- OpenRouter's rankings, checked October 20, 2026, show its top model, DeepSeek V4.1 Flash, down 27% week over week at 29.6 trillion tokens, while its new 1017 version debuted at 7.34 trillion. The page shows no all-model weekly total, and OpenRouter is one router, not the market.
- The '84% of tokens on cheap models' figure comes from Tomasz Tunguz's August 14, 2026 analysis of one week of OpenRouter data. He found the share outside the frontier had stayed at about 82–85% since November 2025, so it's not a new shift.
- Gartner said on October 19, 2026 that only 13% of 2,002 CIOs surveyed from May to July report significant or transformational value from AI. A month earlier, Gartner forecast AI spending of $2.67 trillion in 2026, up 49.5%.
Contents
Is AI usage slowing? Two data points published on October 19, 2026 made it look that way. A Hacker News post titled "Token-maxxing is dead, OpenRouter's token usage plateaus" pointed at OpenRouter's public rankings. The same day, Gartner reported that only 13% of CIOs say AI has delivered significant or transformational value. Read closely, neither shows overall AI use falling. They show companies watching costs and a lot of traffic moving to cheaper models below the frontier, which is a different story.
This analysis is based on OpenRouter's rankings and homepage, Tomasz Tunguz's analysis of OpenRouter data, Gartner's press releases, InformationWeek's report from Gartner's IT Symposium, and earlier 9to5AI reporting, all checked October 20, 2026. The cheap models in question are compared in our guide to cheap, fast AI models.
The short version
- OpenRouter's page doesn't show a total decline. It shows per-model weekly changes. The biggest drop, DeepSeek V4.1 Flash at −27%, came the same week its updated 1017 version took 7.34 trillion tokens.
- The "first weekly decline in about a month" claim comes from a post on X by Brandon Carl, seen in search results. The post couldn't be opened, and OpenRouter's public page doesn't show the total it describes.
- The 84% figure is old and narrow. It's Tunguz's August estimate of tokens going to models that aren't state of the art, and it has been roughly 82–85% since late 2025.
- Gartner's 13% measures value, not usage. The same survey says 58% of CIOs face more pressure to cut AI costs. Gartner's own forecast still has AI spending rising 49.5% in 2026.
- Other signals point up. Anthropic's reported $16.8 billion quarter and TSMC's $46.7 billion quarter both grew strongly. Accenture's $29 Claude cap shows the cost discipline.
What does OpenRouter's data actually show?
OpenRouter is a service that routes API calls to hundreds of models from dozens of providers. Its rankings page publishes token counts for the models that pass through it. As of October 20, the Top Models table covered usage "through Oct 19, 2026."
| Model | Weekly tokens | Change vs prior week |
|---|---|---|
| DeepSeek V4.1 Flash | 29.6T | −27% |
| GLM 5.3 Flash | 11.9T | −6% |
| MiMo-V2.6-Flash | 9.57T | −17% |
| Step 5 Preview | 8.22T | −32% |
| Hy4 preview | 7.82T | −6% |
| DeepSeek V4.1 Flash 1017 | 7.34T | new |
| Claude Opus 5.5 | 6.47T | +21% |
| GPT-6 Luna | 6.23T | −12% |
| GLM 5.3 | 5.29T | +33% |
| DeepSeek V4 Flash 0731 | 4.73T | +1% |
Table 1DeepSeek V4.1 Flash's 27% drop came in the week its 1017 version debuted at 7.34T, while Claude Opus 5.5 and GLM 5.3 grew. Source: OpenRouter rankings, data through October 19, 2026.
Last verified: October 20, 2026 (OpenRouter rankings; data through October 19)
Three things stand out:
- Much of the "decline" is a version switch. DeepSeek released V4.1 Flash 1017 on October 17, partway through the week (our coverage). If the older model fell 27% to 29.6T, it carried about 40.5T the week before. Add the new version's 7.34T and the pair carried about 36.9T. That's 9to5AI's illustrative calculation from rounded figures, but it suggests a dip of under 10% for DeepSeek's Flash line, not 27%.
- Frontier models gained. Claude Opus 5.5 rose 21%, and Anthropic's share of text requests rose 25% to 4.7% in the week beginning October 12. DeepSeek led requests at 24.4%, followed by Google (19.7%) and OpenAI (18.5%).
- The month looks like growth. On the page's trailing 30-day view, DeepSeek V4.1 Flash processed 124T tokens, up 582% from the previous 30 days. OpenRouter's homepage advertises "600T+" monthly tokens, without a date or method.
The page shows no total across all models, so one weekly dip can't be confirmed or measured from it. In the Hacker News thread (Tier 3, community discussion), commenter simonw noted that OpenRouter covers only companies that chose to use it, skews toward open-weight models, and doesn't disclose user counts, so a few heavy users can move the rankings. Others pointed out that teams can drift to direct APIs or their own harnesses.
Where does "84% of traffic on cheap models" come from?
