On 27 August Alibaba posted the price list for Qwen3.8-Flash, and the number that travelled was the last one.
$0.016 per million tokens on a cache hit is a real price and Alibaba's own model page confirms it. It is 10.7% of the $0.15 input price and 3.4% of the $0.47 output price — so a token the model has already seen costs 9.4x less than reading a new one and 29x less than writing one.
It is also, almost exactly, what every other major provider charges. Across OpenAI's current model line, cached input is a flat 10% of the uncached rate. Google's is 10%. Anthropic's is 10%. The 90%-off cache discount is not a Qwen event; it is the settled industry structure, and Alibaba has just joined it on this model.
What has not converged is the second price — what it costs to keep a cache alive long enough to hit it. There the five providers disagree completely, and that is the number worth watching.
The points
- $0.15 input, $0.47 output, $0.016 cached for Qwen3.8-Flash, from Alibaba's own model page, checked 7 September. Not promotional: the page states it shows original pricing excluding limited-time offers.
- Those are the Singapore prices, and Singapore is the expensive region. In Alibaba's five other regions — including US (Virginia) and Germany (Frankfurt) — the same model is $0.113 input, $0.382 output, $0.014 cached, about 25% less. The post quoted the dearer list.
- Alibaba's own published cache schedule says 20% of input for automatic caching and 10% only if you manage the cache yourself. Its other Qwen models charge exactly that. Qwen3.8-Flash is a named exception and charges 10.7% either way.
- So managing the cache on this model now buys nothing on price. Explicit cache creation costs $0.20 per million — 133% of the input price — and reads back at the same $0.016 the automatic tier gives you free.
- Of the five price lists here, only Google charges by the hour to hold a cache. Its explicit context caching is metered at $0.50 to $4.50 per million tokens per hour. Alibaba, OpenAI, Anthropic and DeepSeek all charge zero.
- DeepSeek remains the outlier on the discount itself at 3.2% of the miss price — roughly 97% off, three times deeper than the industry's 10%.
The table above sets the six rows side by side; it is the whole comparison in one place.
Why a cache hit can be nearly free
The mechanism is worth stating because it explains why five companies with different hardware, different margins and different strategies all landed within a rounding error of each other.
Serving a prompt has two phases. Prefill reads the whole input and builds a key-value tensor for every token — that is a large matrix multiplication, and it is what the input price pays for. Decode then generates output one token at a time. A cache hit skips prefill entirely: the tensors for that prefix already exist in GPU memory, so the provider streams them out of memory rather than recomputing them. What is left to charge for is memory bandwidth and the storage the prefix occupies, not arithmetic.
That is why the discount is roughly an order of magnitude everywhere. It is not a competitive concession — it is the ratio between reading memory and doing maths, and it is the same physics for everybody. It is also the token-price side of the hardware story in our note on Nvidia's megawatt: the throughput per megawatt that Nvidia sells is precisely what makes a skipped prefill cheap to give away.
What Alibaba actually changed
Alibaba publishes a two-tier cache schedule. Implicit caching is automatic, cannot be turned off, and bills hits at about 20% of the input price. Explicit caching makes you mark blocks in the request, costs about 125% of input to create an entry, and rewards you with hits at about 10%. Its other models follow that schedule to the decimal: on qwen-flash and qwen-plus, implicit hits are exactly 20% of input and explicit reads exactly 10%.
Qwen3.8-Flash is one of a handful of models Alibaba's documentation names as an exception, and the exception runs one way. Automatic hits and managed hits are billed at the same rate — $0.016 in Singapore, $0.014 elsewhere — while explicit cache creation went the other way, to 133% of input in Singapore and 157% in the other regions.
The result is that explicit cache management on this model is now all cost and no discount. It still buys something real — a guaranteed hit inside a five-minute window that resets on every read, against an automatic cache whose hit probability Alibaba declines to guarantee. But it no longer buys a lower price, and that is a change in what Alibaba is selling: it has stopped charging a premium for developers who do not want to manage a cache.
