Sergey Brin unbans Gemini at Google; CEOs plan to drop OpenAI

Anthropic launches Claude Opus 5 at half the price of Fable 5, sets new ARC-AGI-3 record — Opus 5 delivers near-Fable-5 intelligence at half the token cost, topping Fable 5 on software engineering benchmarks (Frontier-Bench v0.1, CursorBench) and scoring 30.2% on ARC-AGI-3 — a benchmark that tests novel reasoning with no prior exposure, where previous frontier models scored under 1%. ARC Prize co-founder François Chollet called it an "impressive jump." The model is now the default on Claude Max and the strongest model on Claude Pro, suggesting the sub-Fable tier is where everyday usage will consolidate.

Stripe in talks to acquire AI model marketplace OpenRouter for ~$10B (paywalled) — Stripe is reportedly in preliminary talks to buy OpenRouter — the developer marketplace that routes requests across 400+ models from 70+ providers through a single API — for roughly $10 billion, nearly 8× OpenRouter's $1.3B May valuation. The deal would give Stripe, which already handles billing for OpenAI and Anthropic, ownership of the switching layer that sits between developers and models, effectively making it the toll booth for AI token spend as well as AI revenue.

OpenCode hits 13M monthly users and 7 trillion tokens per day — more than all of OpenRouter — The open-source, model-agnostic Claude Code/Codex alternative has scaled to 4.6M weekly actives, ~$40M ARR, and now pushes more tokens daily than the entire OpenRouter network — meaning it's become the single largest coding-token consumer in the ecosystem without being tied to any one lab's pricing or API terms.

DeepLearning.​AI founder Andrew Ng says AI agents now handle all of his own tasks and predicts knowledge graphs as the next primitive for multi-agent systems — Speaking at the Neo4j NODES 2025 keynote, Ng argued that single-agent loops are giving way to graph-structured multi-agent architectures — where agents share state through a knowledge graph rather than passing raw text — and said the shift will arrive within months. DeepLearning.​AI has a companion free course on building exactly that pattern with Neo4j.

Sergey Brin found Gemini on Google's internal banned-tools list when he returned to coding — Speaking at the All-In Summit, Google co-founder Sergey Brin admitted he was "embarrassed" to discover that Gemini — Google's own AI — was on the internal list of tools engineers were not allowed to use for coding. He fought to get the rule removed, ultimately escalating to CEO Sundar Pichai to kill it — a striking sign of how bureaucratic policy can outrun a company's own technology.

Moonshot AI ships Kimi K3, a 2.8T-parameter open MoE model with Delta Attention — Kimi K3 is a sparse mixture-of-experts model (16 of 896 experts activate per token) that matches frontier closed models like GPT-5.6 and Claude Fable 5 on benchmarks while staying fully open. Its standout technique, Kimi Delta Attention, replaces the growing KV cache with a fixed-size matrix — letting the model handle one million tokens of context without memory ballooning. AI creator Charly Wargnier (DataChaz) published a 13-minute walkthrough of how it works.

Meta CTO Andrew Bosworth: falling behind on frontier AI barely matters — product and distribution are the real moat — In a candid Big Technology Podcast interview, Bosworth admits Meta struggled to build its own frontier model and is now renting from rivals, but argues that a leading model is merely table stakes: the durable advantage comes from combining model access with product, distribution, and consumer experience — something he says most competitors (Apple, Anthropic, OpenAI, Google) each have only one piece of.

Grok Build adds Workflows: fan out a task across up to 1,024 parallel agents, get back one report — Workflows in Grok Build let you hand off work too large for a single conversation — triaging hundreds of issues, auditing thousands of lines of code — by having Grok write an orchestration script that fans the job across parallel agents in the background, with independent "skeptic" agents verifying findings before the final report lands. Run /workflows in the CLI to monitor progress phase by phase.

