The frontier got cheaper on Friday. The systems around it became more valuable. That is the cleanest way to read this edition of AI news today. OpenAI temporarily cut GPT-5.6 Sol API prices, while NVIDIA published an agent result built around memory, supervision, tools, and repeated testing—not a magical model swap. At the same time, NVIDIA reportedly put $6 billion behind access to Poolside’s model-building machinery, and xAI placed Grok 4.6 inside Google’s enterprise catalog. The pattern is hard to miss: advantage is moving from “who has the smartest chatbot?” toward who can assemble models, harnesses, data, power, and distribution into a dependable system. For builders, that makes experimentation less expensive. For leaders, it makes architecture and vendor leverage more important. And for anyone hoping the AI market might become simpler, well, Monday has other plans.

AI news today: OpenAI cuts GPT-5.6 Sol prices for three months

OpenAI lowered GPT-5.6 Sol’s API and credit pricing by more than 20% on August 21. Its current model documentation lists $4 per million input tokens and $20 per million output tokens, with the promotion running at least through November 21, 2026. Very large prompts above 272,000 input tokens still trigger higher rates.

The immediate answer is straightforward: frontier-agent experiments just became cheaper, but the discount is temporary. Teams should use the window to measure cost per completed task—not cost per token—and avoid baking promotional economics into a permanent business case. Watch whether rivals answer with their own cuts, and whether OpenAI keeps some portion of the discount after November.

NVIDIA’s AVO shows why the agent harness now matters as much as the model

NVIDIA researchers reported that Agentic Variation Operators, or AVO, completed all 183 levels across the 25 environments in the ARC-AGI-3 public set, earning a 100.00 RHAE score. The same general agent architecture had previously run a seven-day GPU-kernel optimization loop. According to NVIDIA’s technical account, AVO combines persistent memory, tools, evaluation, and a supervisor that redirects stalled work.

This is a notable systems result, with important limits: it covers the public set, the comparison is not a controlled ablation, and it does not prove general intelligence. What it does show is commercially useful. A model’s benchmark score does not tell you how well an agent will recover, preserve state, or improve through feedback. The next test is independent replication on private tasks and real production workloads.

NVIDIA reportedly pays $6 billion for Poolside’s model-building machinery

In a less conventional kind of AI deal, Newcomer reported that NVIDIA agreed to a $6 billion non-exclusive license for Poolside’s model-development software, plus a $1 billion investment at a $12 billion pre-money valuation. The details come from an investor letter obtained by Newcomer; neither company has publicly confirmed the full terms, so treat the structure and figures as reported rather than final corporate disclosure.

If the terms hold, the strategic object is not merely a coding model. It is the factory used to build models. That would give NVIDIA more control over the open-weight software layer surrounding its chips while Poolside remains independent. It also extends the infrastructure-and-financing story we tracked in Friday’s morning brief. Watch for regulatory filings, employee outcomes, and clarity on which intellectual property the license actually covers.

Grok 4.6 lands in Google’s enterprise agent catalog

xAI made Grok 4.6 available through Google Enterprise Agent Platform and Model Garden on August 21. The company announcement lists a 500,000-token context window, four reasoning levels, and prices of $2 per million input tokens, $0.50 for cached input, and $6 per million output tokens.

For enterprise buyers, the distribution channel may matter more than another leaderboard position. Google customers can evaluate Grok inside an existing cloud procurement and governance environment instead of opening a separate model stack. For Google, carrying a rival model reinforces its role as a neutral AI control plane. What comes next is practical: regional availability, enterprise controls, latency, and whether Grok wins workloads when buyers compare it beside Gemini and other Model Garden options.

NVIDIA backs Cloverleaf to accelerate powered AI sites

Cloverleaf Infrastructure announced a strategic partnership with NVIDIA, including an undisclosed minority investment, to advance sites and utility coordination for AI data centers in the United States. The August 21 announcement says Cloverleaf has worked on multiple gigawatt-scale projects; it does not disclose the investment amount or binding capacity commitments.

This is the physical counterpart to the model race. Chips only become revenue-producing systems after developers secure land, transmission, generation, cooling, and permits. NVIDIA is therefore reaching further upstream into the bottlenecks that determine when “AI factories” can actually switch on. The next evidence to demand is project-specific: named locations, interconnection milestones, financing, customers, and energization dates.

ChatGPT and Codex can work with Apple Messages on supported Macs

OpenAI’s August 21 release notes confirmed that the Apple Messages plugin in the ChatGPT desktop app can let Codex and ChatGPT Work read and search iMessage, SMS, and RCS conversations, then prepare or send messages through the Mac’s Messages app. The feature is limited to Apple-silicon Macs, requires the connection to be enabled, and asks for approval of recipients and the message by default. Availability can also depend on plan, region, and workspace settings.

The useful part is obvious: finding an address buried in a thread or drafting a logistical reply becomes an agent task. The risk is equally obvious: private conversations contain other people’s data. Teams should treat message access as a high-trust permission, keep send confirmation on, and test with non-sensitive threads before deciding whether convenience earns that level of access.

Watch & Learn

Tune the Harness, Before Tuning the Model with LangChain | Nemotron Labs

This official NVIDIA Developer tutorial walks through a real agent failure, changes the surrounding prompt and tool setup, then validates the fix with an evaluation. It is best for builders who already understand basic prompting and want a practical view of harness engineering. Budget about 50 minutes; the useful lesson is the debugging method, not any single framework.

AI, Translated

Agent harness

An agent harness is the software wrapped around an AI model that turns a response generator into a working agent. It supplies instructions, tools, memory, permissions, retries, and checks. Imagine an LLM asked to investigate a bug: the harness lets it open files, run tests, remember failed attempts, and stop before making an unsafe change. You should care because reliability and security often improve more from a better harness than from switching to a newer model.

Try This Today

Audit an AI-cost spreadsheet with Grok Skills

Goal: turn a usage CSV into a reviewable cost scenario in 10 minutes. Grok’s built-in spreadsheet skill can generate formatted workbooks and analyze uploaded data, according to xAI’s product documentation. Access may vary by platform, plan, or workspace policy.

• Remove customer names, secrets, prompts, and any data you should not upload.
• Upload the CSV in Grok and ask for a separate untouched raw-data tab.
• Request formula-driven baseline, −20%, and +20% usage scenarios, then reconcile the totals against the source file.
• Open the workbook and inspect several formulas manually before sharing it.

Copy-ready prompt: “Analyze this AI usage file. Preserve the original data on a read-only raw tab. Calculate monthly spend by model, add baseline, −20%, and +20% usage scenarios with formulas, flag missing or inconsistent prices, and create a one-paragraph executive summary. Do not invent rates; mark any missing input clearly.”

One thing to remember

Cheaper intelligence is useful. A system that can use it reliably—with memory, tools, controls, power, and evidence—is where the durable advantage is moving.