The curve nobody quotes: what it costs to keep the cache warm
A cache-hit price is only worth what your hit rate is, and hit rate is a function of how long entries survive. Here the five providers have made genuinely different bets:
- Alibaba gives explicit entries five minutes, renewed on every hit, and describes implicit retention as indeterminate. No storage charge.
- OpenAI caches automatically with a 30-minute sliding window and a 1.25x surcharge on the write. No storage charge.
- Anthropic sells a longer life as a bigger one-off write — 1.25x input for five minutes, 2x for an hour — and then charges nothing to hold it.
- DeepSeek caches automatically, keeps entries "a few hours to a few days" on a best-effort basis, and states on its own documentation that storage is free.
- Google is alone in metering residency. Explicit context caching is billed at $0.50 per million tokens per hour on Gemini 3.8 Flash and $4.50 on Gemini 3.1 Pro. Google documents implicit caching as storage-free, so the rent applies to the managed path only.
The scale of that difference is easy to miss. On Gemini 3.1 Pro, one hour of holding a cache costs 22.5 times one read of it. For a workload that reads a large context a handful of times an hour, storage — not the discount — is the dominant line. Anthropic charges for the same thing, but prepaid at write time rather than metered, which is a materially different shape for bursty traffic.
So the useful reading of the Qwen3.8-Flash announcement is not "cache hits got cheap." It is that the cache discount has stopped being a competitive variable and the cache clock has become one.
What it does to a workload
Take a 200,000-token context — comfortably inside the 991,808-token maximum input Alibaba lists for this model — carried across a 30-turn agent session emitting 1,500 output tokens a turn, at the Singapore prices. The arithmetic below is ours.
Reading that context cold costs $0.0300. Reading it warm costs $0.0032. Over the session that is $0.921 uncached against $0.144 cached — 6.4x — and on Alibaba's published 20% implicit rate it would have been $0.225, so the exception is worth a further 1.56x on top.
The ratio that reframes the design problem is smaller and more useful: at these prices, re-reading the entire 200,000-token context costs the same as generating 6,809 output tokens. Context length has become close to free relative to generation. That is the third demand regime our token-economics piece named — persistent agents carrying a large, mostly-unchanged context across many calls — and it is now priced accordingly by everyone, not by Alibaba alone.
It is also the per-token counterpart to the per-task collapse we measured last week, where the verified cost of a solved reasoning task fell 15,200x in twenty months. Cheaper cached reads are one of the mechanisms underneath that number.
What would change the conclusion
- A published storage fee from anyone other than Google. The claim that residency is the remaining variable rests on four vendors charging zero for it. One more meter changes the shape.
- Realised cache hit rates. Every figure here is a list price. What developers actually pay depends on what fraction of their tokens hit, and no provider publishes that distribution.
- Whether Alibaba extends the exception. Qwen3.8-Flash gets automatic caching at the managed rate; the rest of the Qwen line still pays 20%. If the 10% rate becomes the standard, Alibaba has repriced its whole catalogue rather than launched one model.
- Any disclosure tying Qwen pricing to cloud revenue. Alibaba reports AI Cloud as one line and has never given a token volume, a price realisation or a serving margin. Until it does, the link from this price list to the segment is an assumption, not an estimate.
Qwen3.8-Flash's input, output, implicit-cache, explicit-cache-creation and explicit-cache-read prices, in all six of Alibaba's regions, and the 991,808-token maximum input, are published on Alibaba Cloud's own model page and were read on 7 September 2026; that page states it shows original pricing excluding limited-time promotions. Alibaba's 20% implicit / 10% explicit / 125% creation schedule, the five-minute explicit TTL and the exception naming this model come from its context-cache documentation, last updated 2 September 2026. The OpenAI, Google, Anthropic and DeepSeek figures are each vendor's own current price list and caching documentation, read the same day; the OpenAI and Anthropic cache rates are documented multipliers rather than a single model's price. The 27 August post is Alibaba's own and quotes the Singapore prices. Every ratio, the worked 30-turn example and the equivalent-output-token figures are ours, arithmetic on those published prices. No figure here is a company estimate of demand, and none is a disclosure about Alibaba's cloud revenue.
