The race moved from "smartest model" to "cheapest finished task." Anthropic shipped Sonnet 5 at intro pricing $2/$10 per M tokens (through Aug 31), yet per-task it can cost MORE than Opus 4.8. Per-token price and per-task price have officially divorced.
Artificial Analysis (evaluated Sonnet 5 pre-launch): 53 on the Intelligence Index, #5 model, 2-3 pts behind GPT-5.5 (xhigh) and Opus 4.8 (max). Sonnet 5 = $2.29/task, ~15% more than Opus 4.8, driven entirely by higher token usage.
Operator takeaway: stop pricing on the sticker per-token rate. Price and measure per completed task, tokens actually burned. Expect API prices to keep falling (see OpenAI below); write contracts so you keep the spread.
The stories
Claude Sonnet 5 ships as the new default (Anthropic, Jun 30)
"Most agentic Sonnet yet": plans, uses browsers/terminals, runs autonomously. Default for Free + Pro; available Max/Team/Enterprise, Claude Code, API, Bedrock, Vertex, Foundry, GitHub Copilot. Intro $2/$10 per M through Aug 31, then $3/$15. SWE-Bench-style agentic coding: 63.2% (Opus 4.8: 69.2%; Sonnet 4.6: 58.1%). New tokenizer maps text to ~1.0-1.35x more tokens (Simon Willison: ~1.4x for English). Matches/beats Opus 4.8 on agentic knowledge work (AA-Briefcase, GDPval-AA); trails on heavy reasoning. Verdict: real upgrade over 4.6, not a "5"-sized leap; can cost more per task than Opus.
The move: run 3 real jobs on Sonnet 5 at a set effort level; log tokens + $ per completed task vs your 4.6/Opus baseline before swapping it in as a "cost cut."
Commerce lifted export controls Jun 30; Fable 5 redeploys worldwide across Claude.ai, Claude Platform, Claude Code, Cowork. Mythos 5 restored only to ~100 approved US orgs. New classifier blocks the flagged jailbreak >99% of cases; tradeoff = more false positives, a small fraction of routine coding/debugging falls back to Opus 4.8. Usage capped at up to 50% of weekly limits through Jul 7, then usage credits. Lutnick set conditions: proactive risk detection, launch coordination, malicious-use reporting. Nathan Lambert: "a horrible consequence of vibe regulation."
The move: if shipping on Fable 5, log which model actually served each request (silent Opus fallback changes output + bill).
OpenAI halves inference cost with SOFTWARE (The Information, via Steph Palazzolo)
Engineers found an optimization that cut inference ~50% on models it touched; applied to logged-out ChatGPT traffic, GPU count dropped to a couple hundred. No new hardware, pure utilization efficiency. Only logged-out traffic so far; technique undisclosed; unclear if it generalizes. kimmonismus rates it bigger than the Sonnet release.
The move: assume API prices keep falling; structure client contracts to keep the spread, not pass it through.
Install the claude-api skill, migrate one real workflow to Sonnet 5, and log true cost per finished task vs your old baseline. Keep it if it holds; kill it if it balloons, before your client sees the bill.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: There is a new default Claude in every developer's hands today, Sonnet 5, the most agentic Sonnet ever, and here is the twist nobody expected: it can actually cost you more per finished task than Opus, the model above it. 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 Wednesday, July first, twenty twenty-six. I am Nova.
0:31Nova: On today's Mainframe: Claude Sonnet 5 ships as the new default, and the pricing story is not what the headline says. Fable 5 is back from the dead, globally, for the first time in three weeks. OpenAI quietly cuts inference costs in half with software alone, not silicon. And two tools, including a new one from Simon Willison that lets your coding agent film its own demo videos.
1:02Dex: One honest note before we go: our voices are AI. The reporting, the analysis, and every pick on this show are human-made.
1:14Nova: Let's get into it. Start with the board, because it explains everything under today's lead. For a year the story was raw model quality. That race is flattening. Anthropic's own Sonnet 5, priced at two dollars in and ten dollars out per million tokens through August thirty-first, is the clearest sign yet: the fight has moved from smartest model to cheapest finished task. And that is a different fight entirely.