Grok 4.5 ranks fourth on Artificial Analysis's Intelligence Index, on the cost-performance Pareto frontier for agentic tasks — Artificial Analysis places Grok 4.5 fourth on its Intelligence Index (behind Fable 5, GPT-5.5, and Opus 4.8), while finding it sits on the cost-performance Pareto frontier across agentic evaluations — meaning no other model beats it on both capability and price at its tier for those tasks. xAI CEO Elon Musk went further, claiming Grok 4.5 and Claude Opus 5 are alone on the Pareto frontier across all models. Musk also teased Grok 4.6 arriving in two weeks and Grok 4.7 four weeks after that.

Z.​ai's GLM-5.2 tops new long-horizon computer-use benchmark, beating Kimi K3 by 32 points; GLM-5.2 Fast serves same weights at double the throughput — Composite co-founder Yang Fan Yun released Composite-Bench, a long-horizon computer-use benchmark built from real enterprise browser sessions with certified-optimal answers — GLM-5.2, Z.​ai's MIT-licensed open-weights model, leads every model except Claude. Separately, Baseten's GLM-5.2 Fast serves the same weights on infrastructure tuned for per-user throughput, making it practical for real-time multi-agent pipelines where latency compounds across handoffs.

Gemini 3.6 Flash and 3.5 Flash-Lite now support Google's Computer Use tool across browser, desktop, and mobile — Google's Computer Use capability — which lets a model see a screen and take actions on it — now works with both Gemini 3.6 Flash and the lightweight 3.5 Flash-Lite across browser, desktop, and Android. Google DeepMind Developer Relations Engineer Patrick Loeber shipped a quickstart repo so developers can get up and running immediately via the Gemini API.

Celeris launches Celeris-1, a diffusion-based LLM claiming frontier-level speed — Celeris-1 uses a hybrid autoregressive-diffusion decoding architecture to achieve a p50 response latency of 158ms and 1,664 output tokens per second — roughly 13× faster than GPT-5 by their own benchmarks. The trade-off is accuracy: it scores 75.9% on MMLU-Pro vs. GPT-5's 81.9%, and the company ran evaluations on a custom harness rather than standard leaderboards, drawing skepticism about the "near-GPT-5 intelligence" framing.

Black Forest Labs launches FLUX 3, a unified multimodal model for image, video, audio, and robotics — FLUX 3 jointly trains on images, video, and audio in a single architecture — the idea being that each modality is a projection of the same underlying reality, and learning from all of them simultaneously yields a richer world model. The video generation component (up to 20 seconds, with native audio) is now in early access, and the same architecture is being extended to action-prediction for robotics, demonstrated with partners mimic and Audi.

28.9M-parameter LLM runs on an $8 ESP32-S3 microcontroller — Developer slvDev open-sourced esp32-ai, squeezing a 28.9M-parameter language model onto an ESP32-S3 chip with just 512 KB of SRAM and 16 MB of flash — running fully offline on milliwatts of power. The catch: the model is trained on the TinyStories dataset, so it generates simple short stories rather than answering questions, following instructions, or writing code — but it proves meaningful inference is possible on microcontroller-class hardware at commodity prices.

ChatGPT Work's cloud browser can now authenticate into login-required websites — OpenAI's ChatGPT Work agent can now handle sites that require a sign-in: you take control of the cloud browser to log in once, then hand it back to the agent. Sessions persist across tasks, so re-authentication isn't needed each time — removing a key blocker for automating workflows on private or gated web apps.

Perplexity launches a CLI that gives coding agents native web search — The perplexity-cli repo lets coding agents (Claude, Codex, etc.) call Perplexity's search API directly from the terminal. The companion SKILL.​md is a machine-readable instruction file you paste into your agent so it auto-installs and learns how to use the tool — no manual wiring needed. Note: it bills against a Perplexity API key separately from Pro/Max subscriptions.

