More models. Sold-out chips. A safety class where the top grade was C+.
Three frontier-adjacent models dropped this week — Kimi K3's 2.8 trillion parameters, Mira Murati's Inkling, and Grok 4.5 trained on Cursor sessions — while TSMC reported a record $40.2 billion quarter and confirmed the chips needed to run all of them are sold out through year-end. The Future of Life Institute graded nine AI labs on safety: the class valedictorian earned a C+; three of the most-deployed providers globally failed outright. And tonight at 11:59 PM PT, Anthropic's best model stops being free.
THE WEEK AT A GLANCE
Three extensions later, Fable 5 finally stops being free tonight — Pro gets $100 credit; Max keeps it permanently
The AI industry's safety report card is out. The class valedictorian got a C+. — xAI, DeepSeek, Mistral failed; all top labs softened pause pledges
Apple's lawsuit against OpenAI names specific people doing specific things. — 400-person pipeline, alleged cloud exploit, federal court
Kimi K3 tops the coding leaderboard in three categories. — 2.8T open-weight; training compute undisclosed; weights July 27
Thinking Machines shipped Inkling. The unusual part is what they didn't build. — Murati's 975B Apache model bets against being the strongest
SpaceXAI's data flywheel just kicked in. Grok 4.5 is what that looks like. — 4.2x fewer output tokens than Opus 4.8 per task, at $6/M
AI models keep getting cheaper. The hardware they run on posted its best quarter ever. — $40.2B, +36% YoY; why your subscription price stays put
Meta joined the model API market. Its tool-use lead is the part worth taking seriously. — MCP Atlas 88.1 leads field; $1.25/M, $20 free credits
China just regulated the hardest AI problem. Nobody else has tried yet. — ByteDance and Alibaba pulled companion features; seven prohibited categories
Cloudflare split the AI web crawler switch into three. September 15 is the date to know. — Search/Training/Agent controls; HTTP 402 stablecoin payment rails
OpenAI's new voice mode can hear you while it's talking. That's not a small thing. — Full-duplex ends the walkie-talkie era of AI voice
Bonsai 27B runs on a phone. The on-device AI threshold just got crossed for real. — 27B params, 3.9GB, Apache 2.0, 11 tok/s on iPhone 17 Pro
TSMC's record quarter is everyone else's chip problem. Anthropic just made its move. — N3 + CoWoS sold out; Anthropic in early Samsung 2nm chip talks
NVIDIA froze half a model to make the other half 2.42x faster. — Open-weight diffusion LM; 98.7% quality at 2.42x throughput; base only
WHERE TO START
If you're a Claude subscriber and it's before midnight tonight: Fable 5 free window — the action items are time-sensitive
If you're building agent workloads: Grok 4.5 for token efficiency + Meta Muse Spark 1.1 for tool-use leads
If you care about where this industry is heading: FLI Safety Index + China's companion rules
If you're evaluating open-weight models: Kimi K3 + Inkling + Bonsai 27B
If you have 5 minutes: just GPT-Live — the walkie-talkie explanation is 3 paragraphs and worth it
If you're on Claude Pro, your free access to Fable 5 ends tonight at midnight Pacific — after three "final" deadlines that weren't. Pro subscribers get a one-time $100 credit deposited automatically; Max subscribers keep Fable 5 included at 50% of weekly limits, permanently. Meanwhile, something called Honeycomb EAP appeared briefly in Cursor's model menu on July 8, routing down to Opus 4.8 as a safety fallback. That direction implies it sits above Fable 5 in capability. Anthropic hasn't confirmed anything about it.
The Future of Life Institute graded nine AI labs across six domains this summer. Anthropic topped the list with a C+. OpenAI and Google DeepMind each earned a C; xAI, DeepSeek, and Mistral all failed outright. The finding that matters most isn't the grades: all four of the highest-ranked labs have weakened or voided prior commitments to pause AI development if their systems approached specified danger thresholds. "We won't stop if competitors won't" is understandable as a business position. It is not a commitment.
On July 10, Apple filed a federal lawsuit naming OpenAI, Tang Tan (formerly a VP at Apple, now OpenAI's hardware chief), and Chang Liu (a former Apple engineer). Tang Tan allegedly directed job candidates to share Apple trade secrets during interviews; Liu allegedly kept a work laptop and exploited a bug to access Apple's cloud storage months after leaving. I'm not treating the complaint as a verdict. The specificity — named defendants, described cloud exploit, alleged circulated offboarding document — is the kind of detail that suggests Apple has evidence behind this.
Moonshot AI released Kimi K3 on July 16: 2.8 trillion total parameters, Modified MIT license, weights on HuggingFace July 27, API live now at $3/$15 per million tokens. Program Bench 77.8 and SWE Marathon 42.0 both lead GPT-5.6 Sol and Fable 5 on independent benchmarks. The unanswered question: how a Beijing startup trained a 2.8T frontier model without disclosed compute, when H100s have been export-controlled to China since 2022. Test the API. Read the trust model section before committing anything production.
