Capability at the top of the market has converged. Distribution has not. ChatGPT still takes more than half of standalone chatbot web traffic, Gemini is the only other consumer-scale assistant, and Claude now leads on coding and enterprise revenue even with a much smaller public audience. This page compares the latest flagship models, who uses them, and how much capital sits behind each lab.
Pure labs are shown at last reported private valuation. Public parents are shown at August 2026 market capitalization. These are not the same thing — Alphabet’s $4.3 trillion is Search, YouTube, Cloud, and Android, not just Gemini.
Latest primary or widely reported post-money valuation.
Anthropic’s May Series H ($65B at $965B) passed OpenAI’s March mark ($122B at $852B). Both have confidentially filed for 2026 IPOs. xAI was absorbed by SpaceX in February at a $250B AI-unit value. DeepSeek’s first external round this spring was reported in the $10–45B range; $45B is the high print.
Public parents of a frontier model or major assistant.
SpaceX listed in June (SPCX) after buying xAI. Combined-entity prints since the IPO sit roughly $1.4–2.0T; ~$1.5T is the late-July / early-August cluster. Microsoft has put $13B+ into OpenAI and additional capital into Anthropic. Alphabet, Amazon, Meta, and Microsoft are also spending on the order of $100–200B each in 2026 capex, much of it AI infrastructure.
Companies do not count users the same way. A ChatGPT monthly user, a Gemini app user, a Grok feature user on X, and a Meta AI tap inside WhatsApp are different acts. Treat these as the latest official or well-sourced figures, not a single census.
What each lab is shipping right now — not last winter’s model names. Filter by what you actually need.
The volume leader. Sol is the high-compute flagship; Terra is the everyday workhorse; Luna was cut 80% on 30 Jul to $0.20 / $1.20 per million tokens. Best general assistant most people can actually open. Sol leads several hard math and reasoning boards.
Current intelligence-index leader (Artificial Analysis v4.1: 61). Took Arena’s coding and agent boards. Same $5 / $25 price as Opus 4.8. Fable 5 / Mythos-class sits beside it for the hardest writing and long jobs. Enterprise revenue now outruns OpenAI’s last reported run-rate.
The accuracy and multimodal pick, wired through Search, Android, and Workspace. 3.1 Pro has led GPQA Diamond and cheap ARC-AGI-1 work. Google has been shipping Flash-tier speed this summer rather than a new closed flagship. Gemini Spark is the always-on agent from I/O.
Real-time X context, fewer content limits, and a 2M-token class window on the prior 4.20 line. AA index around 54 — a step behind Opus 5 and Sol, cheaper than both. Distribution is the X graph: 117M feature users, not a ChatGPT-scale destination app.
The price-performance wrecking ball. Flash 0731 is about $0.14 / $0.28 per million tokens with a mid-tier intelligence score. Pro is a 1.6T MoE (49B active) that sits within a few GPQA points of the Western flagships. Open weights, 1M context, China-heavy users.
The default open stack for everyone who will not pay OpenAI or Anthropic. Meta AI is the most-touched assistant on Earth if you count WhatsApp and Instagram taps, and almost invisible as a destination site. Llama remains the ecosystem play, not the LMSYS crown.
Not a single frontier model so much as the workplace surface. GitHub Copilot now auto-routes by task. Microsoft is also shipping cheaper in-house models to reduce OpenAI dependence. Tiny standalone web share, enormous Office and Windows footprint.
The cited-answer engine. Losing consumer web share as ChatGPT and Gemini absorbed search-like answers, still the cleanest product if you want sources first. $22.6B company on roughly half a billion of annualized revenue.
The surprise on the July board. K3 sits ahead of last-gen Western flagships on Artificial Analysis while remaining far cheaper. Evidence that the second tier of Chinese labs is now a frontier participant, not a clone factory.
Alibaba’s open line remains the other default for self-hosting beside Llama and DeepSeek. Strong coding variants, aggressive price, and a large domestic chat surface that never appears in English-language Similarweb pies.
Europe’s lab. Not winning the intelligence index, still the sovereign and on-prem option for buyers who will not send tokens to US or Chinese hosts. Le Chat is a small slice of the “Other 4.2%” in US chatbot share.
A compressed scorecard. Benchmarks move weekly; the strategic shape of the market has been more stable: OpenAI owns habit, Google owns distribution, Anthropic owns the enterprise token bill, xAI owns the X graph, DeepSeek owns the price floor.
| Lab | Latest flagship | Users | Web share | Capital | Revenue signal | Where it wins |
|---|---|---|---|---|---|---|
| OpenAI | GPT-5.6 Sol / Terra / Luna | 1B+ MAU | 53.9% | $852B private | ~$25B run-rate | Default assistant. Math, everyday knowledge work, brand. |
| Anthropic | Claude Opus 5, Fable 5, Sonnet 5 | 56M app MAU | 9.2% | $965B private | $47B run-rate | Coding, agents, writing, API spend. Highest lab valuation. |
| Gemini 3.1 Pro, 3.5/3.6 Flash | 1B MAU | 27.9% | $4.3T parent | Inside Alphabet | Science accuracy, multimodal, free consumer distribution. | |
| xAI | Grok 4.5, Grok Build | 117M feature MAU | 2.4% | $250B unit · ~$1.5T parent | ~$0.5–3B scale | Live X, fewer filters. Still a money-losing compute story. |
| DeepSeek | V4-Pro, V4-Flash 0731 | ~100M | 4.1% | ~$45B | Price leader | Open weights near the frontier at a tenth the token price. |
| Meta | Llama 4, Muse Spark | 1.2B in-app | ~0.1% solo | $1.51T parent | Inside Meta ads | Social distribution and the open-weight default. |
| Microsoft | Copilot + MAI + routed models | ~420M | 1.3% web | $3.7T parent | M365 attach | Work surface. Can switch models underneath. |
| Perplexity | Sonar family | 34–45M | 1.3% | $22.6B | ~$500M ARR | Cited answers. Losing share to the generalists. |
| Moonshot | Kimi K3 | n/a public | — | Private | — | Near-frontier scores without Western flagship prices. |
| Alibaba | Qwen 3.7 | Large in CN | Off this pie | Public parent | Cloud attach | Open multilingual models and domestic chat. |
Still the product with habit, plugins, and a billion people already there. Capability is no longer a unique advantage; the account is.
Coding and agent boards flipped to Anthropic in August. Fable 5 if the job is long and expensive. Sonnet if you are paying the bill.
Frontier-adjacent quality at $0.14 / $0.28. The reason Western labs cut Luna and Terra at the end of July.
A billion app users plus Search and Android. Best science scores among the consumer flagships earlier this summer.
Native firehose, fewer refusals, and a parent with rocket-scale capex. Not the intelligence leader; the context leader for that network.
Consumer fame and API fortune have decoupled. Claude’s user base is a fraction of ChatGPT’s; its run-rate is not.
Flagships are turning over every few weeks. Leave an address if you want a reminder when this page is updated. Your email is sent securely to the list owner and used only to send briefing updates.
We only use your address to send Frontier Brief updates. No spam, and you can ask to be removed at any time.
Compiled 13 August 2026 from public reporting. Figures disagree across trackers because they measure different things (web visits vs. app MAU vs. in-feed taps vs. API seats). Where sources conflict, this page prefers company disclosures, then audited filings, then named research firms.