What happened
Silicon Data runs a daily index, listed on Bloomberg under the ticker SDLLMTK, that tracks the real, usage-weighted price of a million AI tokens across a broad set of providers and models. It is not a list of sticker prices. It blends posted pricing with actual consumption data pulled from multi-provider routing gateways, so it reflects what businesses are really paying on average, not just what a single vendor advertises. On August 31, 2026, that index fell to 97 cents per million tokens, its lowest reading since it launched in late 2025 and less than half the peak it hit earlier this summer. It had already dropped 8.6% in the prior seven days alone. CNBC, which first reported the reading, attributes the decline to several forces stacking on top of each other: cheaper open-source models out of China, such as Moonshot's Kimi K3, undercutting the major labs on price; OpenAI's own price cuts on its GPT-5.6 model line in late July (confirmed separately by Reuters, which reported OpenAI cut its flagship GPT-5.6 Sol model's developer pricing by more than 20% in August); other labs rolling out "dynamic pricing" that moves with demand; and falling underlying compute costs to produce each token. Charles-Henry Monchau, chief investment officer at Syz Group, put the mechanism plainly in comments cited by CNBC: "Token deflation compresses the revenue line while compute commitments stay fixed." Silicon Data's own head of research, Steve Hou, said the drop may signal that supply across both frontier and lower-cost models is already enough to handle most tasks. The timing matters: both OpenAI and Anthropic confidentially filed for IPOs this summer, and this index move lands squarely in the middle of that process. Technology stocks fell the same day, with the Nasdaq down nearly 1% and the S&P 500 down 0.4%, as investors weighed what thinner AI margins mean for the huge infrastructure bets already made around AI.
Why it matters for business owners
Most coverage of AI pricing focuses on a single vendor's announcement: a company cuts a price, or raises one, and that becomes the story. This index is different because it is not one company's decision. It is a market-wide reading of what businesses are actually paying, blended across many providers and models, and it just hit its lowest point ever. That matters for two practical reasons. First, if your business runs meaningful AI workloads through APIs, this is direct evidence that the cost side of that equation is trending down, not up, which changes the calculus on questions like whether to build in-house infrastructure to escape rising cloud AI costs. Second, and less obviously, a market where token prices are falling this fast is a market where the vendors selling you access are under real revenue pressure, even while their underlying costs to run the service stay largely fixed. Companies under that kind of margin pressure do not always respond by cutting prices further forever. Some will change what is included in a plan, restructure tiers, or emphasize new paid features to protect revenue as the raw price of a token keeps falling. Watching this trend is not just about tracking your bill going down. It is about anticipating how the vendor relationship itself might shift.
What owners should not misunderstand
This index does not mean AI is now free, or close to it. Ninety-seven cents per million tokens is a market-wide blended average across many models and providers, weighted toward how the market is actually using them, not a price quote for any specific tool your business uses. A workflow that runs a frontier model on long documents at high volume can still generate a real bill, even as this index falls. It also does not mean prices will keep falling forever, or that they cannot reverse. Index moves like this one reflect current competitive intensity, particularly aggressive pricing from open-source Chinese models and a round of cuts from major labs. If that competitive pressure eases, if a smaller set of providers ends up controlling most of the market after this price war shakes out, or if demand for AI capacity outpaces the current oversupply Steve Hou describes, prices could stabilize or climb again. Trend lines are not guarantees. And this is not a signal to switch every workflow to the cheapest model you can find. Cheaper tokens matter less than whether a model is actually good enough for the specific task, and a cheap model that produces more errors, needs more retries, or requires more human correction time can cost more in the end than a pricier one that gets the job right the first time.
The operational lesson
The useful takeaway is not "AI got cheaper, celebrate." It is that AI pricing is currently moving fast enough, in a single direction, that any long, fixed-rate commitment you sign today carries real opportunity cost if the market keeps moving the way it has for the past several months. This is the same logic that applies to any input whose price is falling: you generally do not want to lock in a long-term contract at today's rate when the trend says next quarter's rate could be meaningfully lower. That argues for usage-based or short-term AI vendor commitments over long annual contracts right now, and for keeping your workflows built in a way that lets you route between models or switch providers without a costly rebuild, so you can actually capture a better price when one shows up instead of being stuck. It also argues for revisiting your AI budget assumptions periodically rather than setting them once. A budget built on token prices from six months ago is likely already stale, in either direction, and the businesses that benefit most from a falling-price market are the ones that check in on it rather than assume last quarter's numbers still hold.
What a serious business should do next
Do not sign a long, fixed-rate AI vendor contract this quarter on the assumption that today's price is a floor. If a vendor is pushing an annual commitment, ask directly what happens if list prices drop during the term, and whether the contract lets you benefit from that or locks you out of it. Do pull your actual AI spend and check what you are really paying per unit of work, not just the sticker price on a plan, the same way this index measures real usage-weighted spend rather than posted rates. If you have not looked at this in the last few months, the number has likely moved. Do keep enough flexibility in how your workflows are built that switching models or providers is a real option, not just a theoretical one. That flexibility is what actually lets a business capture a falling market instead of watching it happen to competitors. Do not treat this as a reason to rush into a new AI project that was not worth doing at the old price. A workflow that did not justify its cost six months ago needs a real reason beyond a lower token price, not just a smaller bill for the same weak case.
The Atlacis view
A record-low pricing index is genuinely useful information, but it is also easy to misread. It is not permission to stop paying attention to AI cost, and it is not proof that any specific tool your business uses just got cheap enough to justify skipping the homework on whether it fits your workflow. Atlacis helps business owners map what they actually spend on AI today, at the unit level, and decide how to structure vendor commitments so a falling market works in their favor instead of leaving them locked into last quarter's prices. That is a conversation worth having before you sign the next contract, not after.
The short version
- Silicon Data's LLM Token Expenditure Index, a market-wide benchmark of real AI token spend, fell to 97 cents per million tokens on August 31, 2026, its lowest reading ever and less than half its peak from earlier this summer.
- The decline is driven by cheap open-source Chinese models like Kimi K3, OpenAI's own GPT-5.6 price cuts, dynamic pricing from other labs, and falling compute costs, not a single vendor's promotion.
- The drop lands as OpenAI and Anthropic both weigh IPOs after confidentially filing with regulators this summer, and it raises real questions about pricing power and returns on AI infrastructure spending.
- The index is a blended market average, not a quote for your specific workload. A heavy, frontier-model use case can still cost real money even as the broad market average falls.
- A falling-price market is a bad moment to lock into a long, fixed-rate AI vendor contract. Usage-based terms and the ability to switch providers let a business capture further price drops instead of missing them.
- This is not a signal to chase the cheapest model for every task, and not a reason to greenlight a weak AI project that only looks better because tokens got cheaper.
Where ATLACIS can help
Sources
- Silicon Data: LLM Token Expenditure Index (SDLLMTK), checked September 2, 2026
- CNBC: Artificial intelligence token prices are hitting new record lows (Alex Harring, September 1, 2026)
- Seoul Economic Daily: AI Token Prices Hit Record Low, Squeezing OpenAI and Anthropic Ahead of IPOs (Kim Jeong-Uk, September 2, 2026)
- Reuters: OpenAI cuts developer pricing for frontier GPT-5.6 Sol model by more than 20% (August 21, 2026)