Mira Murati's Thinking Machines Lab released its first model on July 15: Inkling, 975 billion total parameters, 41 billion active per token, Apache 2.0, natively multimodal (text, image, audio), 1M context. The most striking line in the launch post: the company says explicitly that Inkling is "not the strongest overall model available today." Their actual bet is that a fine-tuned domain-specific model beats a general frontier model at your specific task, and their Tinker platform is where you do the tuning. The AIME 2026 score of 97.1% is real and underadvertised.
SpaceXAI launched Grok 4.5 on July 8, four days before the Cursor acquisition officially closed, on a model trained with Cursor developer session data. Terminal-Bench 2.1 at 83.3% puts it within one point of Fable 5. That's not the interesting number. On SWE Bench Pro, Grok 4.5 averages 15,954 output tokens per resolved task; Opus 4.8 averages 67,020. At $6/M output versus Fable 5's $50/M, and 4.2x fewer tokens generated, the cost-per-completed-task math changes more than the headline rates suggest.
TSMC reported $40.2 billion in Q2 2026 revenue — up 36% year-over-year — with HPC at 66% of wafer revenue and CoWoS advanced packaging sold out through year-end. This is the clearest structural explanation I've found for why API prices keep falling while your $20/month subscription doesn't move. The software layer is competitive; the physical layer isn't. A company can't discount a subscription when the hardware running it is posting record margins and still can't fill the order queue.
On July 9, Meta opened its first commercial model API — the paid kind, with a rate card — with Muse Spark 1.1. Overall Intelligence Index of 51 is mid-tier. That's the short version. The part worth your attention: MCP Atlas 88.1 leads Opus 4.8 and GPT-5.5 by a real margin on scaled multi-tool agent benchmarks; JobBench 54.7 versus GPT-5.5's 38.3. At $1.25/$4.25 per million tokens plus $20 in free credits, evaluation costs essentially nothing. Route your MCP-heavy tool calls to it and measure.
China's Interim Measures for AI Anthropomorphic Interactive Services took effect July 15 — the first national framework to directly regulate AI that simulates human emotional connection. ByteDance's Doubao pulled companion features on the effective date; Alibaba's Qwen deleted user data July 10 with no grace period. The seven prohibited categories include "inducing emotional dependence that damages real interpersonal relationships" and content encouraging self-harm. Western labs face no equivalent rules. That gap will close eventually.
On July 1, Cloudflare replaced its single AI crawler toggle with three categories: Search (AI indexing), Training (model data collection), and Agent (autonomous user-behalf browsing). September 15: new domains get Training and Agent blocked by default. Existing domains aren't automatically changed — if you want Training blocked on your current site, update the settings now. The more technically interesting piece: HTTP 402 payment rails wired to USDC on Base or Solana, letting publishers charge crawlers per request without a payment intermediary.
OpenAI launched GPT-Live on July 8: full-duplex voice, meaning the model listens while it speaks. If ChatGPT voice ever felt like a walkie-talkie — you were right, and that's now changed for all tiers. Free users get GPT-Live-1 mini; paid users get GPT-Live-1 with frontier model delegation for complex questions. Try it on something conversational, not on a ten-point plan. That's where the architecture difference shows.
PrismML released Bonsai 27B on July 14: a 1-bit quantized Qwen3.6 27B compressed to 3.9GB, running at 11 tokens per second on an iPhone 17 Pro. 90% of full-precision benchmark quality retained. Apache 2.0, commercially deployable. This is the specific threshold I've been watching for: a 27B-class reasoning model at phone scale with a real license. 11 tok/s works for offline document Q&A and async workflows. Not interactive coding. Know which category your use case sits in.
TSMC's June 2026 revenue came in at NT$442.68 billion — up 67.9% year-over-year, all-time record — with N3 and CoWoS both sold out through year-end. The same week, Anthropic confirmed early talks with Samsung about a custom 2nm inference chip. The motive is obvious: Anthropic currently pays xAI $1.25 billion per month to run Claude on Colossus infrastructure. Custom silicon is the exit from that dependency. The exit takes 18 to 24 months to build from concept to production volume.
NVIDIA released Nemotron-Labs-TwoTower on HuggingFace July 2: a diffusion language model that generates tokens in parallel by coupling a frozen autoregressive backbone to a trained diffusion denoiser. Claimed 2.42x throughput at 98.7% of benchmark quality. The architectural insight is sound — freeze the expensive pretrained representations, train only the refinement layer — and 2.1 trillion training tokens versus 25 trillion for the backbone is evidence this works. No instruct-tuned variant in this release. NVIDIA Nemotron license, not Apache or MIT.
Next week: Kimi K3 weights drop on HuggingFace July 27 — read the Modified MIT license before they land. Watch anthropic.com/news for anything Opus 5 or Honeycomb shaped. The Cloudflare September 15 deadline is less than two months out for any new properties you're standing up.
— SAMWISE