Consumer agent wars: the scoreboard after week one

The consumer agent market just played its first full season in ten weeks. OpenAI shipped ChatGPT Work on July 9, Meta shipped Muse on September 8 and took the #1 free US App Store slot in twelve days, Apple shipped Siri AI on September 14 inside iOS 27, and Google's Gemini Spark has been live since May. Four capability stories, told in keynotes; here is the other document: the scoreboard. Who has distribution, who has capability, where the checkout lands, and which cells are blank. The pattern is simple to state and expensive to unlearn: this market is a distribution race wearing a capability costume.

• Distribution: Muse crossed 2.5M downloads in two weeks and hit #1 free on the US App Store (Fox Business, Sept 22). ChatGPT sits near 1B weekly users (The Information, July 29). Gemini's app claims 900M+ monthly users (Google, May 19). Siri AI ships to every Apple Intelligence-capable iPhone, adoption unstated.
• Capability: the honest spread is enormous. Stanford's AI Index puts OSWorld task success at 66.3% (April). The OSWorld record is now 90.2% (announced July 27). On real white-collar tasks, the best model managed 24% first-attempt success (Mercor's APEX-Agents benchmark, via TechCrunch, Jan 22).
• Checkout: Stripe's Agentic Commerce Protocol underwrites ChatGPT (Sept 2025). PayPal signed Microsoft (Jan 8), Google (Jan 11), and Meta (Sept 22).
• Price: Muse is free. Gemini Spark requires $19.99 or $99.99 per month (PCMag, July 24). ChatGPT Work launched on top tiers first.
• The blank cells: agent weekly actives and agent revenue. They are the numbers that decide this market, and nobody publishes them.

the scoreboard

player	agent surface	distribution (latest, dated)	capability bet	checkout rail	price
Meta	Muse	2.5M+ downloads in two weeks, #1 free US App Store (Fox Business, Sept 22)	personal agent: one VM per user, phone calls, connectors	PayPal (Sept 22)	free
OpenAI	ChatGPT Work	~900M weekly users (OpenAI, Feb 27); near 1B (The Information, July 29)	work platform: chat plus Codex on GPT-5.6	Stripe ACP (Sept 2025)	top tiers first
Google	Gemini Spark	900M+ monthly Gemini users (Google, May 19); Spark in 160+ countries (Thurrott, July 31)	Workspace and Chrome agent on Gemini 3.5	PayPal via UCP (Jan 11)	$19.99 or $99.99/mo
Apple	Siri AI	every Apple Intelligence-capable device from Sept 14 (Apple); no adoption figure	on-device beta, English first	none announced	free with hardware
the field	Claude Cowork (Jan 2026)	Claude at 56M monthly app users (Sensor Tower via Reuters, June 2)	professional multi-step work	none published	n/a

The capability column looks differentiated; the distribution column says OpenAI's base dwarfs the consumer field and Muse took the quarter's only new #1.

distribution won week one

The Muse numbers, in order. 730,000 downloads in the first five days (Sensor Tower via CNBC, Sept 21). #1 free on the US App Store by September 18, ahead of ChatGPT (CNBC). 2.5M cumulative through September 21 (Sensor Tower via ChosunBiz). 642,000 US daily active users at two weeks old, against 231,000 for ChatGPT at the same age (Apptopia via TechCrunch, Sept 21). 1.8M iOS downloads across the US and Canada in twelve days, against 1.3M for ChatGPT's opening (Apptopia via Mint, Sept 22).

Two honesty notes. ChatGPT's 2022 launch did 3.1M downloads over the same post-launch window (Sensor Tower via ChosunBiz, Sept 22), so Muse hit #1 with fewer downloads than the 2022 record run: that says more about the 2026 chart than about Meta. And the estimators disagree: Appfigures had Muse crossing 1M when Apptopia counted 1.8M iOS.

OpenAI needed three years to assemble 900M weekly users (OpenAI, Feb 27); The Information reported in late July that it is nearing 1B, seven months behind its internal target (via TNW, July 29). Google's Gemini app crossed 900M monthly users by I/O (Google, May 19), but its agent, Spark, launched behind the $99.99 AI Ultra plan, reached the $19.99 tier on July 24, and hit 160+ countries at month end, the EEA, UK, and Switzerland excluded (PCMag, May 30 and July 24; Thurrott, July 31). Apple shipped Siri AI on September 14 to a floor of iPhone 16 and 15 Pro devices, English only, five more languages promised in October (Apple Newsroom; The Verge; MacObserver, Sept 9). The defaults channel ships silently; no adoption figure ever follows.

capability is the costume

Every keynote above sold capability. The published numbers tell a stranger story.

