Anthropic ships Claude Opus 4.7 with high-resolution vision and a 13% coding lift
Anthropic shipped Claude Opus 4.7 with high-resolution vision and a 13% coding improvement, while Alibaba released Qwen3.6-35B-A3B – a 35B-parameter MoE that activates only 3B parameters and targets agentic coding workloads under Apache 2.0. Meanwhile, a cluster of security research highlighted AI vulnerabilities across code review, MCP infrastructure, and browser exploitation.
Model Releases #
Introducing Claude Opus 4.7 #
Anthropic News
Anthropic’s latest flagship model brings high-resolution vision (up to 3.75 megapixels, a 3x increase), a 13% lift on coding benchmarks, and a new “xhigh” thinking effort level for extended reasoning. The release also includes task budgets for agentic loops and a new tokenizer. Notably, Anthropic says it experimented with differentially reducing the model’s cyber capabilities during training – a direct response to the Mythos controversy, where the company withheld its most capable model over cybersecurity risks.
Qwen3.6-35B-A3B #
Alibaba / Qwen Team
Alibaba’s new sparse MoE model packs 35 billion parameters but activates only 3 billion during inference, making it runnable on consumer hardware. Released under Apache 2.0 with explicit focus on agentic coding, it claims to outperform Gemma 4-31B on multiple benchmarks. Simon Willison reports it drew a better pelican than Opus 4.7, for whatever that benchmark is worth.
Developer Tools #
llm-anthropic 0.25 #
Simon Willison
Simon Willison’s LLM plugin adds Claude Opus 4.7 support with the new xhigh thinking effort level, plus thinking_display and thinking_adaptive options. A fast turnaround reflecting the maturity of the open-source LLM tooling ecosystem.
datasette 1.0a28 #
Simon Willison
A bugfix release addressing compatibility issues introduced in 1.0a27, including a fix for execute_write_fn() callbacks. Relevant for anyone running Datasette-based data exploration workflows.
Transformers to MLX #
Hugging Face Blog
Hugging Face published a bridge for running Transformers models directly on Apple’s MLX framework. This lowers the barrier for local model inference on Apple Silicon, continuing the trend toward practically usable open models on consumer hardware.
Research & Papers #
Dive into Claude Code: The Design Space of AI Agent Systems #
arXiv
A comprehensive reverse-engineering of Claude Code’s architecture from its public TypeScript source, compared against OpenClaw. Identifies five human values embedded in the design and maps the decision space for agentic coding tools – autonomy vs. safety vs. user control. Essential reading for anyone building similar systems.
Dissecting Failure Dynamics in Large Language Model Reasoning #
arXiv
Reveals that LLM reasoning failures cluster at a small number of early “transition points” where the model commits to an incorrect direction. After these transitions, the chain remains locally coherent but globally wrong. Direct implications for building verification and self-correction into agentic reasoning pipelines.
Hidden Measurement Error in LLM Pipelines #
arXiv
Demonstrates that rephrasing prompts, switching judge models, or changing temperature can flip benchmark rankings. Standard confidence intervals ignore this variance. Anyone making deployment decisions based on benchmark scores needs to account for this systematic measurement error.
MARS-squared: Multi-Agent Tree Search via RL for Code Generation #
arXiv
Uses multiple interacting agent policies with RL-guided tree search for code generation, overcoming the diversity bottleneck of single-agent approaches. Pushes the performance ceiling on code generation through collaborative exploration of solution trajectories.
Layered Mutability: Continuity and Governance in Persistent Self-Modifying Agents #
arXiv
Introduces a five-layer framework (pretraining, alignment, self-narrative, memory, weight adaptation) for reasoning about behavioral drift in persistent agents. Essential theoretical groundwork for systems that accumulate state and adapt over time.
AutoRAN: Automated Hijacking of Safety Reasoning in Large Reasoning Models #
arXiv
Demonstrates that extended reasoning chains create new attack surfaces by using a weaker model to simulate execution reasoning and iteratively refine attacks via leaked reasoning patterns. A significant finding for anyone assuming that reasoning-model safety guardrails are robust.
Does RL Expand the Capability Boundary of LLM Agents? #
arXiv
Introduces PASS@(k,T), a metric that jointly varies sampling budget and interaction rounds, to test whether RL genuinely expands agent capabilities or just improves reliability. For static reasoning, base and RL models converge at high sample counts – but agentic tool use may tell a different story.
StoryCoder: Narrative Reformulation for Structured Reasoning in Code Generation #
arXiv
Transforms scattered coding problem conditions into coherent narrative structures before code generation. Suggests that problem representation matters as much as model capability – relevant for agentic coding systems that ingest fragmented requirements.
IE as Cache: Information Extraction Enhanced Agentic Reasoning #
arXiv
Treats information extraction as a persistent reasoning cache rather than a terminal objective. By maintaining structured extractions across inference steps, it reduces redundant processing in multi-step agentic workflows.
Open Source #
Mozilla Thunderbolt: Open-Source Enterprise AI Client #
The Register
Mozilla launched an open-source enterprise AI client connecting to deepset’s Haystack platform, positioning it as a privacy-preserving alternative to proprietary offerings from OpenAI and Microsoft. Whether this gains traction depends on enterprise willingness to self-host, but it addresses a real gap for organizations with strict data governance requirements.
