The article warns against wasteful token use: the top models from Anthropic and OpenAI cost 100 times more per token than their smallest ones, yet many users default to the flagship for everything. It cites Meta employees who burned 60 trillion tokens in 30 days chasing an internal leaderboard, and Uber, which used up its 2026 AI tools budget within a few months. The article recommends choosing smaller models for simple tasks, writing one clear prompt instead of several back-and-forth rounds, and using techniques such as trimming context, caching and batching, noting that switching simple jobs to a small model can cut API bills by up to 99%. It also reports that Anthropic's revenue run rate rose from about $9 billion at the end of 2025 to more than $47 billion by May.
Summary written by Upside AI from the original article. Read the source for full details.
Read the original on Decrypt ↗
Keywords
artificial intelligenceAnthropicOpenAI
Open a keyword to see other news on the same topic.
Related news
- OpenAI, Anthropic's AI Lead Is 'Very Perishable,' IMD Professor Says: 'You Don't Need to Drive a Ferrari to Go to Walmart'
IMD professor Howard Yu says OpenAI and Anthropic may stay ahead of Chinese rivals, but a lead of roughly four to six months is 'very perishable' because most use cases don't need top-tier capability. He notes many companies use AI for simple tasks and weigh token spending against hiring, while cheaper Chinese open-source models can run inside a company's own data centers, a demand especially strong in Europe. Yu believes OpenAI and Anthropic will not be easily displaced in Western consumer markets thanks to easy access and clean interfaces, though enterprise buyers are more sophisticated. Polymarket traders still see U.S. labs leading, with Anthropic at about 43%, Alphabet at 42%, OpenAI near 8%, and Chinese labs around 1% or less to have the best AI model at end-2026.
- Will AI Companies Generate Enough Revenue to Continue the Build-Out? Here's What Nvidia CEO Jensen Huang Says.
AI stocks fell sharply in October amid higher interest rates and doubts about returns on AI spending. Back-to-back leaks suggested slowing growth: OpenAI reported $50 billion in ARR, versus a $70 billion figure leaked in September that was later attributed to accounting differences, while Anthropic reported about $16.8 billion in Q3 revenue, or $67.2 billion annualized, below the $65 billion ARR reportedly reached by end-July. On a podcast with Brad Gerstner, Nvidia CEO Jensen Huang argued that other frontier labs, including xAI, Meta and Google, are growing too, and that frontier labs account for only 25% to 33% of Nvidia's chip demand. Axios reported OpenAI's revenue run rate grew roughly 70% in Q3 versus 46% quarter-over-quarter growth for Anthropic, which the article reads as share loss rather than a broad AI slowdown.
- Jensen Huang: Token Costs Are Falling “30, 40 Times per Year.” That’s Nvidia’s Plan
Nvidia CEO Jensen Huang says the cost of an AI token is falling 30 to 40 times a year and wants it to keep falling, since rising usage should outpace lower prices and boost demand for Nvidia's chips. OpenAI CFO Sarah Friar said an 80% price cut on its cheapest model series lifted usage 10x, which the article estimates roughly doubled revenue on those models. Analysts expect Nvidia's revenue to nearly double from $216 billion to $412 billion in the year ending January 2027. The risk falls on token sellers and heavily indebted neoclouds: an Epoch AI report cited in the piece shows token costs falling about 13-fold a year, and CoreWeave dropped 15% on Oct. 14 as its bond yields neared 15%. Nvidia's next quarterly report will be an early test after that selloff.
- “Suddenly Meta Is in Second Place” in AI, and Unlike OpenAI It Doesn’t Need the Money
Meta has launched Muse Wonder and Muse Wonder Full, which it calls its most capable AI models yet, and Dragonfly's Haseeb Qureshi said the release puts Meta in second place. Meta has not set API pricing yet and the models are only available within its own products, while the article argues that ad revenue lets Meta price aggressively in a way OpenAI and Anthropic cannot. Meta's operating margin was 41.4% in 2025, and its $116 billion in operating cash covered $69.7 billion of capex, but 2026 capex guidance of $115 billion to $135 billion could absorb most of that cash. The article also cites a Financial Times report that OpenAI expects to burn $278 billion of cash from 2026 through 2030. The main uncertainty is whether Meta's eventual pricing will be low or premium.
- Accenture Expands Federal AI Push With Google's Gemini Agents, Months After OpenAI Partnership
Accenture and Google Public Sector announced Gemini launchpads to help U.S. federal agencies move AI agents from experimentation into operations, integrating them with existing government systems and datasets. A centralized team of forward-deployed engineers will work alongside government staff, and pre-built agents are being developed for human resources, finance and supply chain tasks. The announcement did not name federal customers, disclose contract values or give revenue projections. The initiative builds on the Gemini Enterprise Business Group, where Accenture committed to 1,000 forward-deployed engineers, and follows Accenture Federal's May partnership with OpenAI. So far this year, ACN stock has fallen about 11% while GOOG has risen nearly 7%.
