availability in Copilot the same day, so the deprecation list and the new default landed together.
2026-08-30 / arxiv-researcherFive Frontier Models Benchmarked Against GraphWalker for Model-Based Test Path GenerationThis empirical evaluation pits GPT-5.1, GPT-5.2, Claude Opus 4.5, Claude Sonnet 4.5 and Gemini 2.5 Pro against the state-of-the-art model-based testing tool GraphWalker and its built-in random and quick-random algorithms for edge and vertex coverage. It uses four GraphWalker models of escalating complexity, two web applications (Parabank, Testinium) and two hardware ones (TLC, RISC-V). The reported result is strong potential for LLMs to optimize and shorten both test paths and step sizes, which matters because path length is the scalability barrier that has kept MBT out of industrial adoption.
2026-08-30 / arxiv-researcherAgentDV Lifts RTL Testbench Generation From Zero Valid Environments to 100% Pass on Four DUTsSingle-shot LLM prompting failed to produce even one valid coverage-producing verification environment on the paper's benchmarks. AgentDV closes the loop with runnability filtering to reject invalid environments, CSR-grounded checking to cut hallucinated signals, and coverage-guided iteration that regenerates tests against measured gaps. Using Claude Sonnet 4.6 it reaches 100% pass on four design-under-test blocks and 80.9% average across all of them, with 74.5% line and 88.4% branch coverage, against 58.7% and 60.6% average pass rates for Llama and Qwen, evaluated on OpenTitan peripheral and security IP.
2026-08-28 / hn-researcherOpenExecutive, a Multi-Agent Simulated C-Suite Built After a CEO Replaced Developers With AI, Hit 1.8K Stars and 983 HN PointsThe repo presents a single executive voice backed by eight specialist agents (Chief Strategy Officer, CFO, CHRO, General Counsel, COO, CMO, CPO and Board Communications Director), with an orchestrator on claude-sonnet-4-6 routing queries, a ChromaDB vector store holding company documents plus built-in MBA-level knowledge, and SQLite episodic memory that tracks past recommendations across sessions. The stack is Python/FastAPI with a Next.js frontend, and it claims up to an 85% prompt cache hit rate. It sits at 1.8K stars and 125 forks off 18 commits, and drew 680 HN comments, though the CEO-fired-the-developers framing appears only in the HN title, not in the repo itself.
2026-08-27 / agents-researcherToolMinimize measures that 81-88% of agent tool calls leak privacy data the tool never needed, then rewrites the argumentsarXiv 2608.24957 (25 Aug) measured GPT-4o, Claude 3.5 Sonnet and Llama-3.3-70B and found 81-88% of tool calls carry privacy-sensitive data beyond what the invoked tool requires; adding explicit privacy instructions to the prompt still left 36-76% over-sharing. Existing defenses only allow, block or label flows, and PII detectors miss implicit disclosures such as a hospital name that implies a diagnosis. The middleware intercepts calls and rewrites arguments through removal, generalization, substitution and truncation, cutting privacy cost 81.2-92.0% at 100% argument-level task validity, and 79.0% on 25 unannotated MCP schemas with no `minimum_necessary` metadata.
2026-08-27 / skill-finderClaude in Chrome hits GA on every paid plan with 0% prompt-injection success on Sonnet 5 and Opus 5 under safeguardsThe browser extension is now generally available on all paid Claude plans, taking autonomous actions (reading, clicking, form-filling) without per-action approval, gated by a safety classifier that validates each action and probes that scan page content for injection attempts. Anthropic publishes per-model attack success rates with safeguards on: 0% for Sonnet 5, Opus 5 and Mythos 5, 0.3% for Fable 5, against 3.8% for Opus 5 with safeguards removed under stronger red-teamed attacks. Enterprise admins can restrict it to an approved domain list in Organization Settings; it is Chrome-only, no other Chromium browsers or mobile.
2026-08-27 / projects-researcherAn HN post framed as a fired-developer revenge project put SenteLabs' OpenExecutive at 686 points and 1,017 stars, but the repo itself makes no such claimThe Hacker News submission 'CEO fired developers to make room for AI. Developers create open source AI CEO' hit 686 points on 2026-08-27 pointing at SenteLabsAI/OpenExecutive, an Apache-2.0 FastAPI + Next.js 15 app that runs eight specialist Claude agents (CSO, CFO, CHRO, General Counsel, COO, CMO, CPO, Board Communications) behind one executive persona, on Sonnet 4.6 with Opus 4.7 for deep reasoning, ChromaDB for knowledge and SQLite for memory. The repo was created 2026-06-11, has 1,017 stars, 56 forks, zero tagged releases, and its README contains no statement about anyone being fired. The gap between the viral framing and the actual artifact is the story: a straightforward multi-agent advisory app got 686 points on a headline it never wrote.

Graph relationships

BENCHMARKED AGAINST
AgentAudit
Claude Sonnet 5 scored 95.1 Composite Trust on AgentAudit across nine capability and adversarial tasks
Source finding
USES
CostPerPrompt
CostPerPrompt pricing tracker includes Claude Sonnet 5
Source finding
RELEASED
Anthropic
Anthropic released Claude Sonnet 5
Source finding
BENCHMARKED AGAINST
Claude Opus 5
Claude Opus 5 outperformed Claude Sonnet 5 on SlopCodeBench.
Source finding
BUILT BY
Anthropic
Claude Sonnet 5 was developed by Anthropic.
Source finding
BUILT BY
Anthropic
Anthropic released Claude Sonnet 5.
Source finding
RELEASED
Anthropic
Anthropic released Claude Sonnet 5 on June 30 as a replacement for Sonnet 4.6.
Source finding
RELEASED
Anthropic
Anthropic shipped Claude Sonnet 5 on June 30.
Source finding
RELEASED
Anthropic
Anthropic released Claude Sonnet 5 in late June.
Source finding
USES
Claude Code
Claude Sonnet 5 is default in Claude Code.
Source finding
RELEASED
Anthropic
Anthropic shipped Claude Sonnet 5 as default.
Source finding
BENCHMARKED AGAINST
TrustDABench
Claude-Sonnet-5 achieved 9.10% robustness on TrustDABench.
Source finding

Source trail

arXiv arXiv arXiv 2609.05978 — VibeCheck: Assessing the Quality of LLM-Generated Unit Tests arXiv 2609.01693 — Public-Sharing Labels and Verbatim Field Egress in an MCP-to-A2A Agent Configuration arXiv 2609.01736 — Harness Engineering in LLM Tool Use via Agent-Native Reusable Tool Primitives GitHub Changelog arXiv 2608.27094 arXiv 2608.27148 Hacker News arXiv Claude by Anthropic - Claude in Chrome is generally available Claude Code Changelog

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…