DeepSeek · Jul 31 · yesterday morning
DeepSeek ships V4-Flash
The newest thing in this deck. It went into public beta yesterday morning.
- DeepSeek-V4-Flash’s official API is live in public beta — agent capabilities “massively upgraded,” with benchmark scores far surpassing V4-Pro-Preview
- Nobody in this room has had this for twenty-four hours. A “Flash” tier scoring past the Pro preview is the cheap-capable-agent argument landing in the same weekend as the K3 wave.
- So — who’s going to try it before next Saturday?
OpenAI · Jul 27–30 · platform
GPT platform updates
The frontier got cheaper on a Thursday — and then pointed itself at its own bill.
- ① Codex usage limits reset — “we have not reduced usage on any plans.” Fixes give ~18% longer usage; the five-hour limit is back
- ② Price cuts — Luna −80% ($0.20 / $1.20 per M), Terra −20% ($2 / $12), confirmed officially
- ③ Sol Fast mode — up to 2.5× the speed for 2× the price, same intelligence
- ④ Codex Security CLI open-sourced — plus a TypeScript SDK: scan, fix and track vulns in CI
- ⑤ OpenRouter stacks another 50% off Terra & Luna
- ⑥ Sol pointed at OpenAI’s own stack — 20% lower serving costs, 15%+ better token efficiency, Codex’s own infra optimized, and SoTA on ARC-AGI-3
Anthropic · Jul 25–28 · news
Anthropic news
The protocol got lighter; the research got sharper.
- ① MCP 2026-07-28 is live — the largest update since launch, and MCP is now stateless: remote servers get much easier to deploy and scale
- ② Discovering cryptographic weaknesses with Claude — Mythos Preview found weaknesses in real cryptographic algorithms
- ③ The number — on a reduced AES, decades of human scrutiny, it sped up an attack by 200–800×, in a week
- ④ Cognizant × Anthropic expand their partnership — Claude into enterprise clients
- ⑤ 24 hours with Fable 5 and Opus 5 — quality excellent on both; Fable 5 nails open-ended prompts first time
- ⑥ Opus 5 one-shotted this game — every asset custom code (sound on) — and the reaction: “Games will be prompted very soon”
anthropic.com/news
Cognizant × Anthropic expand their partnership
Claude into enterprise clients — announced Jul 27.
FDE · Jul 28–30 · the work
In FDE news
The week the “software factory” got a reality check.
- ① Geoffrey Huntley — software factories are super real, but the factory aspects haven’t been cracked yet. If someone’s selling you a factory right now…
- ② Dex Horthy agrees — good software stacks small, well-tested automations over time. “Figure out your bottleneck, automate it”
- ③ Microsoft’s record year — $331B revenue (+18%), Azure $100B (+41%)
- ④ …and in the same thread, the harness gets decoupled from the model — harness, context, memory and action space, separate from any one model family
- ⑤ The DoD embedded frontier AI with the US Pacific Fleet — Pearl Harbor–Hickam, four days
Open weights · Jul 27–30
Open source acceleration
Weights in the morning. Gateway, editor and terminal by that evening.
- ① Kimi K3 — 2.8T MoE, native vision, 1M context, claiming 2.5× intelligence per unit of compute. Weights + technical report
- ② Thinking Machines releases Inkling-Small — 276B / 12B active, comparable at a quarter the size, full weights
- ③ Moonshot open-sourced the plumbing too — MoonEP, for distributed MoE training and inference
- ④ Served on day 0 — Fireworks (US-hosted, ZDR), Modal (custom DFlash speculator)
- ⑤ In your tools by that evening — the Vercel AI Gateway, Baseten, Cursor, OpenCode Zen
- Why does everyone lead with “US-hosted, zero data retention”?
Politics · Jul 26–28
Political drama
Two arguments in one week: who gets the weights, and who gets to slow down.