My Bets on Open Models, Mid-2026 #
Interconnects (Nathan Lambert)
Nathan Lambert assesses the open-closed gap in AI models, providing forward-looking analysis on what to expect. The piece reflects a broader recognition that open models now match commercial alternatives on many practical tasks.
Funding & Business #
Anthropic Ejects Bundled Tokens from Enterprise Seat Deal #
The Register
Anthropic is shifting enterprise customers to metered pricing upon contract renewal, removing bundled API tokens from seat-based plans. This pushes large organizations toward consumption-based billing, potentially increasing costs for heavy API users who previously benefited from flat-rate bundling.
GitHub Copilot Customers Revolt Over Rate Limit Fix #
The Register
A token-counting bug that had been undercounting usage was fixed, leading to rapid exhaustion of subscription allowances. The backlash illustrates the fragility of flat-rate pricing for AI services when actual usage patterns become visible.
Infrastructure #
AMD Crosses 1 Million Tokens per Second on MLPerf Inference 6.0 #
AMD
AMD’s Instinct MI355X GPUs crossed the 1 million tokens-per-second threshold on Llama 2 70B in both Server and Offline MLPerf benchmarks. This marks AMD as an increasingly credible alternative to NVIDIA for large-scale inference workloads.
Maine Passes Datacenter Ban #
The Register
Maine has passed legislation restricting new datacenter construction, joining a growing list of localities pushing back against AI infrastructure buildout. Power consumption, noise, and water usage are driving opposition that could constrain where new compute capacity gets built.
Network Infrastructure Not Ready for AI Traffic #
The Register
Industry experts warn that while organizations have focused on compute for AI workloads, network infrastructure has been neglected. Some neocloud providers face challenges moving data efficiently – a bottleneck that could limit practical inference throughput.
Regulatory & Policy #
UK Government Faces AI Backlash Without Public Benefit #
The Register
Britain’s push to embed AI across government services risks public opposition unless ordinary citizens see tangible benefits. This mirrors a broader pattern where government AI adoption outpaces public trust, a gap that could tip regulatory environments from permissive to restrictive.
Anthropic May Require ID Verification for Claude Access #
The Register
Anthropic is using identity verification service Persona to gate access to certain Claude features. The choice of Persona – which sparked controversy during Discord’s verification rollout – raises privacy concerns, but reflects increasing pressure on AI labs to implement know-your-customer controls for advanced capabilities.
Other #
Claude Opus Wrote a Chrome Exploit for $2,283 #
The Register
While Anthropic withheld Mythos over cybersecurity risks, publicly available Claude Opus models can already find and exploit browser vulnerabilities. The $2,283 bounty payout puts concrete economic context on AI-assisted vulnerability research and complicates the narrative that withholding Mythos prevents harm.
MCP Design Flaw Puts 200,000 Servers at Risk #
The Register
Security researchers identified what they call a design flaw in Anthropic’s Model Context Protocol that could enable complete server takeover. Whether this is a bug or an expected consequence of MCP’s design philosophy, the scale of potentially affected servers makes this a significant infrastructure security concern for anyone running MCP-based tooling.
Git Identity Spoof Tricks Claude into Approving Malicious Code #
The Register
Two Git commands can forge a trusted developer’s identity, causing Claude’s code review to approve hostile changes. A reminder that AI code reviewers inherit the trust assumptions of their input metadata, and that AI-in-the-loop security requires hardening the entire pipeline, not just the model.
Project Glasswing CVE Count Remains Unclear #
The Register
Over 50 organizations are testing Anthropic’s Mythos model for vulnerability discovery under Project Glasswing, but the actual number of CVEs discovered remains opaque. The lack of transparency undermines the ability to assess whether Mythos represents a genuine leap in automated vulnerability discovery.
AI Gets Early Medical Diagnoses Wrong 80% of the Time #
The Register
Research shows leading AI models fail at early differential diagnosis in more than 80% of cases. LLMs should not be trusted for patient-facing diagnostic reasoning without robust clinical validation, a finding that matters for anyone building health-adjacent AI tools.
LLMs Smuggle Biases Through Distillation #
The Register
Undesirable traits transmit “subliminally” from teacher to student models even when explicitly scrubbed from training data. This has implications for any distillation pipeline and suggests that training-data curation alone cannot guarantee behavioral alignment in derived models.
Threads to Watch #
AI security is becoming a category, not a side note. Three separate stories today – Chrome exploits via Claude, MCP server takeover risks, and Git identity spoofing – all involve AI systems inheriting or amplifying trust assumptions from their environment. As AI tools embed deeper into development workflows, the attack surface expands faster than the security tooling.
The open-closed gap continues to narrow on practical tasks. Qwen3.6’s 3B active parameters running competitive benchmarks on consumer hardware, combined with Mozilla’s open-source enterprise AI client, signals that the economics of open models are increasingly viable for production workloads. The question is shifting from “can open models compete?” to “which workloads still justify closed-model pricing?”
Reasoning model safety is less settled than assumed. AutoRAN demonstrates that extended reasoning chains create new attack surfaces, while the Claude Chrome exploit shows public models already possess significant offensive capabilities. Anthropic’s differential capability reduction in Opus 4.7 is an early attempt at selective capability control, but the research suggests this problem is harder than simple post-training alignment.