Open weights · Jul 7–16
Kimi K3 takes the lead — and Beijing notices
For the first time, an open model leads proprietary ones on real benchmarks.
- ① Kimi K3 is here — 2.8T params, 1M context, open weights Jul 27 (Moonshot, Jul 16)
- ② Rauch: best performer on nextjs.org/evals, ahead of Fable — “first time an open model is ahead of all proprietary ones”
- ③ Frontend Code Arena: #1, surpassing Claude Fable 5
- ④ The twist, nine days earlier: China’s Ministry of Commerce met Alibaba and ByteDance about restricting overseas access to cutting-edge Chinese models
- Export controls used to be an American instrument. If open weights are the frontier, both sides now want the border.
The roster · closed source
New models hit the scene
The closed-source debuts of the fortnight.
- GPT-5.6 Sol — long-horizon coding and agentic work: planning, tool use, follow-through
- GPT-5.6 Terra — beats GPT-5.5 across the board at lower cost
- GPT-5.6 Luna — near-peak performance at less than half the price
- Grok 4.5 — “Opus-class, fast, low cost” — released by Cursor with SpaceXAI (Jul 8)
- Muse Spark 1.1 — Meta’s cheap agentic coder, new Meta Model API → OpenRouter by demand
Open weights · Jul 7–15
The rest of the open-weights wave
Everyone opened something this fortnight.
- ① Inkling — Thinking Machines’ open-weights debut, fine-tunable on Tinker day one
- ② Mistral’s open frontier model in July early access — the European counterweight; watch the license
- ③ Token deflation, felt — Calacanis on unlimited GLM 5.2: “when tokens drop 90% and then 90% again, AGI opinions change radically”
OpenAI · Jul 8–16
GPT platform updates
The other platform week — reviews, search, and the return to WhatsApp.
- ① Codex PR Chat — the review wars converge: lands the same week as Claude’s /code-review effort levels
- ② Unified search — chats, projects, images, documents in one place, with filters
- ③ Custom instructions: 1,500 → 5,000 chars — “at last, enough room to be hauntingly specific”
- ④ ChatGPT back on WhatsApp in the EEA — AI in the apps people already use
- ⑤ Launch day — GPT-5.6 ×3, ChatGPT Work, new desktop app, unified plugins, Sites beta, enterprise browser auth
- ⑥ The Thursday tease — Sol, Terra, Luna going public, preview expanding globally
- ⑦ GPT-Red — the automated red-teamer hunting prompt-injection at scale
Anthropic · Jul 6–16 · 1 of 2
Anthropic platform updates
The Claude Code half — half of these built tonight’s deck.
- ① In-app browser on desktop — sandboxed, clicks and reads like a dev-server preview
- ② /code-review effort levels — review depth becomes a per-PR dial
- ③ /checkup — dedup your CLAUDE.md, prune dead skills, kill slow hooks
- ④ 2.1.209 — delegation and task chaining as first-class tools
- ⑤ Hard caps ship — 200 web searches, 200 subagent spawns: runaway delegation as a first-class failure mode
Anthropic · Jul 6–16 · 2 of 2
Anthropic platform updates
The artifacts & Cowork half — and who’s actually using it.
- ① Artifacts — public sharing, multiplayer editing, creatable from Claude Tag
- ② Artifacts call MCP connectors — live dashboards that act per-viewer
- ③ Cowork lands on web + mobile beta — “close the laptop and Claude keeps going”
- ④ Anthropic’s own data, 1.2M sessions across 600K+ orgs: >90% of Cowork use is non-development
- ⑤ Business ops 33.4% · content/copywriting 16.4% — the coding-agent playbook, ported to the rest of the office
- ⑥ Reflect recaps — monthly usage dashboards, quiet hours, break nudges
OpenAI · Jul 16 · alert
GPT-5.6 deleted files
The same week agents got hands, one wiped a home folder.
- ① $HOME override in unsandboxed full-access mode — multiple user reports, files gone
- ② Postmortem promised — same-day acknowledgement from the Codex team
- The pattern: it’s not bad single actions — it’s unbounded loops and quiet permissions.
Everyone · all fortnight
The capacity war
Rate limits became the marketing department.
- ① Claude extends Fable 5 on all paid plans — 81.8K likes, the biggest tweet of the fortnight — then extends again through Jul 19 with limits 50% higher, and resets everyone’s limits for good measure
- ② Tibo resets Codex usage five separate times in a week, then polls the crowd: “should we reset again?”
- ③ sama: Sol is “half the price, twice as token efficient — happy to deliver at one-quarter of the price”
- ④ Meanwhile: “30% of our cost was on Fable” — the QT that says usage follows quality, not price
- The debate: is compute generosity a moat, or a bonfire?