Stanford's 2026 AI Index put agent task success on OSWorld, a benchmark of real computer tasks, at 66.3%, up from roughly 12% a year earlier: still one failure in three (April 2026). In July an agent called Z-Agent pushed the record to 90.2%, the first entrant past 90% (GlobeNewswire, Aug 5). Those are structured benchmarks, the friendly case. Mercor's APEX-Agents benchmark ran 480 real white-collar tasks from banking, consulting, and corporate law: the best model succeeded 24% on the first attempt, 40% within eight (TechCrunch, Jan 22). On the tasks consumers want delegated, the frontier is a coin you lose three times in four.

The cost side moves faster than the success side: the smallest GPT-5.6 model completes these tasks about as well as the largest, at one-fifth the cost (Max Weinbach, Creative Strategies, via Reuters, July 9). Capability per dollar is compounding; capability per task is crawling.

The tell that the frontier is not fully machine: the capability Muse demos most vividly, the phone call, is in some cases completed by a human contractor (Reuters, Sept 22). That reveal is analyzed in our concierge piece; here it earns one line: the cell labeled autonomous task completion is, at the margin, still staffed.

the checkout is the prize

Where does value land? Follow the payment integrations, the only votes cast with money.

OpenAI moved first: Instant Checkout inside ChatGPT, on the Agentic Commerce Protocol built with Stripe, open sourced, merchants paying a fee on completed purchases (OpenAI, Sept 29, 2025). PayPal then spent nine months signing everyone else's surface: Microsoft's Copilot Checkout (Jan 8), Google's Universal Commerce Protocol (Jan 11), a merchant-discovery acquisition in Cymbio (Jan 22), and finally Muse (Sept 22), where PayPal customers check out inside Meta's agent.

Two rails, four fronts: the payment partners are bidding to sit under every agent, and the loop owner picks the rail while the rail takes a fee. The task-completion loop is the asset: it sees intent first, holds the context, and presses the buy button. Distribution wins the install; the loop wins the checkout.

where the value lands

Score the quarter on the two indexes that exist. Attention per install: Muse's 642K US daily actives at two weeks against ChatGPT's 231K at the same age (Apptopia via TechCrunch, Sept 21) says the free agent is winning habit formation, not just downloads. Capability per dollar: free Muse, Spark at $19.99 to $99.99, Work top-tier-first. The consumer front door is being given away while the workflow layer is sold: Meta is buying the audience OpenAI already has and Google rents by subscription.

The compounding asset is the loop. Distribution gets a trial, capability keeps the task, the checkout monetizes it, and the data makes the next task better. Week one says distribution is the scarcest input: capability converges because everyone licenses comparable models, but an installed base and a daily habit do not transfer. Muse proved reach can be manufactured in twelve days; whether it compounds is what the blank cells would answer. Nobody publishes agent weekly actives or agent revenue: the scoreboard's loudest empty row.

the fine print

Sources, with dates. Muse figures: Fox Business (Sept 22) and CNBC (Sept 21) carrying Sensor Tower estimates; Apptopia via TechCrunch (Sept 21) and Mint (Sept 22); competitor comparisons via ChosunBiz (Sept 22). The estimators disagree, so read Muse numbers as ranges. OpenAI user figures: company statements (Feb 27) and The Information on internal figures (July 29). ChatGPT Work, GPT-5.6, and the Weinbach quote: Reuters (July 9), VentureBeat (July 10). Google: company statements (May 19); Spark availability via PCMag (May 30, July 24) and Thurrott (July 31). Apple dates: Apple Newsroom and The Verge (Sept 14); Apple publishes no AI adoption figures. Benchmarks: Stanford AI Index 2026 (April), Z-Agent via GlobeNewswire (Aug 5), APEX-Agents via TechCrunch (Jan 22); Mercor sells into the market it benchmarks. The Siri-Gemini report is the Daily Star's (June 9), unconfirmed by Apple. The concierge finding is Reuters's (Sept 22). Blank cells: Apple adoption, Muse engagement beyond the estimates above, agent revenue, Muse availability outside North America.