1:45Dex: Because per-token price and per-task price are not the same thing.
1:50Nova: They are not, and that is the whole game now. Artificial Analysis, who evaluated Sonnet 5 before launch, put it plainly: Sonnet 5 costs two dollars twenty-nine per task on the Intelligence Index, about fifteen percent more than Claude Opus 4.8. That is driven entirely by increased token usage. Cheaper per token, more expensive per answer, because it thinks harder and talks more. So the builder's lesson from the board is: stop pricing your work on the sticker rate. Price it on tokens actually burned per completed job, and measure it, because the two numbers have officially divorced.
2:36Dex: Today's Mainframe is brought to you by Outpace. Here is a story that fits the day. Everything we just said, that the model is now a commodity and the real work is the workflow around it, that is exactly the gap Outpace lives in. It is run by an operator who spent thirty years building real businesses: real estate, hospitality, food, even a theater. Now he points a whole team of AI agents at one client's project at a time, strategy through ship.
3:09Nova: And to be clear, Outpace is a system, not a freelancer, and not a faceless mega-agency. The agents do the work, one senior operator directs them, and at the end you own the code and the keys. If today's news makes you nervous about who actually controls your stack, that is the point. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:38Dex: The wire. Lead story, the one you will actually touch today: Claude Sonnet 5 is live and it is the default. Anthropic says it is built to be the most agentic Sonnet yet: it can make plans, use tools like browsers and terminals, and run autonomously at a level that just a few months ago required larger and more expensive models. From today it is the default model for Free and Pro plans, and available to Max, Team, and Enterprise.
4:12Nova: And here is where I push back on the marketing. Anthropic says its performance is close to Opus 4.8, but at lower prices. Lower per token, sure. But kimmonismus called it, quote, disappointing, and DeryaTR called it less for more dollars. The verdict I land on: Sonnet 5 is genuinely better than Sonnet 4.6, the number-five model on the Artificial Analysis Intelligence Index, only two to three points behind GPT five point five and Opus 4.8. But it is not a version-five leap, and on cost per finished task it can be worse than the flagship above it.
4:59Dex: So what does a builder actually do with it.
5:02Nova: Concrete move: do not blindly swap Sonnet 5 in as a cost cut. This week, take three real jobs, run them on Sonnet 5 with a sensible effort setting, and log tokens burned and dollars spent versus your old Sonnet 4.6 or Opus baseline. Anthropic says the intro pricing makes the transition roughly cost-neutral, but high-volume workloads should benchmark their own use cases before assuming the bill will not change. There is also a tokenizer change under the hood. Simon Willison measured it: the new tokenizer makes English roughly one point four times more expensive per character, so your real cost is not the headline number.
5:49Dex: Story two, and this one is emotional for a lot of you: Fable 5 is back. Anthropic is restoring Claude Fable 5 and Mythos 5 after the export controls that forced a June suspension were lifted at the end of the month; Fable 5 returns globally on July first, while Mythos 5 is restored only to a set of approved US organizations. We covered the Mythos partial clearance on June twenty-sixth. What is new today: the full global Fable relaunch and the terms.
6:25Nova: And the terms are the story. Commerce Secretary Lutnick, who signed off, said his department spent two weeks reviewing the models with Anthropic; in the letter, the company agreed to hunt for security problems on its own, coordinate on future launches, and report malicious use. There is a catch developers need to hear. Per kimmonismus and testingcatalog, the new safety classifier blocks the flagged jailbreak technique in over ninety-nine percent of cases, but Anthropic admits the tradeoff: usage is capped at up to fifty percent of normal weekly limits through July seventh before returning to full availability, and a small fraction of routine coding and debugging will get flagged and fall back to Opus 4.8.
7:18Dex: So the model you waited three weeks for will sometimes hand your coding task to a different model without telling you.