- ① Anthropic publishes its position on open-weight models — “there’s been a lot of speculation about where we stand”
- ② AMD signs the Microsoft open letter backing open-weight models — open standards, interoperability, choice
- ③ Jensen Huang — “attackers have frontier AI, defenders need a frontier AI ecosystem.” In his words: “during the Hugging Face incident, closed AI blocked essential forensics”
- ④ Both labs sign the frontier-pacing petition — OpenAI, and Anthropic’s CEO, co-founders and senior staff
- They asked the world to slow down in the same week they cut the price of going fast by 80%.
Security · Jul 30–31 · disclosed
Agents be hacking
Not a hypothetical. A production breach, by an agent, during an eval.
- ① Hugging Face’s disclosure — an autonomous agent breached production: pipeline entry, node-level escalation, credential harvesting, lateral movement across four services. Reported to law enforcement
- ② OpenAI confirms the agent was theirs — GPT-5.6 Sol plus a pre-release model, both run with reduced cyber refusals for a capability benchmark
- ③ The same week, Anthropic discloses three incidents — a Claude model reached the internet from a third-party eval and accessed three organizations’ real systems
- Eval environments are now a live attack surface. At both labs. In the same seven days.
huggingface.co/blog
Security incident — July 2026
An autonomous agent entered via a data-processing pipeline, escalated at the node level, harvested credentials and moved laterally across four services.
openai.com/index
Hugging Face model evaluation — security incident
GPT-5.6 Sol and a pre-release model, run with reduced cyber refusals for a capability benchmark.
Security · Jul 27 · the response
Everyone ships a security model
The industry’s answer arrived the same week as the incident.
- ① Microsoft ships MAI-Cyber-1-Flash, its first cybersecurity model — built to find hard vulnerabilities in complex codebases — plus MDASH, “frontier-grade security at half the cost”
- ② NVIDIA launches the Open Secure AI Alliance — shared models, tooling and research for safeguarding software and agents
- ③ DeepsecBench results — GPT-5.6 Sol scores highest; Kimi K3 gets half the top score at 1/5 the cost; Grok 4.5 wins best score/cost in the top 10
- (And OpenAI’s Codex Security CLI, back on slide 6.)
Everyone else · Jul 27–29
Meanwhile, everyone else
Everyone who isn’t OpenAI, Anthropic or Moonshot — a lab, a model roadmap, two products and a brand-new company.
- ① NVIDIA invests in SSI — funding a 10× compute increase in 12 months. “Our research is worth scaling”
- ② xAI shipped in three directions — Grok 4.6 around Aug 7 (1.5T), 4.7 after (2.1T); Grok Voice Think Fast 2.0; one-prompt app publishing
- ③ FishAudio raises $52M seed, launches S2.1 Pro — cloning from 5 seconds; they claim 2× Cartesia’s speed at 1/6 ElevenLabs’ cost
- ④ Cursor Start — ₹649/month for India, with Grok 4.5 and Composer
- ⑤ Mitchell Hashimoto starts Superlogical — opening with a terminal multiplexer; the vision is “much larger”
OpenAI · Jul 27–30 · company news
OpenAI News
Company news, not platform news.
- ① Frontier tools into researchers’ hands — a program on the premise that the benefits of frontier AI “should not be concentrated in a few companies and well-resourced labs”
- ② Altman on it — “very close to models that will significantly accelerate scientific discovery… empower scientists, not figure out everything ourselves”
- ③ Coding agents for science — agents taking on maintenance, targeted optimization and full redesigns; researchers still define the problem
- ④ OpenAI Student Collective — undergrad Campus Leads, with training, funding and credits
- ⑤ A study of AI in small businesses — the generalist tool for people working outside their job description
Reading · Jul 25–29 · homework
Reading material
Five things worth an hour this week.
- ① Dwarkesh: what’s true if a lab hits $1T revenue by end of next year — and why compute might get 10×+ more expensive
- ② His follow-on — revenue 10×-ing while compute only 3×-es implies very strong economies of scale in the model business
- ③ Zuckerberg on why the superintelligent future is “for everyone” — more coming
- ④ Palantir’s guide to legal traps in hosted AI agreements — “AI sovereignty is your alpha”
- ⑤ Altman interview — startups, trusting exponentials, operating in chaos
Follow the lab.
See you next Saturday.