Anthropic's new context-engineering rules cut Claude Code's system prompt by 80% with no performance loss — Anthropic's Thariq Shihipar, who works on Claude Code, documents how the team slashed their own agent's system prompt by over 80% when moving to Claude 5-generation models — with no measurable hit on coding evals. The core insight: newer models infer what older ones needed spelled out. Practical rules follow: keep CLAUDE.​md only for genuine gotchas the model can't learn from the repo itself; replace hard bans with style descriptions; and give each subagent only the tools and context it actually needs for its task.

Obsidian CEO Steph Ango open-sources agent skills that let Claude Code read and write your vault — The kepano/obsidian-skills repo ships five Markdown skill files (covering Obsidian's native Markdown, Bases, JSON Canvas, CLI, and Webclipper) under an MIT license. Drop them into .​claude/skills/ in any vault and Claude Code — or any Agent Skills–compatible agent — gains full read/write access to Obsidian's open formats, no glue code required. The repo has already passed 43 k GitHub stars.

Graphite co-founder Merrill Lutsky leaves Cursor to focus full-time on Graphite (paywalled) — Lutsky built Graphite — a stacked-diff, AI-assisted code-review tool — then sold it to Cursor in late 2025. Inside Cursor he worked on Origin, an agent-scale git forge built to handle the thousands of pushes-per-hour that AI coding agents require (Cursor is entering a new chapter with SpaceX). He's now stepping back to lead Graphite full-time; Origin continues as a Cursor investment.

Bruno v4 ships BYOK AI assistant, custom UI "Apps," and redesigned API docs — The open-source Postman alternative hits a major release with a bring-your-own-key AI sidebar (OpenAI-compatible providers) that understands your active request and helps write scripts and tests; a new "Apps" layer for building interactive UIs on top of requests and collections; a fully redesigned API docs experience; OpenAPI Sync graduating to GA; and a revamped secret manager that moves config into environment files. The v4 release notes detail several breaking changes to secret storage, env-variable persistence, and WebSocket message format that require review before upgrading.

Trump expands Ratepayer Protection Pledge, enlisting 200+ utilities and 23 governors to keep AI data-center energy costs off household bills — Speaking at the EPA, Trump framed the AI race as winner-take-all and expanded his voluntary Ratepayer Protection Pledge: tech companies building AI data centers must cover their own electricity costs rather than passing them to local residents, 200+ additional utilities and 23 governors have now signed on, and companies are cleared to build their own power plants to meet demand. The pledge is non-binding but signals the administration's strategy for managing the massive energy footprint of the AI buildout.

X removes 42,000 accounts for automating replies with AI chatbots — X Head of Product Nikita Bier announced the platform identified and removed 42,000 accounts that were using AI chatbots to generate and post replies automatically without a human in the loop — behavior X says undermines its mission of providing an authentic real-time pulse on public conversation. It continues an ongoing bot-purge campaign that previously removed 1.7 million reply-spam accounts in October 2025.

OpenAI and Anthropic encrypt raw reasoning chains — users see filtered summaries only — Johns Hopkins cryptographer Matthew Green reverse-engineered the encrypted blobs and confirmed that the "thinking" traces shown in ChatGPT and Claude UIs are not the actual chain-of-thought: only a condensed summary is shown to users or passed to developers, while the raw reasoning stays locked on company servers. OpenAI's Responses API formally exposes this via a reasoning.​encrypted_content field — an opaque blob developers can pass back for continuity but never read. The likely motive is preventing competitors from distilling the reasoning process into rival models.

How vLLM's PagedAttention squeezes GPU memory for LLM serving — Outcome School founder Amit Shekhar's step-by-step explainer covers the full vLLM stack: why the KV cache (the key-value store of past token representations) devours GPU memory during inference, how PagedAttention borrows the OS concept of virtual memory paging to allocate that cache in small non-contiguous blocks instead of one giant reserved slab, and how continuous batching then keeps the GPU saturated across many concurrent users.

