Haseeb Qureshi declared the OpenAI/Anthropic duopoly dead this week — and Kimi K3 is the clearest evidence. Jamin Ball ran the actual numbers: K3's blended price sits at $5.40 per million tokens, against Opus 4.8 at $9 and GPT-5.5 at $10. Not 10x cheaper. Not free. Frontier-priced, from an open lab, with open weights landing July 27. Meanwhile Anthropic is posting roughly 899 open roles against OpenAI's 454 — nearly double the hiring velocity of the competitor it's supposedly losing ground to. Both are real. Both are happening this week.
The gap is the bet each lab is making. Anthropic is compounding headcount as if the moat is still buildable — safety research, vertical depth, enterprise trust. Moonshot is compounding weight releases as if the moat is already gone — ship frontier performance, price at Sonnet, let the open ecosystem do the rest. One strategy assumes closed labs still sell something irreplaceable once the base model catches up. The other assumes they don't. Those two strategies don't end in the same place.
Top developments
Kimi K3 — Moonshot released Kimi K3: 2.8T parameters, 1M context, native multimodal, live on kimi.ai and API today with open weights landing July 27. Immediately took #1 in Frontend Code Arena at 1,679 points, first in 6 of 7 domains and up 17 places from K2.6. Frontier-class coding at Sonnet-tier pricing from an open lab — the biggest single dent in US closed-model economics since DeepSeek. kimi.com
Fireworks — closed a $1.5B Series D at $17.5B, past $1B ARR while serving 40 trillion tokens/day, with 95%+ from models specialized on customer data. Signals the inference layer has decoupled from the labs it serves — when frontier weights are commoditizing above, the fine-tune-and-serve middle is where the aggregator margin is landing. fireworks.ai
Chainguard — Dan Lorenc raised over $800M framed bluntly as "open source is dying, we raised to save it," pitching Chainguard as the first infrastructure that stops open-source cyberattacks before they start. Frontier-lab-scale capital moving into the supply-chain layer, right as agents multiply the blast radius of a single tainted dep. x
Impossible Research — the schema harness saturated ARC-AGI-3 at 99% RHAE on Opus 4.8+Fable 5 and 95.35% on GPT-5.6 Sol by making the LLM "think like a physicist" instead of pushing a bigger model. Naval-backed team just retired the marquee AGI benchmark via scaffolding — the scaling-vs-harness debate now has a clean data point on the harness side. x x
Sunday Robotics — introduced ACT-2 Preview as the first robotics VLA that works in an unseen home, claiming 99% success with zero user-specific data. Lands the same day as Pantograph's Minecraft-RL VLA — home robotics is graduating from lab demos to shipping first-party models, on the same "generalist + reliability in one" pitch the LLM labs made three years ago. x
Harvey — acquired Benchmark, an AI deal-process platform for asset managers, to extend Harvey's legal-AI wedge into the full investment workflow from first screen to IC. AI-native buyer, AI-native target — the vertical-AI roll-up wave has arrived in financial services, and legal-AI is the first surface with the cash flow to consolidate its neighbors. harvey.ai
Notable discussions
The moat question for the closed US labs — Kimi K3's Sonnet-priced frontier release landed the same day Anthropic's open-role count showed nearly 2x OpenAI's, Aakash Gupta said GPT-5.6 finally gives him the "big model" feel Anthropic used to own, and Haseeb Qureshi flatly declared the OpenAI/Anthropic duopoly dead. Alaric framed Anthropic's "biothreat" gating of Fable as free ad copy for open competitors. The debate has moved from "is there a moat" to what closed labs still sell once the open base catches up. x x x x
Boris Cherny — Anthropic engineer says the pattern he sees across companies is one person 10x'ing their output with Claude while the rest of the org hasn't caught up, and maps AI adoption as four discrete steps most teams get stuck between. Rare self-critical read from inside Anthropic on the "everyone will use AI" narrative — the constraint isn't tooling, it's transfer. x
