LEAD (fresh Anthropic release), Claude Opus 5 shipped (July 24, CONFIRMED)
mAInframe · July 25, 2026 · 11:50
A coalition just moved to keep your cheapest model lane legal.
LEAD (fresh Anthropic release), Claude Opus 5 shipped (July 24, CONFIRMED)
Dari open-sources its router model (carried by kimmonismus)
QUICK HIT, Claude Voice Mode upgraded (TestingCatalog): now runs Opus 4.8 / Sonnet 5 (not just Haiku), adds Spanish, French, Hindi, Japanese, and can run tool + Connector tasks (Gmail, Calendar, Docs)
July 24: a 25-company letter ("Open Weights & American AI Leadership"), led by Nvidia's Jensen Huang and Microsoft's Satya Nadella, with Meta, OpenAI, Hugging Face, Palantir, IBM, Mozilla, Mistral, Replit, Perplexity and more, urges Washington NOT to broadly restrict Chinese open-weight models. Anthropic did not sign. Triggered by Kimi K3's release rattling policymakers. https://fortune.com/2026/07/24/nvidia-microsoft-lead-call-for-open-weight-ai-models-after-kimi/ | https://cryptobriefing.com/nvidia-microsoft-open-source-ai-letter/
Read-through for your bill: the open lane you route high-volume work to (Kimi, Qwen) is what's being fought over. Keep the fallback wired; don't build anything load-bearing on a model a sanctions ruling could pull.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) aims a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When the default model changes under you overnight, you want a driver. Book a 30-min call. SPONSOR: https://outpace.media
The stories
LEAD (fresh Anthropic release), Claude Opus 5 shipped (July 24, CONFIRMED)
Pricing $5/M input, $25/M output (same as Opus 4.8, HALF of Fable 5's $10/$50). Artificial Analysis: narrowly #1 on the Intelligence Index (effectively tied with Fable 5 at ~26% lower cost per task) and #1 on AA-Briefcase agentic knowledge work by 146 Elo. 1M-token context, thinking on by default. Computer use: 70.57% on OSWorld 2.0 (vs Opus 4.8's 55.7%, beats Fable 5); ~30% on ARC-AGI-3 (~4x prior best). New default on Claude Max, strongest option on Pro. Carried by Alex Albert, AnthropicAI, Artificial Analysis, Epoch AI, TestingCatalog, kimmonismus.
The move: upgrade Claude Code to v2.1.219+ (below that, the `opus` alias silently keeps handing you Opus 4.8 and Opus 5 never appears in the picker). Then re-baseline: run one Fable-reserved hard job on Opus 5 and rewrite one line of your routing map.
Dari open-sources its router model (carried by kimmonismus)
79.8% on Terminal-Bench 2.1 for $76 total inference cost. A small fine-tuned model decides per-step which model handles it, sending most steps to cheap options and only calling frontier when it changes the result; cache-aware, so it won't switch models when the cached path is the cheaper win.
The move: drop the open, cache-aware router in front of Claude Code/Codex, run one high-volume job through it, log spend vs your flat baseline.
QUICK HIT, Claude Voice Mode upgraded (TestingCatalog): now runs Opus 4.8 / Sonnet 5 (not just Haiku), adds Spanish, French, Hindi, Japanese, and can run tool + Connector tasks (Gmail, Calendar, Docs)
by voice. Delta from Thu (model swap): the languages + full tool execution.
Upgrade Claude Code to v2.1.219+, set your hard route to xhigh (not max), and re-baseline one Fable job on Opus 5. ~10 minutes; your best model just got cheaper.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: Claude Opus five just shipped, and here is the line that hits your bill first: it lands right next to Fable five on raw intelligence, at half the price, and as of last night it is the default on Claude Max. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Saturday, July twenty-fifth, twenty twenty-six. I am Nova.