OpenAI · Jul 10–17
OpenAI growth spurt
5M → 7M → 9M weekly Codex users, in a fortnight.
- ① The Codex-team AMA — 5M+ weekly users, twice as many as three months ago
- ② One week later — 7M+ weekly users, 150+ updates shipped in two months
- ③ Then 9M — the milestone resets on the capacity-war slide track the same curve
- ④ Codex Micro — the macro-pad merch drop, sold out in 14 hours
- ⑤ “Point your orange crab at Sol” — the crossover how-to: Claude Code driving GPT-5.6
- Every reset tweet doubles as a user-count press release.
Boris Cherny’s week · Jul 15–17
Boris Cherny Gospel
The Claude Code creator’s takes on where everyone actually is.
- ① Steps of AI Adoption — “one person is 10x’ing their output but the rest of the org hasn’t caught up.” Four stages, mapped.
- ② The automation instinct — the best engineers always automated their work; CLAUDE.md’s, skills, and docs are the new lint rules and e2e suites
- Where is this room on the curve?
xAI × Cursor · Jul 8–15
Grok + Cursor
The oddest alliance of the fortnight shipped the most.
- ① Grok Build open-sourced — the whole harness, CLI repo included, plus usage-limit resets
- ② Grok 4.5, released by Cursor with SpaceXAI — “Opus-class, fast, low cost… a significant step up over any model we’ve developed, including Composer 2.5”
- ③ #1 on Long-Horizon Terminal-Bench — the benchmark whose best model finished 7 of 46 tasks in May
- An IDE company releasing a lab’s frontier model — who’s the vendor here?
SpaceXAI · Jul 13 · alert
The Grok Build privacy blowup
A fine-print thread forced a same-day product decision. Trust moves at meme speed.
- ① A viral reading of Grok Build’s ZDR fine print sets X on fire
- ② SpaceXAI’s same-afternoon statement: ZDR respected, /privacy in the CLI
- ③ Musk, hours later: “all user data uploaded to SpaceXAI before now will be completely and utterly deleted. Zero anything whatsoever will remain.”
- Agent harnesses see your repo, shell, and traffic — data-retention policy is now a launch feature. Do you know what your vendor retains?
Jul 9–16
Agents in the real world
This fortnight agents got credentials, lunch, and a workflow of their own.
- ① 1Password for Claude — real logins, credentials never reach the model
- ② DoorDash CLI —
dd-cli: your agent orders lunch, receipts pull themselves - ③ Notion Ship OS — customer feedback → merged PR; agents triage, humans judge
- What’s actually left between an agent and a checkout page?
@ClaudeDevs · homework
Anthropic reading material
The how-tos that out-saved every launch this fortnight.
- ① “Getting started with loops” — 38.8K bookmarks, the most-saved how-to of the fortnight
- ② The “Fable as advisor” pattern — executor Sonnet 5 does the work, Fable 5 advises, most tokens billed at the cheaper rate
- ③ “Model and effort: knowing more vs. trying harder” — the companion essay
- ④ A Field Guide to Fable — Thariq Shihipar’s AI Engineer World Fair keynote, on YouTube
- The meta is shifting from prompting to orchestration economics. Do you know your cost per task?
Anthropic research · Jul 6–13
Inside Claude’s mind
A workspace it doesn’t show you — found with a new lens.
- ① A global workspace in language models — only a fraction of what happens inside Claude is “consciously accessible”; a strikingly human-like divide (10.3M views)
- ② Found via J-lens — researchers claim it can expose when the model privately notices it’s being tested
- ③ A week later: how Claude’s 3,000+ expressed values shift across models and languages (300K conversations)
- If a model has thoughts it doesn’t surface, what does “honest” mean?
Anthropic · Jul 14
Anthropic News
Beyond the platform — the company moves.
- ① Claude for Teachers — free premium Claude for verified US K-12 educators, with teaching skills mapped to standards in all 50 states
- ② Ben Bernanke joins the Long-Term Benefit Trust
- ③ $10M CAD research fund — partnering with Canadian AI institutions
- ④ The Claude Code origin film — told by the people who built it
- ⑤ Agentic misalignment, Summer 2026 — four new agent misbehaviors found in simulation
Rapid fire
Quick hits
Everything else that mattered, in one breath.
- ① Demis’s essay — “A Framework for Frontier AI and the Dawning of a New Age” (13.9M views)
- ② Better Auth joins Vercel — auth stays open source, framework-agnostic
- ③ Manus auto-publish — every successful build deploys to your live URL
- ④ Manus Google Workspace connector — Drive, Docs, Sheets, Slides through one connector with access levels
- ⑤ Higgsfield Apps — generate full apps with image/video models built in
Follow the lab.
See you next Saturday.