The Signal — August 22, 2026

Nobody released a frontier model this week. Nvidia, Anthropic, and OpenAI each shipped something that changes what their AI can do anyway, and in all three cases the change lives outside the weights.

Nvidia's harness took Claude Opus 5 from 30% to a perfect score

Nvidia published research on Friday describing AVO, an agent architecture built around persistent memory, a supervisor component that watches the work, and structured tool use. Running Claude Opus 5 inside that system, Nvidia reports it completed all 183 levels across all 25 environments of the ARC-AGI-3 public set, and did it with 12% fewer environment actions than VISTA, the previous approach.

The comparison that carries the weight is the other number: the same model, handed the same benchmark without the harness, scored 30%. That was still the best bare-model result Nvidia measured across everything it tested. So the gap between a leading frontier model and a solved benchmark was closed entirely by engineering that sits around the model, not by anything inside it.

Two things temper this. It is the public set rather than the private evaluation ARC uses to guard against overfitting, and Nvidia is grading its own homework on a system it would like people to build with. TechCrunch's Julie Bort noted that ARC-AGI-3 is a benchmark OpenAI has publicly bristled at, which is worth holding in mind when reading any triumphant score on it. Nvidia also reported separate results where AVO explored more than 500 optimization directions on GPU kernel work, committed 40 versions, and beat FlashAttention-4 by up to 10.5% on DGX B200 hardware.

What this does not show is a step toward general intelligence, and Nvidia does not claim it does. What it shows is that a lot of the capability people attribute to models is currently sitting in scaffolding that anyone can build and almost nobody benchmarks.

Sources: NVIDIA Technical Blog · TechCrunch · Techmeme

Anthropic gave its most restricted model a scan button instead of a prompt box

Claude Security now runs on Claude Mythos 5. That is the tier Anthropic had been keeping behind Project Glasswing since April, available only to a small group of organizations securing critical software, on the reasoning that a model good enough to find novel vulnerabilities is also good enough to write the exploits for them.

The product connects to a GitHub repository, traces how data moves across files, reads git history instead of matching against a rule list, and then challenges its own findings in a verification pass before surfacing anything. What comes back is a CWE category, confidence and severity ratings, and a suggested patch. It is in public beta for Claude Enterprise customers and bills as ordinary token usage with no model add-on.

The interesting design choice is what users do not get. There is no chat window. Patching happens separately in Claude Code using whatever models the organization already pays for, so Mythos never becomes something you can talk to. Anthropic's position is that risk concentrates at the prompt, and that handing someone a fixed artifact instead of a steerable model is the control. That is a real argument. It is also entirely Anthropic's argument, and nobody outside the company has tested it. Every patch still requires a human to approve it.

Anthropic paired the release with the Defender Advantage Fund, $35 million in Claude credits for organizations patching vulnerabilities in open-source projects, and said it plans to widen its Cyber Verification Program over the coming weeks. Enterprise-tier only, though. Teams on Pro, Max, or Team plans get a Claude Security plugin that runs on their existing models, which is a different product wearing a similar name.

Sources: Anthropic · MarkTechPost · The Decoder

OpenAI cut its frontier price by a fifth, for ninety days

GPT-5.6 Sol now costs $4 per million input tokens and $20 per million output tokens for standard short-context work, down from $5 and $30. That is a 20% cut on input and a 33% cut on output. The lower rates also flow through credit consumption on ChatGPT Work and on Codex, while Pro, Plus, and Business subscription prices stay where they were.

The three-month expiry is the part most coverage mentions once and then forgets. This is not a repricing of the frontier tier, it is a discount window, and anyone building cost models on the new numbers is building on something with a stated end date. When OpenAI cut prices in late July, it took 20% off Terra and 80% off Luna and left Sol alone. Sol was the model it did not discount. Now it is, temporarily, and Reuters reports the move against a backdrop of pressure from Anthropic and from Chinese labs shipping capable models at aggressive prices.

Nothing about the model changed; the weights are what they were on Thursday. What changed is a number in a billing table, which for a lot of teams deciding what to run in production is the variable that actually moved.

Sources: Reuters · Storyboard18 · OpenAI

On the Editor's Desk

DeepSeek released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to its Flash tier, and it nearly made the edition. It got cut for space rather than quality. The comparison everyone is repeating, that it approaches Opus 4.8 on multimodal agent work, comes from DeepSeek's own benchmark table and varies by task, so we would want a couple of outside evaluations before treating it as settled.

We also looked at a Bloomberg report that Anthropic hired Amir Salek, who ran Google's TPU program until 2022, as groundwork for building its own chips. Good story, but it rests on one outlet and the chip effort is a hiring decision rather than a roadmap. If a second source or an actual program surfaces, we will come back to it.