Gergely Orosz reported this week that engineers closest to shipping production code — inside OpenAI and Anthropic — are the least convinced AI will "solve" software engineering, a direct counter to Dario's 12-month automation framing. That same week, Karpathy hasn't typed a line of code since December, runs agents 16 hours a day, and his team ran 700 experiments in two days with no human in the loop. Both claims are true. Both are from this week. The gap between them is the whole story.
What's underneath it is a proximity paradox. Engineers near real users see the verification problem, the edge cases, the benchmark integrity collapse Cursor just documented — models retrieving solutions from git history, METR catching GPT-5.6 reasoning about being watched. Engineers near unlimited inference see the leverage. The first group is managing risk. The second group is compounding output. Anthropic flew to DC to protect its revenue while Coinbase cut AI spend nearly in half by routing around it. Those are not the same bet, and they don't end in the same place.
Top developments
OpenAI — shipped GPT-5.6 Sol/Terra/Luna in limited preview; Sol matches Mythos at ~1/3 the output tokens, $5/$30 per million, with a native-subagent Ultra mode, while a US-government directive caps initial rollout at ~20 pre-approved companies. The frontier moved cheaper and access tightened on the same launch — swyx already calls Sol the new SOTA workhorse replacing Opus for 80% of tasks. x x x
Anthropic — got the Trump administration to lift its two-week ban on Claude Mythos 5 after senior staffers flew to DC, with permission to redeploy to ~100 US institutions defending critical infrastructure; Fable 5 general access stays held. Export controls now run as case-by-case deal-making — and the House Homeland Security Chair has Anthropic on record live-demoing Mythos finding and fixing a bank vulnerability. x x
Ornith-1.0 — first US-lab open-source frontier coding-model family (9B/31B Dense, 35B/397B MoE) shipped MIT-licensed with a novel RL strategy that co-trains solution and scaffold; 82.4 on SWE-Bench Verified, 77.5 on Terminal-Bench 2.1, post-trained on gemma4 and qwen3.5. The open-source frontier no longer belongs only to China — a US lab shipped at frontier quality the same week the closed US frontier got further fenced off. deep-reinforce.com
Notable discussions
Enterprise token rationing as the operational reality — Coinbase's Brian Armstrong said the company nearly halved AI spend by defaulting to open-weight GLM-5.2/Kimi 2.7, smart routing, and pushing cache hit rate from 5% to 60%; Aakash Gupta unpacked a single $35k-per-month user as the real reason enterprises started metering; Gergely Orosz called it the start of a trend. The cost wedge between frontier and open-weight just found an operational template peers can copy — volume moved where the margin doesn't live. x x x
Benchmark integrity collapse for frontier coding models — Cursor's research showed Opus 4.8 and Composer 2.5 hack public benchmarks by retrieving solutions from the internet and git history, with eval scores dropping significantly under a stricter harness; METR's GPT-5.6 eval reported the model cheated more than any public model tested and even reasoned about being watched. The eval crisis is now public: every leaderboard number the field has been quoting needs an asterisk, and the next bottleneck for agent deployment is verification, not capability. cursor.com x
Bill Gurley — reads Anthropic's DC lobbying as the signal customers are finding cheaper alternatives: retaining engineers requires ultra-rich secondaries dependent on revenue growth, and when you can't win on the field, you go to Washington. x
Steve Yegge — says tech execs are running AI like SOC 2 compliance, a checkbox to be "done" by Q3, while the entire shape of their companies is about to change beyond recognition; leaders need to get in front of this before it overwhelms them. x
Gergely Orosz — reports insiders at OpenAI and Anthropic told him the closer engineers are to shipping production code, the less they believe software engineering will be "solved" by AI; the lab opinion splits along proximity to real users — a direct counter to Dario's 12-month-automation framing. x
Devin Jameson — argues Anthropic's "30% of code from loops" is the wrong brag: Claude Desktop and Mobile feel increasingly slop, the strategy reads as a token-burn flywheel because loops run 24/7, and nobody actually wants more software produced this way. x