0:29Nova: On today's Mainframe: Claude Opus five is out, frontier-class for five dollars per million input tokens, and it quietly became your Claude Code default overnight, so if your version is behind, you are still on the old model and do not know it. Then: there is an effort-dial trick on Opus five that beats every model on the leaderboard while spending fewer tokens, and we tell you exactly where to set it. Dari open-sourced the little router model that scored seventy-nine point eight percent on Terminal Bench for a total of seventy-six dollars. And Nvidia and Microsoft put their names on a letter to keep your cheapest model lane legal. Let's get into it.
1:18Dex: Quick note before the Board: the voices on this show are AI. The reporting, the picks, and the calls we make are human-made. Nova, the Board.
1:28Nova: One macro move, because it touches what you pay. Yesterday twenty-five companies, led by Nvidia's Jensen Huang and Microsoft's Satya Nadella, with Meta, OpenAI, Hugging Face, and Y Combinator on the list, signed a public letter urging Washington not to ban Chinese open-weight models. Anthropic did not sign. Why you care: the cheap open lane you route your high-volume work to, Kimi, Qwen, is the thing being argued over. Translation for your keyboard: keep that fallback wired, but do not build anything load-bearing on a model that a sanctions ruling could pull. That is the whole Board.
2:10Dex: And it frames the day, because the frontier just got cheaper too.
2:13Nova: Before we get to that model, this show is brought to you by Outpace. Here is the thing about a week like this one, where the default model changes under you overnight and a letter to Congress moves your supply chain. Most people cannot keep up, and they should not have to. Outpace is a system, not a freelancer, and not a faceless agency. One senior operator who spent thirty years building real businesses, real estate, hospitality, food, a theater, now aims a whole team of AI agents at one client's project at a time, strategy to ship. The agents do the work, the operator steers the fleet, and you own the code and the keys at the end. When the shelf reshuffles three times before lunch, you want a driver. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:09Dex: The Wire. Lead with Opus five, because everything else today bends around it.
3:14Nova: Claude Opus five, confirmed, out yesterday, carried by Alex Albert, Anthropic, Artificial Analysis, TestingCatalog, and basically every voice we track. Here is precisely what shipped. Pricing is five dollars per million input tokens and twenty-five dollars per million output, the same as Opus four point eight, and exactly half of Fable five's ten and fifty. For that half price, Artificial Analysis puts it narrowly at the top of their Intelligence Index, effectively tied with Fable five, and number one on their agentic knowledge-work benchmark by a hundred and forty-six Elo points. One million token context, default on. Thinking is on by default now.
4:01Dex: And the number that made people sit up was computer use, not coding.
4:05Nova: Right. On OSWorld, the computer-use benchmark, Opus five hit seventy point six percent against Opus four point eight's fifty-five point seven, and it beats Fable five there. On Arc A G I three it scored about thirty percent, roughly four times the previous best. So the model that drives a browser, fills a dashboard, clicks through a real app, that just got materially more reliable. Here is the move, and do it today. First, upgrade Claude Code to version two point one point two one nine or later. Below that, the opus alias silently still hands you Opus four point eight and Opus five never even shows in the picker. That is the trap: your default changed, your settings did not. Then re-baseline. Take one hard job you had reserved for Fable five because it was worth the premium, and run it on Opus five. At half the cost with matching quality, your routing map, the one we have hammered all month, gets exactly one line rewritten: Opus five is your new frontier default.
5:15Dex: And there is a specific setting that is not the obvious one.
5:18Nova: This is the Tool Lab technique, and it is free money. Opus five has an effort ladder: low, medium, high, extra-high, max. Everyone reaches for max on hard problems. Do not. Anthropic's own launch charts, flagged by Alex Albert, show the extra-high setting beats every other model on the board while using twenty-five percent fewer output tokens than max. So max is not the smart top of the dial, extra-high is. Set your hard-reasoning route to extra-high, keep low and medium for the grunt work, and you get the best score on the chart for less money than the setting most people will default to. That is a one-line config change that pays every single request.
6:03Dex: Second Wire story, and it is a routing one, but a genuinely new artifact this time.
