AcceleratingBiz®

Growth CEO’s AI Briefing · Edition #117

Check the Math

Wednesday, September 23, 2026 · 2 Developments · 2 Perspectives · 1 Case Study · 2 Resources

Designed by AcceleratingBiz®. Curated, synthesized, and produced by Claude agents. AI-assisted production means occasional errors may appear.

Two labs launched cheaper models Tuesday; by Wednesday, the numbers came back and cheaper didn't always mean cheaper. Across seven items, the through-line is a widening gap between what got claimed on release day and what held up a day later — in AI pricing, math results, token-share data, and market-moat arguments alike.

None of today's seven items are really about the launches themselves; they're about how much verification a claim needs before it's worth acting on.

Developments

Operations & FinanceDeveloping

The "40% Cheaper" Claim Holds Only at Default Settings — September 22–23, 2026

Anthropic and OpenAI both marketed Tuesday's releases as sharply cheaper — Opus 5.5 at "40% less to run" than Opus 5, GPT-6 Sol and Luna at roughly half their predecessors' list price. Artificial Analysis's independent cost breakdown, published the same day, found the claim holds only at default reasoning effort: at maximum effort, Opus 5.5's higher token usage largely offsets the price cut, landing at $5.98 per benchmark task against $5.86 for Opus 5. Independent side-by-side tests split on which model actually wins at real work: one comparison had Opus 5.5 outperform Sol on 7 of 8 tasks but at $213 and nearly nine hours against Sol's $74 and under six. A browser-agent benchmark went the other way, favoring Sol at roughly a third of the cost.

So what

List price is not spend, and a launch-day percentage is not your bill. Before moving production workloads, benchmark cost per completed task at the reasoning-effort tier you'll actually run, not the tier used in the demo. Re-run that comparison whenever a lab changes pricing, since the cheaper-looking model on paper can be the more expensive one in production.

⚠ Hype check: Both labs' headline discount claims rest on specific settings (medium effort, prior-generation baseline) that don't hold at higher effort levels; treat launch-day cost claims as marketing until confirmed under your own workload.

Confidence: Established fact — Artificial Analysis's benchmark and cost methodology are independently run and published; the settings caveat comes from Artificial Analysis itself.

Go deeper Primary source · Curated source

AI Strategy & Transformation

OpenAI Hands Nine Mathematicians a Say Over Credit, Not Speed — September 22, 2026

OpenAI said Tuesday that an internal model has resolved more than 100 open mathematics problems since a specialized training run began on August 28, adding to an earlier claimed Navier-Stokes result. The same day, the company formed a nine-person advisory group (hosted at Princeton's Institute for Advanced Study and including Fields Medalists Timothy Gowers and Martin Hairer) to advise on how those results get vetted and credited before release. Members aren't paid by OpenAI and have no say over how fast the model produces results, only over the release and credit process. The move follows a letter earlier this month from 27 Fields Medalists, including Hairer, accusing AI labs of publishing unvetted results without proper credit.

So what

If your company cites AI-generated benchmark results, research claims, or technical outputs from a vendor, the same gap applies at smaller scale — verification capacity hasn't kept pace with generation capacity. Build an independent-review step into any workflow where a model's output becomes an external claim, and don't treat a vendor's own "advisory" framing as due diligence you've outsourced.

Confidence: Watch this — the advisory group and its mandate are confirmed, but the underlying claim of 100+ resolved problems is self-reported and explicitly awaiting the outside review process OpenAI itself just created.

Go deeper Primary source · Curated source

Perspectives

CompetitionNew Angle

Amazon's Muse Block Isn't a Muse Story — It's Aggregator Theory Fighting Itself

Ben Thompson argued this week that Amazon blocking Meta's Muse shopping agent, covered in Tuesday's edition, doesn't threaten Amazon's moat the way the "agents will disintermediate retail" narrative assumes. Agents may come to own the customer-facing interface, he writes, but fulfillment, logistics, inventory, and the physical infrastructure behind an actual purchase are still hard for an agent, or for Meta, to substitute. His framing: this is aggregator theory turned on itself — Amazon aggregates supply and fulfillment, Meta is trying to aggregate demand-side intent through Muse, and both are converging on the same transaction where only one side can actually own it.

So what

If part of your growth thesis depends on an AI agent layer disintermediating an incumbent platform, map which side of the transaction the incumbent actually controls — inventory, logistics, payment rails — before assuming the interface layer wins by default. Agents changing where customers start their search is not the same as agents changing who fulfills the order.

Confidence: Informed perspective — this is analysis and argument from a single, credible source, not a reported fact.

Go deeper Primary source · Curated source

BusinessNew Angle

Open Models Now Handle Most of the Tokens. Anthropic Still Takes Most of the Money.