- Breakfast News: The Sunday Edition
Nvidia CEO Jensen Huang posted in September that "AGI has arrived," while many other tech leaders are more cautious about when AGI will be reached. AGI is described as AI capable of a skilled worker's performance in most jobs, whereas today's models still need supervision and cannot work reliably over long stretches. METR research found that top models succeed only about half the time on tasks that would take a person two workdays. Sam Altman says OpenAI is "not quite yet" at AGI but expects an internal system he would call AGI by year-end, while Sundar Pichai expects it shortly after 2030. Andy Jassy has not set a date, but AWS customers have signed $496 billion in contracts for future services.
Related markets
Latest news
- Amazon (AMZN) Stock Costs $250 a Share. Here's Why I'd Still Buy It.
The article argues that Amazon, trading at about $250 a share, is still worth buying for the author, who would buy it if he did not already own it. The main reason is that the stock looks appealingly valued: its forward P/E of about 24 is well below its five-year average of 39, while its price-to-sales ratio of 3.6 is slightly above the five-year average of 3. Amazon ranked fourth in the 2026 Kantar BrandZ Global Top 100, with a brand value of $1 trillion, and its shares have averaged annual gains of 22% over the past 15 years. The author stresses that a stock's absolute price means little on its own, so investors should compare it with earnings and sales and research each company in depth.
- Amazon Just Made Its AI Chips Easier to Use. Should Nvidia Investors Be Worried?
Amazon announced TorchNeuron, a free open-source tool that lets developers run existing PyTorch code on its Trainium chips with minimal changes, with a public beta due on Oct. 20. This challenges a key part of Nvidia's moat, because switching away from CUDA typically carries large costs in time and labor. However, the tool only covers PyTorch work, custom CUDA code will not port easily, and the release is still in beta, with the ease-of-switching claims coming from Amazon itself. Google is also developing a similar tool, TorchTPU. The author sees it as a chink in Nvidia's narrative and does not expect a major near-term effect on order flow or pricing power, but says the market may no longer pay quite the same premium for Nvidia stock.
- Cisco Systems (CSCO) Unveiled 102.4T AI Networking Platforms
Cisco Systems unveiled new AI networking platforms built on Silicon One P200, G300 and G302 switching chips, with 51.2T and 102.4T systems. It also introduced 1.6T optical modules and integrated liquid cooling aimed at high-density AI data centers, and disclosed a collaboration with NVIDIA, TE Connectivity and Amphenol to co-develop a 400Gb/s per lane USFP optical module and open specification. The company has a market cap of US$423.4b. The article says the launch supports Cisco's position as an AI networking provider competing with Arista and Juniper, but the richer hardware mix adds to margin pressure that has already contributed to non-GAAP gross margin compression.
- J.P. Morgan Stays Top of Extel’s All-America Equities Research Table
Extel's 55th annual All-America Research Team survey ranks J.P. Morgan first in the research team table with 47 published positions, followed by BofA Securities with 44 and Morgan Stanley with 43. UBS posted the largest gain in positions, reaching 42 and climbing to fourth place. In the analyst leaders table, Jefferies takes first place with 43 positions, while UBS leads Extel's weighted ranking of research teams ahead of J.P. Morgan and Evercore ISI. The results reflect the views of 4,451 portfolio managers and analysts at 1,608 institutions across 60 sectors, with 129,919 votes counted. The number of participating institutions fell from 1,705 in 2025, but more firms now sit in the highest commission bands, rising to 89 from 82.
- Visa, Mesh partner to enable crypto spending at 175M merchant locations
Visa and crypto payments network Mesh have partnered so users can spend cryptocurrency and stablecoin balances directly from participating digital wallets at merchants on Visa's network, without transferring funds to a bank account or using a separate crypto card. Merchants receive payments through Visa's existing network without changing their checkout systems. The network covers more than 175 million merchant locations in over 200 countries and territories, and participating wallets and launch dates will be announced later. Mesh will handle funding and conversion between crypto, stablecoins and fiat, as well as onboarding and KYC requirements. The partnership follows Mesh's September integration with Bybit Pay and its $75 million raise in January at a $1 billion valuation.
- Top 3 Founder Led Stocks To Watch In October 2026
The article presents three founder-led stocks from a screen: monday.com, Palantir Technologies and Figure Technology Solutions. monday.com has revenue of about US$1.37b and a market cap of US$3.8b. Palantir earns US$3.21b from Government customers and US$2.95b from Commercial clients, with a market cap of US$447.04b. Figure Technology Solutions reports US$619 million of revenue and a market cap of US$6.1b. The article argues that founder ownership can align leadership with long-term shareholder interests, but each stock depends on key assumptions to watch, such as customer expansion, pricing power or the pace of blockchain adoption. The full screen surfaced 26 more businesses with similarly strong founder ownership.
Get the briefing on Telegram or Zalo
Join the Upside Telegram channel or Zalo group for the Vietnam stocks and gold briefing at 8:00 and 19:00 (Vietnam time) on trading days, plus a weekly wrap on Saturday morning.