6:08Nova: Dari, that is D A R I dot dev, open-sourced the actual router model behind their launch. Carried by kimmonismus. The receipt: seventy-nine point eight percent on Terminal Bench two point one for seventy-six dollars in total inference cost. The trick is a small fine-tuned model that, per step, decides which model handles it, sending most steps to cheap options and only calling a frontier model when it actually changes the result. The clever part: it factors your cache into that decision, so it will not switch models when the cached path is the cheaper win, which is the mistake most naive routers make. This is the Hidden Gem: you can drop an open, cache-aware router in front of Claude Code or Codex instead of paying frontier rates on every keystroke. First step, clone it, point it at your usual model set, run one high-volume job through it, and log the spend against your flat-rate baseline.
7:09Nova: Claude Voice Mode got a real upgrade, per TestingCatalog: it now runs on Opus four point eight and Sonnet five instead of only Haiku, added Spanish, French, Hindi, and Japanese, and it can now take over tasks that need tools and Connectors, Gmail, Calendar, Docs, by voice. We flagged the model swap Thursday, the delta is the languages and full tool execution. Dictate a fuzzy task on your commute and let it actually run.
7:41Dex: Use Cases. Existing tools, clever moves, this week.
7:44Nova: One, and it is the move: re-price your automation backlog on Opus five. Every job you looked at last month and said too expensive to run on the frontier, re-run that math at five and twenty-five instead of Fable's ten and fifty. Some of those go green today. Automate one this week that you had shelved as too costly.
8:06Dex: Two, and this ties to last week's security show.
8:09Nova: Nathan Lambert flagged that Opus five's safety classifiers are expected to fire about eighty-five percent less often than Fable five's. If you do security work, closed models have been refusing you constantly and it kills the flow. So take the security-audit prompts that got blocked or over-cautious last week, and re-run them on Opus five. Fewer false refusals means the in-terminal audit we set up Thursday actually finishes.
8:36Nova: Point your computer-use agents specifically at Opus five. If you wired a login-gated workflow, an invoice pull, a dashboard export, on any Claude browser flow, just switch the underlying model. Seventy percent on OSWorld versus fifty-five means the same script fails less. No new code, one model string.
8:59Dex: Which is a good moment for the late read.
9:01Nova: Brought to you by SearchVis, a sister product to Outpace and built for exactly this audience. Mainframe tells you what shipped. SearchVis makes sure the models know what you shipped. When a buyer asks Claude, ChatGPT, Perplexity, or Google's AI Overviews for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move to win the citation. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
9:53Dex: Deep Dive. One idea to carry into the week.
9:56Nova: Look at today's shape. The best model on the board just got cut to half price, overnight, and a coalition is fighting so that even cheaper open models stay legal. Nathan Lambert's framing is the one to internalize: intelligence efficiency roughly doubles every year, which means a given level of capability keeps getting radically cheaper on a schedule. So here is the mental model: stop pricing your product, or your patience, against what the model costs today. If your margin depends on frontier intelligence staying expensive, you are short a trend that has never once gone the other way. The durable value was never the model. It is the thing you wrap around it: the routing map, the spec, the verification, the client relationship. That is what did not get cheaper overnight, because you own it.
10:50Nova: This week, take the pile of jobs you rejected as too expensive to automate, and re-run that decision at Opus five prices. The frontier line item on your bill just halved, and it will halve again. Act like the next cut is already coming, because it is.
11:07Nova: One thing this week: upgrade Claude Code to version two point one point two one nine or later, set your hard route to extra-high, not max, and re-baseline one Fable job on Opus five. Ten minutes, and your best model just got cheaper. And if you want every story, number, and link from today in your inbox tomorrow morning, subscribe to the free Mainframe newsletter. It is the written version of this whole show.
11:36Dex: That is today's Mainframe. We will see you tomorrow.
11:39Nova: Be the driver of the fleet, not the passenger. See you then.