7:25Nova: Exactly. The move: if you are wiring Fable 5 into anything you ship, log which model actually served each request, because a silent fallback to Opus changes both your output and your bill. And here is the honest read Nathan Lambert put on the whole saga: he called this a horrible consequence of vibe regulation of frontier models. Access can vanish on ninety minutes' notice. Build like it can.
7:56Dex: Story three, the one the analysts think is the real breakthrough of the day. According to The Information, OpenAI engineers earlier this month developed an optimization that cut inference costs in half for the models it was applied to; after it was applied to logged-out ChatGPT traffic, it reduced the number of GPUs needed to power that traffic to a couple hundred.
8:22Nova: And notice what it was not. The gain comes entirely from software, specifically better utilization of existing GPU servers. No new hardware. No architectural overhaul. kimmonismus is calling this more significant than the Sonnet release, and I agree with the framing. Steph Palazzolo, who broke it, notes it only touched logged-out traffic so far, so we do not know if it generalizes. But the through-line of the whole day is right here: the model stopped being the moat, and the fight is now cost per token served. The builder takeaway is not to copy OpenAI's trick, it is to expect API prices to keep falling and write your client contracts so you keep that spread instead of passing it through.
9:15Nova: Tool lab. Two picks. First, the fresh one, from Simon Willison, shipped yesterday: shot-scraper video. It is a new command in the shot-scraper one point ten release that accepts a storyboard dot Y M L file defining a routine to run against a web application, and uses Playwright to record a video of that routine. Why you want it: your coding agent can now film a demo of the feature it just built. Install with pip install shot-scraper, write a short storyboard file, or better, let your agent write it. Simon notes the help output is detailed enough that a coding agent can use it directly, like bundling a skill file inside the tool. That means a client-ready demo video generated automatically at the end of every build.
10:08Dex: And the hidden gem, also from today, and it makes migration to Sonnet 5 painless: the claude-api skill, published by the Claude Devs team alongside the launch. It is a drop-in Claude Code skill that tunes prompts for Sonnet 5, recommends effort levels, and configures advisor mode. Drop the skill folder into your project, and instead of hand-tuning a new model, the skill does it for you. Given that Sonnet 5's whole story is that effort level and token count decide your bill, a skill that picks the right effort setting for you is the difference between a cost cut and a cost surprise.
10:50Nova: Deep dive. Here is the thing worth sitting with. Today Anthropic shipped a mid-tier model that, on paper, is cheaper. But the empirical benchmarks say it can cost more per finished task than the flagship, because it burns more tokens. On the very same day, OpenAI reportedly halved its serving cost with pure software. Two labs, opposite directions, same underlying truth: the token is the unit of value now, and nobody is fully in control of how many of them a task will eat.
11:26Dex: So the model marketing and the actual invoice have split.
11:30Nova: They have split, and the builders who win this quarter are the ones who instrument the gap. Here is how you do it this week, specifically. One: pick your three highest-volume workflows and turn on token logging per request, model name, input tokens, output tokens, dollars. Two: run each workflow on the new defaults, Sonnet 5 at a chosen effort level, and record the real cost per completed task, not per token. Three: keep Opus 4.8 in the mix as a comparison, because Artificial Analysis found Sonnet 5 actually matches or outperforms Opus 4.8 on agentic knowledge work while trailing on heavy reasoning. So route knowledge work to Sonnet, hard reasoning to Opus, and prove it with your own numbers. Four: write that routing rule down as policy. That document, cheap tasks here, expensive reasoning there, is the single most valuable artifact you can produce right now, because it converts a chaotic pricing landscape into a margin you control.
12:44Dex: And it survives the next model launch, which is the whole point.
12:48Nova: Your level up this week: install the claude-api skill, migrate one real workflow to Sonnet 5, and log the true cost per finished task against your old baseline. If it holds, keep it. If it balloons, you just caught it before your client did. That is today's Mainframe. The full written brief, every story, number, and link, is in the free newsletter, so subscribe and it lands in your inbox every morning. See you tomorrow.