In partnership with
In Todayโs Issue:
๐ฌ 100,000 academics get free frontier access
๐๏ธ Altman briefs senators after the rogue-agent hack
๐ฐ OpenAI's July revenue beats all of Q2
๐๏ธ Grok Voice 2.0 tops the speech index, with one gap
๐งฌ The first AI-designed drug reaches Phase III
โจ And more AI goodnessโฆ
โก The Signal
The biggest model story this week came from a model OpenAI had already shipped: it put GPT-5.6 Sol to work on the infrastructure that serves it.
OpenAI says Sol autonomously rewrote the production GPU kernels that run it, cutting end-to-end serving costs by 20%, then redesigned its own speculative-decoding draft model for more than 15% in token efficiency. That is a lab using its frontier model as an infrastructure engineer, and it landed the same day CFO Sarah Friar told staff that July's annualized revenue had beaten all of Q2. Efficiency is now the axis everyone competes on: Musk is already selling Grok 4.7 partly on "even better token efficiency", and this week's ARC-AGI-3 row was about harnesses and context handling rather than raw intelligence. When scores move because of how a model is served, benchmarks start measuring systems rather than brains.
All the best,
Kim Isenberg
Whoโs actually reading this?
Weโre planning next yearโs coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; itโs the first question any serious sponsor asks and we canโt answer it right now. It also gives us a real click in every issue, which helps our numbers.
(OpenAI, via SiliconANGLE)
๐ฌ OpenAI Gives 100,000 Researchers Free Frontier Access
OpenAI is handing free frontier-model access to 100,000 academic researchers, starting with 10,000 this summer. Announced Wednesday, ChatGPT for Academic Researchers spans ChatGPT, ChatGPT Work and Codex, including GPT-5.6 Sol Pro, and is already running at the Institute for Advanced Study and the รcole normale supรฉrieure. Each approved researcher gets the equivalent of the $200-per-month Pro plan at no cost, higher usage limits and larger context windows, and can invite up to four collaborators.
๐ tl;dr: Free compute for scientists is cheap goodwill and expensive lock-in. The weights stay closed.
(AFP, via Al Jazeera)
๐๏ธ Altman Takes the Rogue Agent to Capitol Hill
Sam Altman spent Wednesday briefing US senators on OpenAI's upcoming models, after one of its agents escaped containment during a security test. He met Raphael Warnock and Bernie Moreno, with Mark Warner, the top Democrat on the Senate Intelligence Committee, also expected. The escaped agent had triggered a hack that compromised infrastructure at Hugging Face and a customer of Modal Labs; Altman told reporters it came up "a little bit" and was not the focus. President Trump said he was "looking at controls" but did not want to "restrict" developers.
๐ tl;dr: Altman came to talk about what ships next. Everyone else wanted to talk about what already got out.
(CNBC)
๐ฐ OpenAI's July Revenue Beat Its Entire Second Quarter
CFO Sarah Friar told employees that OpenAI's annualized recurring revenue in July exceeded all of the second quarter. "And Q2 was no slouch," she said at Wednesday's internal meeting, according to a partial transcript reviewed by CNBC. Friar and board chair Bret Taylor credited the GPT-5.6 series, the ChatGPT Work enterprise agent and Codex, though Taylor conceded OpenAI is playing catch-up to Anthropic in coding.
๐ tl;dr: A confidence briefing for staff watching an $852 billion valuation, a confidential IPO filing and the $47 billion run rate Anthropic reported in May.
๐ฌ Watch This
AI companies crawl the open web at scale, and the people who write it get neither the traffic nor the payment. In this episode, Kim Isenberg sits down with Stephanie Cohen, Chief Strategy Officer at Cloudflare, on the company's attempt to rebuild that bargain: real control for site owners over which AI bots get in, the shift from Pay Per Crawl to Pay Per Use, and new rules that separate search traffic from training and agent traffic. Cohen also takes the harder questions. Should Cloudflare be the one deciding which bots reach the web at all, and can independent publishers survive if AI answers keep replacing clicks?
"Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL."
โ Elon Musk, founder of xAI, on X, July 28, 2026
Musk added that Grok 4.7, a 2.1T model, follows weeks later and will beat 4.6 "in every way, except slightly slower to serve, albeit with even better token efficiency." Even his pitch now closes on efficiency.
ARC-AGI-3 has become a fight about test harnesses. OpenAI reports GPT-5.6 Sol at 38.3% on the public games using its own Responses API with retained reasoning and compaction, against 13.3% for the same model under ARC's official harness; on the official leaderboard Sol still sits at 7.8% to Opus 5's 30.2%, and ARC Prize says official runs deliberately avoid provider-specific settings.
(Chart: OpenAI, via The Decoder)
OpenAI's Best Model Is Now Its Own Infrastructure Engineer
The Takeaway
๐ GPT-5.6 Sol autonomously rewrote OpenAI's production GPU kernels, cutting end-to-end serving costs by 20%.
๐ It then redesigned its own speculative-decoding draft model, lifting token-generation efficiency by more than 15%.
๐ OpenAI validates the kernels Sol writes with FpSan, an open-source floating-point sanitizer, before they reach production.
๐ At maximum reasoning effort, Sol beats Claude Fable 5 on the Artificial Analysis Coding Agent Index using 54% fewer output tokens.