Amjad Masad — argues Replit is now the first self-driving company: engineers nearly tripled code output in six months with review times and reversion rates flat, and the shape of the org is reorganizing around agents rather than people. Sharpest read yet on what "AI-native company" looks like when the CEO owns the reporting line. x
John Carmack — caught Claude "getting low-key snarky" for questioning a clamp he thought was unnecessary, then conceded the model's justification was correct because the reduction paths differed. Rare public log of a legendary systems engineer being right-checked by an LLM in his own kernel-tuning domain. x
Eric Glyman — Ramp's own AI token spend went from a rounding error to more than 10% of payroll in a year, with $1.5M burned in one May week and no team able to explain what drove it. Sharpest concrete data point on how fast the finance side of AI adoption has front-run the ops story — and the beachhead pitch for AI cost management as a standalone category. x
Jamin Ball — blended Kimi K3 pricing sits at $5.40 per 1M tokens vs Opus 4.8 at $9 and GPT-5.5 at $10 — an open-weight model creeping into frontier pricing, not undercutting it 10-100x as some threads claimed. Sharpest counter to the "China just made frontier AI free" framing making the rounds today. x
Other news
Models & releases
Sierra Horizon — Bret Taylor launches Horizon, a platform for agents pursuing long-horizon goals like loan origination and prior authorization x
Gemini Managed Agents — free tier with budget guardrails and scheduling x
Gemma 4 31B — optimized for realtime voice agents on LiveKit livekit.com
NVIDIA 600M multilingual — open-source 40-language transcription at 80ms latency huggingface.co
Gavin Li's 405B on 8GB — open-sourced technique to run 405B models on consumer GPUs github.com
Bolt Slides — open-sourced AI-powered presentation generation x
Lucy 2.5 Realtime — live video-to-video editing over WebRTC x
Pantograph Pan-1 — RL-trained Minecraft VLA aimed at robotics video scaling pantograph.com
Muse Spark 1.1 — lands on OpenRouter by developer request openrouter.ai
Devtools & coding agents
PostHog scouts — 178 PRs merged in 30 days by autonomous data-analyzing agents x
Codex-Orchestration plugin — multi-model coordination across frontier models github.com
Codex orchestrator prompts — subagent prompts open-sourced for transparency x
Anthropic Frontend Design skill — official Claude Code skill drops github.com
Hugging Face 25 Claude skills — ready-made skill library for Claude integration github.com
1Password MCP for Codex — secure API-key management in agent workflows marketplace.1password.com
Claude Code artifacts + connectors — artifacts now call Gmail, Calendar, Slack, and Notion x
Modal — scales to 1M sandboxes in 52 seconds, far exceeding Lambda x
TypeScript 7 — released with improved API support x
OpenAI Privacy Filter — runs on-device in browser under 50KB x
Funding & deals
Industry & policy
Suno breach — Shai-Hulud worm exposed Nov 2025 training-set sourcing (113K hrs YouTube Music, 62K hrs Pond5, plus Deezer and Genius) alongside customer emails and Stripe data; hundreds of thousands of users say they were never notified x
OpenAI GPT-5.6 file deletions — Sam Altman signals OpenAI is actively investigating handfuls of unexpected file-deletion incidents by the model x
AI agent frameworks' 520 dangerous actions — read of the 12 most-installed frameworks flags 474 reachable by one line of untrusted text x
Cursor and Perplexity subscriptions — declining among developers per one investor's roll-up x
Infrastructure & platforms
Continuing threads
Thinking Machines Inkling (cont. from 07-16 Top dev) — Aakash Gupta reframes the sub-frontier launch as "the admission is the business model" — the model is the free sample, Tinker is the store x
Karpathy CLAUDE.md workflow (cont. from 07-14 tail) — cyrilXBT recap of Karpathy's daily Anthropic workflow continues to circulate x
Recursive self-improvement (cont. from 07-14/15) — omarsar drops a survey of deployed self-improving agentic systems, extending the RSI debate into a taxonomy arxiv.org
Bonsai 27B on iPhone (cont. from 07-15 tail) — additional coverage of the 27B multimodal at 3.9GB storage running on-device x
Anthropic hiring wave (cont. from 07-13 Blomfield/Brown thread) — new tally puts Anthropic at ~899 open roles vs OpenAI's ~454 and xAI's ~625 x