Issue #7 · July 25, 2026 · ~5 min read

Frontier performance, half the price

Welcome back to Zero to MVP Weekly. Same deal: one signal, one tool, one MVP idea — five minutes, every Saturday. No link dumps. Let's go.

1. The signal of the week

Four cost stories landed in five days. The frontier labs have stopped competing on "smartest" and started competing on cost per finished task — which is the number you were quietly paying all along.

On July 24, Anthropic released Claude Opus 5 at $5 / $25 per million tokens — unchanged from Opus 4.8, and half of Fable 5's $10 / $50. On Anthropic's own Frontier-Bench v0.1 it scores 43.3%, against Opus 4.8's 18.7% and Fable 5's 33.7%. The company's framing is entirely economic: on CursorBench 3.2 it lands "within 0.5% of Fable 5's peak score, but at half the cost per task," and on OSWorld 2.0 it beats Fable 5's best result "at just over a third of the cost." It ships with a 1M-token context window, in the API as claude-opus-5, and in Claude Code.

Three days earlier, Google shipped Gemini 3.6 Flash at $1.50 / $7.50, claiming 17% fewer output tokens than 3.5 Flash and a jump on DeepSWE from 37% to 49%. On the 24th, Microsoft put two in-house models into public preview under Satya Nadella's "frontier diffusion" doctrine — route traffic to your own models wherever they match frontier quality — reporting GPU costs down up to 89% in Dynamics 365 Contact Center. And on the 21st, Poolside released Laguna S 2.1, a 118B open-weight model activating 8B parameters per token, scoring 70.2% on Terminal-Bench 2.1 — ahead of DeepSeek-V4-Pro-Max at 1.6 trillion parameters, and ahead of Inkling at 975 billion.

Here's what ties them together. Every one of these launches is priced and benchmarked in cost per completed task, not tokens per dollar and not points on a leaderboard — because that's the number that decides whether an agent product has a margin. A model that costs twice as much per token but finishes in a third of the turns is cheaper. This is the metric flip that makes agent loops viable at small scale. Last quarter the honest advice was "an agentic workflow will eat your budget"; this quarter four vendors shipped independently against exactly that complaint. The practical consequence for you: model choice is now a measurement problem, not a reading problem. Whatever you picked three months ago on vibes is probably wrong on price today, and the only way to know is to run one of your real tasks through two models and count dollars-to-done.

One caveat to keep us honest: almost every number above is vendor-reported, on benchmarks the vendor designed. Frontier-Bench v0.1 is Anthropic's own, and each lab defines "cost per task" for itself. The one independent read cuts the other way — CodeRabbit found Opus 5 makes "a stronger case as a builder than as a reviewer," reading about 50% more and writing about 65% more per review call than baseline frontier models. Thorough, but token-hungry on that particular workload. Efficiency is workload-shaped. The only benchmark that settles it is yours.

2. My take

The price of capability is falling fast enough that you should build for it. This is the fourth significant drop in the price floor in the six issues of this newsletter, and it has changed how I write things. I'm increasingly convinced the model should be just another parameter — one you can swap when the numbers say so. One thin seam, with prompts and evals in git.

But go easy on that. Premature abstraction is real, and model-agnostic code is code written to the lowest common denominator: you give up what each model is sharpest at in order to be compatible with all of them. So the split I'd make today is try now (one seam), prototype only (automatic routing), skip (a framework before PMF).

3. Tool worth trying

Claude Opus 5.

I've already had a chance to use it, and the impression is very good. For me it replaces Opus 4.8 rather than competing with Fable 5 — I treat Fable as the model for more creative work, and reach for Opus when it's code.

Yesterday I put it on the codebase of my own password manager, Kivarion, and together we shipped a batch of improvements. You can judge the work yourself: every commit to that repo from yesterday was written with Opus 5.

And at half of Fable 5's per-token price, with a 1M-token context window and an optional fast mode at 2.5× the speed for 2× the price, the cost per finished task comes out very good indeed.

4. MVP idea of the week

Cost-per-task receipts for your agent runs.

Problem: four vendors now advertise "cost per task," each measured on a benchmark they designed, and no builder can check any of it against their own work. You know your monthly API bill. You don't know what a finished feature costs you — or which model is cheapest at actually finishing one.
Solution: a wrapper that treats every agent session as a billable job — capturing task, model, tokens in and out, cache hits, wall-clock, turns, human interventions and the final outcome (tests pass / merged / abandoned) — and prints a per-task receipt plus a rolling "cost per completed task, by model" table.
For whom: solo builders and small teams running agents daily; anyone about to switch models on a launch-week claim; consultants who need to price agent-assisted work honestly.
Minimal v1: a hook or wrapper writing JSONL, plus a receipts command that prints the table. Two models. No dashboard, no cloud.
Demand check: publish your own three-month table — what a shipped feature actually costs you, by model, in real dollars. Then watch which people ask for: your numbers, or your harness. Only the second one is a product.
Why now: the metric the whole industry adopted this week is one nobody can independently verify, and switching costs are low enough that being right about it is worth real money. Related: Stripe is reportedly in talks to buy OpenRouter for around $10 billion — the routing layer is consolidating into a payments company, which makes vendor-independent measurement more valuable, not less.

5. Useful links

Anthropic — Introducing Claude Opus 5 — the primary for this week's signal; read it as a pricing document, and note which benchmarks are Anthropic's own.
CodeRabbit — Claude Opus 5 for code review — the independent counterweight: where the efficiency claim holds up and where it doesn't.
Google — Gemini 3.6 Flash, 3.5 Flash-Lite, 3.5 Flash Cyber — the other half of the efficiency turn, with prices and token-reduction numbers stated plainly.
Poolside — Introducing Laguna S 2.1 — a capable open-weight coding model, plus every raw evaluation trajectory published for anyone to audit. That's the standard to hold other labs to.
Microsoft AI — Introducing MAI-Image-2.5-Pro and MAI-Voice-2-Flash — the fourth vendor in the pattern, and the clearest statement of the doctrine behind all of it: route to the cheaper model wherever it matches. Read it as the routing policy you should be writing for your own stack.