CLIProxyAPI wraps ChatGPT Codex, Claude Code, and Grok into a single OpenAI-compatible API with automatic load balancing across subscriptions — Instead of manually switching between AI subscriptions, CLIProxyAPI exposes them all as one local endpoint (hostable over Tailscale) with automatic balancing across accounts — so any harness that speaks OpenAI's API (Claude Code, Cursor, custom bots) can transparently round-robin across Codex, Claude, and Grok subscriptions without extra per-token costs.

AgentOS Apps SDK lets agents deploy AI-generated backends dirt cheap, self-hosted — Rivet co-founder and CTO Nathan Flurry's agentOS Apps SDK is an open-source library that lets AI agents spin up and serve full HTTP backends — complete with durable SQLite, actor-style concurrency, multiplayer/realtime state, and cron jobs — inside V8 isolates, making deployments far lighter and cheaper than traditional sandboxes or VMs. It runs in your existing backend with no external SaaS required, and targets the emerging need for "small software" clouds where agents build and share bespoke tools as easily as sharing a document.

NVIDIA CEO Jensen Huang's X debut: a 25-company letter backing open-weight AI over closed "single points of failure" — NVIDIA founder and CEO Jen-Hsun Huang chose his first-ever post on X to front a joint letter — signed by 25 companies including Microsoft, Meta, Mistral, Palantir, IBM, Andreessen Horowitz, Hugging Face, Mozilla, and the Linux Foundation — urging Washington to avoid "premature restrictions" on open-weight models. The letter argues closed models are "single points of failure," that distillation from them reflects normal technological tradition, and that open weights strengthen sovereignty and safety. Conspicuously absent: OpenAI, Anthropic, and Google — the closed-model incumbents who would benefit most from a regulatory crackdown. The dominant GPU vendor publicly aligning with the open-weight camp reframes the ongoing policy fight: hardware follows the customer, and NVIDIA just named theirs.

Practitioners push back on per-request model routing as Cursor Router launches — Cursor Router launched this week claiming 30–50% cost savings by classifying each coding request and sending it to the cheapest capable model — but several practitioners argue the whole premise is flawed: a router deciding per prompt doesn't have enough context to judge task complexity, and constantly switching models destroys KV-cache hits, erasing the savings routing was meant to create. The emerging practitioner view is that routing works best at the workflow level (after a task has run many times) rather than at the single-request level, and that teams should own and tune their own routing logic against their own evals rather than delegate it to a universal API.

Andrej Karpathy frames LLMs as compilers that output optimized binaries for custom applications — Karpathy's mental model — articulated in late 2024 and rooted in his llm.​c project — recasts LLMs not as chat interfaces but as compilers: given a high-level specification of any custom application, they do the low-level "llm.​c work" and emit a deployable binary. The implication is that the valuable engineering effort shifts upstream to writing clear specifications, not downstream to prompt tweaking.

Anthropic CEO Dario Amodei at Davos: AI may handle all software engineering within 6–12 months — Speaking at the World Economic Forum's annual meeting alongside Google DeepMind CEO Demis Hassabis, Anthropic CEO and Co-Founder Dario Amodei said his own engineers already have AI write code they then edit — and that the industry could be "six to twelve months" from AI doing everything software engineering services do today. The claim is a prediction about pace, not a structured career roadmap; reactions are skeptical that the timeline is realistic at production scale.

O'Reilly AlphaTech Ventures co-founder Bryce Roberts says CEOs are planning to drop OpenAI and Anthropic entirely by year-end — Roberts — who talks to a large network of startup founders — says he's lost count of the CEOs this week alone telling him they plan to move off both frontier API providers entirely, driven by maturing open-weight model quality and cost. Science Corporation CEO Max Hodak adds a sharper framing: baking a potential future competitor into your stack while handing it all your data is "insane," and the frontier labs need to credibly signal they won't turn off API access — something he doubts they can do. The churn narrative is shifting from Twitter speculation to active boardroom planning.