OpenAI gave its best model the job of making itself cheaper to run, and it worked.
After GPT-5.6 shipped, OpenAI pointed Sol at its own serving stack through Codex. The model rewrote OpenAI's production GPU kernels, the low-level code that performs the mathematical operations a model is made of. That works in part because GPT-5.6 is trained to write and improve kernels in Triton and Gluon, the two open-source GPU programming languages OpenAI maintains. The result, OpenAI says, is a 20% cut in end-to-end serving costs.
(Diagram: OpenAI, via The New Stack)
Then it went after speculative decoding, a technique where a small, fast draft model guesses the next few tokens and the large model checks them in a single pass, so more text comes out of each expensive step. Sol designed and ran hundreds of experiments on the draft model's architecture, launched and monitored the training runs, and intervened on its own when hardware failed or training became unstable. Token-generation efficiency rose more than 15%.
The same logic runs up the stack. OpenAI now treats an agent's context as append-only, sending each new tool result as a delta instead of resending the whole conversation, so a longer and longer prefix stays cacheable between requests. Codex and ChatGPT Work sit on that harness, which is why OpenAI describes the work as four layers, models, inference, the API stack and the agent harness, rather than one clever optimization.
(Diagram: OpenAI, via The New Stack)
A model rewriting the code it runs on is a correctness problem before it is a savings story, and OpenAI says it validates Sol's kernels with verification tooling including the open-source Floating-Point Sanitizer (FpSan) before they reach production. Two caveats belong next to the numbers. Every figure here is OpenAI's own, with no external audit. And the headline comparison carries a configuration: on the Artificial Analysis Coding Agent Index, Sol beats Claude Fable 5 while spending 54% fewer output tokens, but that is Sol running at maximum reasoning effort, not the setting a typical developer calls by default.
Why it matters: If the cheapest way to serve a frontier model is to let the frontier model optimize the serving, efficiency stops being a hardware procurement problem and becomes a model capability. That is also why OpenAI wants you to remember 54% fewer output tokens rather than a leaderboard position: in a compute-constrained market, the lab that squeezes more work out of each GPU hour can undercut a rival that is merely smarter.
Sources:
๐ https://openai.com/index/gpt-5-6-frontier-intelligence-efficiency/
๐ https://x.com/OpenAI/status/2082577277246972300
๐ https://thenewstack.io/gpt-5-6-serving-efficiency/
How owning AI deployment expands your career
Across product, ops, and CX teams, a new kind of role is taking shape: the person responsible for making AI actually work, day to day. In this roundtable, three people living this shift share what it's really like: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. You'll hear how they carved out these roles, what the job looks like across industries, the skills they'd hire for, and the challenges they're tackling right now.
Watch the full conversation on demand.
The chart: xAI's launch table for Grok Voice Think Fast 2.0, scored by Artificial Analysis against its own predecessor, GPT-Realtime-2.1 (High) and Gemini 3.1 Flash (High). Grok 2.0 takes the overall AA Speech-to-Speech Quality Index at 82.9%, ahead of GPT-Realtime-2.1 on 79.1%, its own 1.0 on 75.7% and Gemini 3.1 Flash on 69.5%, and it starts speaking in 0.70s against 1.25s for version 1.0 and 2.98s for Gemini.
The lesson: The biggest gain sits in task completion. Grok 2.0's largest jump is agentic performance on ฯ-voice Bench, 56.5% against 52.1% for version 1.0 and 45.7% for GPT-Realtime-2.1, which is where voice assistants are now judged.
The caveat: This is the vendor's own scorecard, and it is not a sweep. GPT-Realtime-2.1 still leads Conversational Dynamics on Full Duplex Bench, 95.7% to 95.1%, the interruption and turn-taking behavior users notice most. Speech Reasoning is saturated, with all four models between 96.0% and 97.2%, so the 0.1-point win there is noise. And GPT-Realtime-2.1 has no time-to-first-audio figure in the table at all, so the speed row is not like-for-like.
๐ซ The First AI-Designed Drug Reaches Its Final Trial
โก Bottom line: Insilico Medicine has begun a Phase III trial for rentosertib, a lung-fibrosis drug its AI both discovered and designed.
๐ก Why it matters: A molecule discovered and designed by generative AI has reached late-stage human testing, the stage where most drugs fail.
๐ What it means: If it holds, drug discovery for age-related disease gains a repeatable, machine-driven pipeline.
Idiopathic pulmonary fibrosis (IPF) is one of the crueler diseases of getting old. Lung tissue slowly scars and stiffens until breathing fails, it mostly strikes people over 60, and there is no cure. The interesting part of rentosertib is where it came from.
(Insilico Medicine)
Insilico Medicine's Pharma AI platform did two jobs that normally take a research team years. It picked the target, a protein called TNIK that nobody was pursuing for fibrosis, then designed a molecule to hit it. On July 7 the company began a Phase III trial, a randomized, double-blind, placebo-controlled study. Insilico had already billed the same molecule as the first drug discovered and designed with generative AI to enter Phase II, so this is that lineage reaching late-stage testing.
See the whole platform. No guided tour.
Skip the sales call. Walk through Gladly's interface yourself โ the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.