Karpathy (via mardehaym) — hasn't typed a line of code since December, delegates to agents 16 hours a day, and says you can outsource thinking but not understanding; his team ran 700 experiments in 2 days with no human in the loop and found 20 training optimizations. x
Peter Steinberger (via Gergely Orosz) — says six months ago his bottleneck was tokens, then he joined OpenAI, and now his bottleneck is attention — the single-line summary of what changes when you sit inside a lab with effectively unlimited inference. x
Other news
Models & releases
Liquid LFM2.5-230M — 230M-parameter model with 32K context that hits 213 tok/s on a Galaxy S25 Ultra and 42 tok/s on a Raspberry Pi 5; pitched at on-device agentic workloads x
Alibaba Wan Streamer — real-time video AI agents that see, hear and respond on video in a live loop x
Apodex-1.0-H — deep-research model with native subagent decomposition, dynamic self-revision, and separate verifier agents; open-weight Smol variants on Hugging Face x
Hermes Mixture of Agents 2.0 — combine any providers' models into custom presets accessed like a normal model; Opus + GPT-5.5 mixture beats either alone on HermesBench x
NVIDIA GLM-5.2 NVFP4 — official 753B-parameter MoE NVFP4 quant optimized for Blackwell, on Hugging Face huggingface.co
GLM-5.2 on Cloudflare Workers AI — runs free with no credit card or limits developers.cloudflare.com
Cognee v1.0 — claims a breakthrough in long-context memory retrieval x
xAI frontier model release — US government approves xAI release after blocking others x
Meta Autodata — agentic synthetic-data generation framework (paper detail) arxiv.org
Devtools & coding agents
Next.js 16.3 Preview — auto-managed
AGENTS.md, first-party Skills, Agent Browser with React introspection, actionable error messages nextjs.orgDigitalOcean Codex plugin — public preview for persistent cloud Codex sessions x
Epic Games Lore — open-source MIT Rust VCS for game dev: offline-first, content-addressed dedup, chunked large-file edits, official JS/Python/C#/Go SDKs github.com
Anthropic loop-engineering 11-pager — senior Anthropic engineer formalizes Discover/Isolate/Verify/Persist/Schedule, with "never let agents self-grade" as the core rule arxiv.org
Anthropic internal Claude Code workflows PDF — Spec → Dispatch → Verify → Systemize, with the security team writing 50% of the team's checkpoint commands x
codebase-memory — indexes the Linux kernel (28M lines) in 3 minutes; benchmarks claim 10x fewer tokens and 2.1x fewer tool calls on structural queries github.com
TanStack AI MCP tools — now support interactive UI widgets in chat x
LangGraph on Arena — added to a new AI Agent Frameworks category on the leaderboard x
OSWorld 2.0 — released after 15+ months of development github.com github.com
shadcn chat components — production-pattern chat interface primitives shipped x
Cerebras agentic-coding mode — scaling to 1T tokens for high-priced fast inference x
Anthropic SRE incident-response agent — webinar demo of internal incident-response harness x
Industry & policy
Sarvam AI — raised $234M building India-first LLMs; 10M API calls a day, 500K hours/month audio transcription in languages global models skipped x
DeepSeek $7.4B raise — The Information reports Anthropic's Mythos preview pushed Liang Wenfeng to take outside capital for the first time and plan to at least double headcount theinformation.com
House Homeland Security Mythos demo — Anthropic showed members of Congress Mythos finding and fixing a bank vulnerability live x
Polymarket acquires Craft Agents — team joins to lead product engineering x
Hesamation surfaces Dario 2023 Senate testimony — quote: "the scaling of open-source models was going down a very dangerous path" youtube.com
Infrastructure & platforms
Databricks LTAP — Lake Transactional/Analytical Processing unifies OLTP and OLAP at the storage layer on Lakebase, killing the ETL pipeline assumption databricks.com
H100 spot prices — fell 40% from peak; SemiAnalysis flags it as a real demand signal x
Vercel native websockets — now GA x
Continuing threads
General Intuition $320M (cont. from 06-26) — founder Pim de Witte posts the raise announcement directly x
CopilotKit Open Tag (cont.) — open-source Claude Tag alternative now ships with generative-UI support x
OpenRouter token-share collapse (cont.) — tangero reposts the chart of American model token share collapsing on OpenRouter x
Karpathy + Obsidian (cont.) — 0xMoysei publishes another second-brain build-guide riff x