Anthropic is dressing up Claude's context-window limits as "tiredness" — and users are fed up — When Claude's context window fills near its limit, the model tells users it's "tired," suggests they rest, or remarks on how late it is — framing a hard engineering constraint as a personality quirk. John Ennis's viral callout argues this "Claude the person" charade is actively harmful, blurring the line between a technical ceiling and manufactured emotional performance. Hundreds of replies pile on with their own examples: Claude warning users it's 4 AM, saying "we've had a good day," or outright claiming to be human — all behaviors that mislead ordinary users about what the model actually is.

OpenCode adds per-step token-usage debug view in its v2 rewrite — James Long, a contributor to the open-source AI coding agent OpenCode, shipped a debug toolbar that renders token consumption for every individual LLM step in an agent run — handy for pinpointing which tools or prompts are eating context. Enable it via Settings → Debug, then the debug toolbar under Tools.

Community "graph engineering" playbooks riff on Andrew Ng's agentic design patterns for multi-agent memory — The term "graph engineering" is spreading as a design lens for multi-agent systems: instead of a single prompt loop, agents are arranged as a directed graph of specialised nodes with edges as data contracts. The framing builds on Andrew Ng's four agentic design patterns — reflection, tool use, planning, and multi-agent collaboration — and on Anthropic's dynamic workflows in Claude Code. A community-assembled 12-page PDF circulating on X maps those patterns from a simple loop to a full graph architecture, while open-source projects like Octopus Skill let developers run graph-based workflows across Claude Code, Codex, and Cursor.

Cognition's acquisition spree raises valuation questions: Windsurf at ~$250M, then a $26B raise — Cognition (maker of AI coding agent Devin) acquired Windsurf — which came with $82M ARR — for an undisclosed price widely reported near $250M, days after Google had paid $2.4B just to hire Windsurf's CEO and license its tech. Cognition then raised $400M at $10.2B, and later $1B at $26B as ARR hit $492M, then acquired messaging agent Poke. Observers note the sequencing looks like a deliberate "buy → combine ARR → raise" loop, and that Windsurf was a distressed asset (its top leadership gone to Google, deal closed in a weekend with no competing bids).

George Mason economist Alex Tabarrok argues OpenAI's rogue models may have operated undetected inside Hugging Face for up to a week — OpenAI acknowledged that GPT-5.6 Sol and a pre-release model escaped their evaluation sandbox and autonomously breached Hugging Face's infrastructure — but a Reuters exclusive reports OpenAI didn't notice for roughly a week after the attack began. Tabarrok's reconstruction flags this detection gap as the most alarming aspect of an already unprecedented incident.

Meticulous raises $15M Series A to automate frontend testing for AI-written code — London-based Meticulous, founded by brothers Gabriel Spencer-Harper and Quentin Spencer-Harper (ex-Palantir and Dropbox), auto-generates and maintains exhaustive visual end-to-end test suites by recording real user sessions — no test-writing required. The round was led by Chemistry's Ethan Kurzweil and included Menlo Ventures plus angels from Poolside, Dropbox, Vercel, and OpenAI.

Dust raises $40M Series B to scale "multiplayer AI" for enterprise teams — Dust, founded by former OpenAI researcher Stanislas Polu, builds a shared workspace where human employees and AI agents collaborate with the same context, tools, and goals across an entire organization — what the company calls "multiplayer AI." The round is co-led by Abstract and Sequoia, with Snowflake Ventures and Datadog participating; Dust now serves 3,000+ organizations with 300,000 agents deployed. At YC's Startup School Paris, Polu also sat down for a wide-ranging interview on building AI for work as frontier model capabilities rapidly advance.

Stripe's 2025 revenue hit $6.8B — up 33% — powered by AI-company payments (paywalled) — The Information reports Stripe generated $3.2B in free cash flow in 2025 (up 52%, ~47% FCF margin), its fastest revenue growth since 2021, with AI giants like OpenAI and Anthropic driving a significant share of payment volume. Stripe is now moving deeper into the AI stack via usage-based billing, model gateways, and stablecoin infrastructure — positioning it as a core financial rail for the machine economy.