From venture capitalist Tomasz Tunguz's August 14, 2026 post, not from OpenRouter. He joined OpenRouter's weekly top named models with their Artificial Analysis Intelligence Index scores. He found that "84% of tokens on OpenRouter aren't state of the art" in the week of August 10, defining "near the frontier" as within 10% of that week's best-scoring named model.
Key details that the shorter versions drop:
- Six models carried about 80% of named volume, at a blended price of about $0.50 per million tokens, versus $20 for Claude Fable 5.
- The share outside the frontier was about 82–85% in weekly data from November 2025 to May 2026. So 84% isn't a new shift.
- The sample covers the head of OpenRouter's rankings only. It excludes the "Others" bucket and first-party traffic to OpenAI, Anthropic and Google, which Tunguz says biases the data.
Crypto Briefing and TokenPost repeated the figure on October 13, which is likely why it resurfaced. It describes the mix on one router, and it says nothing about whether total usage is rising or falling.
What did Gartner's CIO survey find?
Gartner's October 19 press release covers its 2027 CIO and Technology Executive Agenda survey of 2,002 CIOs and technology executives, fielded May to July 2026:
| Finding | Share |
|---|---|
| Enterprise has realized significant or transformational value from AI so far | 13% |
| Reporting increased pressure to deliver AI cost improvement | 58% |
| Expect AI to increase total cost of ownership of technology solutions | 51% |
| Have guidelines for when an AI pilot should not go to production | 31% |
| Have guidelines for abandoning low-value AI already in production | 17% |
| Plan to redesign roles for a human-agent workforce | 4% |
Table 2Only 13% report significant value while 58% face more pressure to cut AI costs; none of these figures measures usage. Source: Gartner, October 19, 2026 (2,002 CIOs, surveyed May to July 2026).
Last verified: October 20, 2026 (Gartner press release)
Gartner doesn't define "significant or transformational value" in the release, and the results are self-reported. "Volume is not value," said Gartner analyst Alicia Mullery. She also noted that "AI costs are probabilistic and determined at run time," unlike most technology costs.
At the keynote in Orlando, InformationWeek reported, Gartner analysts said AI spending had grown about eighteenfold in two years. They also said 21% of CIOs have agentic AI governance today, while 71% expect to deploy agentic AI within 12 months. Gartner's analysts told CIOs to tie AI spending to measurable outcomes and to stop using AI "when you don't need to."
None of that says usage is falling. It says most CIOs haven't seen the payoff yet and are under pressure to control costs. Gartner's September 16 forecast still puts worldwide AI spending at $2.67 trillion in 2026, up 49.5%, and about $3.64 trillion in 2027.
What does the counter-evidence say?
Each of these has been re-checked against its source:
- Anthropic's revenue. On October 14, the FT, Reuters and Bloomberg reported that Anthropic told prospective investors its third-quarter revenue was $16.8 billion, up about 45% from about $11.6 billion in the second quarter. Anthropic hasn't published the figure (our analysis).
- TSMC's results. On October 15, TSMC reported Q3 revenue of $46.72 billion, up 41.2% year over year in US dollars, and guided Q4 to $52.3–54.3 billion (our report).
- Accenture's cap. 404 Media reported on October 15 that Accenture limits central Claude funding to $29 per employee per month, after some staff spent up to $1,000 (our report). That's cost control, not abandonment. Usage above the cap moves to project budgets. In August, 404 Media also reported that Microsoft told engineers "tokenmaxxing is not what we are optimizing for" as it set AI spending limits.
- Market jitters. The October 14–15 sell-off, in which CoreWeave's bonds fell until they yielded about 15%, reflected worries about debt and financing, not falling token demand (our explainer).
So, is AI usage slowing?
The evidence supports a narrower conclusion:
- Supported: companies are capping and scrutinizing AI spending (Accenture, Microsoft, Gartner's 58%), and most CIOs don't yet report big returns (Gartner's 13%).
- Supported: on OpenRouter, most tokens go to low-cost models below the frontier, as they have for about a year.
- Not supported: that overall AI usage is falling. OpenRouter's public page shows no all-model total, its biggest weekly drop coincides with a version switch, and frontier models grew on it the same week. Revenue at Anthropic and TSMC kept climbing.
The end of "token-maxxing" looks like a shift from using as many tokens as possible to paying for the right ones. That matches the setup our model-routing explainer describes: send most calls to a cheap model and escalate only the hard ones to a frontier model.
What's confirmed, reported and uncertain
- Confirmed (primary sources): OpenRouter's per-model figures and request shares on October 20; Gartner's survey figures and spending forecast; TSMC's Q3 results.
- Reported: Anthropic's $16.8 billion quarter (FT, Reuters, Bloomberg); Accenture's $29 cap (404 Media); Gartner keynote figures (InformationWeek).
- Third-party analysis: Tunguz's 84% estimate, with the methodology limits above.
- Uncertain: whether OpenRouter's total weekly tokens fell. The X post claiming a "first weekly decline in about a month" couldn't be opened, and OpenRouter doesn't show the total publicly.
About this storyBased on the sources linked below. Editorial standards