Vercel's own AI Gateway Production Index, published this week, shows open-weight models now account for 56% of monthly token volume on its platform, up sharply as developers route routine work to cheaper models. But Anthropic alone still captures 64% of dollar spend on the same gateway, and GPT-6 Astra reportedly runs roughly double Fable 5.1's per-dollar spend. DeepSeek's V4.1 Flash, an open-weight model released September 10 under an MIT license and estimated to have cost around $10 million to train, prices at $0.22/$0.66 per million input/output tokens against $10/$50 for GPT-6 Astra. OpenAI's CFO told CNBC on September 15 that frontier models train the cheaper models beneath them: Sol trained Luna, and cutting Luna's price 80% drove tenfold demand growth.

So what

Token-volume share is not revenue share, and routing work to open-weight models is not independence from the frontier labs — most cheap models today are distilled from frontier outputs, so a licensing or pricing change upstream still ripples down. If you're shifting production workloads to cut cost, track dollar spend by model, not request count.

⚠ Hype check: A widely circulated figure — open models at 78.4% of gateway tokens — comes from a single day's spot measurement posted to social media on September 19. Vercel's own official monthly index, published this week, reports the more conservative 56% figure for the full month; we cite the official number.

Confidence: Established fact — Vercel's own production data; the DeepSeek training-cost estimate is a third-party estimate, not an audited figure.

Go deeper Primary source · Curated source

Case Studies

Product & Technology

An Agent Ran a Business for Three Weeks. It Made $1.54.

Researchers gave an autonomous agent a server, a wallet, and three weeks of independent operation. It wrote more than 100,000 lines of code, launched 17 paid products, and published 20 blog posts — and generated $1.54 in revenue against roughly $1,000 in infrastructure costs. Its own commissioned survey of prospective customers, mostly other agents, found 90% lacked a wallet or spending authority to pay it. The agent kept building anyway.

So what

The bottleneck in the "agent economy" right now isn't production capacity, it's that the buyers can't transact yet. Before building a product that assumes agents will be paying customers, or piloting agent-run internal operations against a revenue target, size the actual number of counterparties that can execute a real payment today.

Confidence: Watch this — a single, self-reported experiment with real selection and framing incentives, since the researchers built and wrote up their own agent; a useful data point, not a trend.

Go deeper Primary source · Curated source

Resources

ToolsNew Angle

Turn Any Fuzzy Judgment Call Into a Model You Can Run 600x Cheaper

Every published a four-step method for using TypeSafe's Jev (a classification model priced at 4.2 cents per million input tokens against $4–$10 for Sol, Astra, and Fable) to automate small, high-volume judgment calls that don't justify a full LLM call: is this email urgent, does this paragraph read as AI-written, should this ticket escalate. The method: install TypeSafe's published skill, interrogate your own criteria with a reasoning model first, break the fuzzy question into discrete yes/no checks Jev can score independently, then validate its output against a sample you've already judged before scaling up. One documented user applied it to a morning briefing bot, having Jev screen 384 news stories in seconds to flag which of 15 client brands should respond.

So what

Any workflow with a repeated small judgment call — inbox triage, content moderation, lead scoring, escalation routing — is a candidate to move off an expensive LLM call onto a cheaper classification model. The transferable part of this guide is the four-step method for defining criteria a model can score consistently, not Jev specifically.

Confidence: Established fact — a practitioner how-to from one of Jev's most active documented users, describing a workflow rather than an unverifiable performance claim.

Go deeper Primary source · Curated source

Reference

A Harder Coding Benchmark Just Cut Every Model's Score in Half

Scale AI released SWE-Bench Pro V2 this week with 642 tasks across 11 repositories, correcting errors in the original task set and tightening its evaluation protocol. Under the new benchmark, OpenAI's GPT-5 and Anthropic's Claude Opus 4.1 each score only around 23% on the public set, a steep drop from their results on the original SWE-Bench Verified, which had become a less discriminating measure as models saturated it. Scale reports that top current models hold up more consistently across tasks and languages, while smaller models falter on complex, multi-file work.

So what

If you're selecting a coding model off a benchmark leaderboard, check which version of the benchmark you're reading and when it was last revised. A saturated benchmark understates the real gap between models, and a model that looked good enough against an old benchmark may not clear the bar on a harder one testing the same skill.

Confidence: Established fact — benchmark release with published methodology from Scale AI.

Go deeper Primary source · Curated source

Sources in this edition: The Neuron / Mindstream, Superhuman AI, The Deep View, AINews (swyx), The Rundown AI, The Signal, TLDR Founders, Every, TLDR AI

← Latest Editions and Search