By asahifile 
Between 26 June and 29 September, OpenAI and Anthropic rewrote their API price sheets eight times. GPT-5.6 Luna fell 80 percent in one announcement on 30 July. Claude Sonnet 5’s temporary launch price became permanent on 10 August. Claude Fable 5.1 cut its cache-read price by 75 percent on 1 September. On 22 September the two labs shipped flagship releases within hours of each other, and a week later OpenAI halved the GPT-6.1 Sol cache price while leaving its input and output rates untouched.
Every summary of that run now repeats the same line: LLM prices have fallen by up to 80 percent in a few months. The claim is accurate about one model tier and one way of counting. To see what a buyer actually saves, I re-costed two fixed workloads against 16 published price points, using dollar prices per million tokens. The flagship discount lands at 62 to 71 percent, not 80. And the line item that moved the invoice most was not the one in the headline: it was the cache-read price.
What the price sheets said
- 26 June — GPT-5.6 Sol at $5 / $30 / $0.50 per million input, output and cache-read tokens; GPT-5.6 Luna at $1 / $6 / $0.10.
- 30 July — Luna cut 80 percent to $0.20 / $1.20 / $0.02.
- 1 September — Fable 5.1 keeps its $10 / $50 headline, cutting cache reads from $1.00 to $0.25.
- 22 September — GPT-6 Sol at half of GPT-5.6 Sol’s discounted rates ($2 / $10 / $0.20); Claude Opus 5.5 at $4 / $20 / $0.20 from $5 / $25 / $0.50.
- 29 September — GPT-6.1 Sol keeps $2 / $10, halves cache reads to $0.10.
- Open weights — DeepSeek V4.1 Flash 1017 at $0.15 / $0.60 / $0.003; Kimi K3.1 Pro at $2 / $10 / $0.20.
Workload A: one 200-page document, no caching
180,000 input tokens in, 12,000 out. On GPT-5.6 Sol in June that job costs $1.26; on GPT-6.1 Sol it costs $0.48, a 61.9 percent drop. Opus 5’s $5 / $25 sheet priced it at $1.20 and Opus 5.5 at $0.96, exactly 20 percent off. The 80 percent headline only reproduces on Luna, where $0.2520 becomes $0.0504 because all three rates were cut by the same amount.
The other surprise is convergence. GPT-6.1 Sol, Sonnet 5 and Kimi K3.1 Pro all land at $0.48 to the cent, while Fable 5.1 and GPT-6 Astra charge $2.40 — five times more for identical token counts. DeepSeek’s open-weight 1017 checkpoint prices the job at $0.0342, fourteen times below the closed frontier tier.
Workload B: one 100-turn agent session
I modelled the shape agent products actually bill: 100 turns, each carrying 45,000 cached context tokens, 1,500 fresh input tokens and 800 output tokens. Cache reads are 95.1 percent of the tokens. They are not 95 percent of the bill. At Opus 5 pricing the cache line is 45 percent of the total; at Opus 5.5 and GPT-6.1 Sol, 29 percent; at Fable 5.1 and DeepSeek, 17 and 16.1 percent.
The session costs $5.00 on Opus 5 and $3.10 on Opus 5.5, down 38 percent; $5.40 on the June GPT-5.6 Sol and $1.55 on GPT-6.1 Sol, down 71.3 percent; $10.00 on Fable 5 against $6.63 on Fable 5.1, down 33.8 percent; and $0.084 on DeepSeek. The cheapest documented change was also the most effective: the 29 September move from GPT-6 Sol to GPT-6.1 Sol altered nothing but the cache price and still cut the session bill by 22.5 percent.
That is the arithmetic behind the idea that cache costs are a footnote. In agent traffic the cache-read rate is effectively the price of the conversation, and it ranges from 2 percent of the input rate at DeepSeek to 10 percent on the OpenAI and Anthropic flagships — a five-fold spread no headline advertises.
Why aggregate AI spend still rises
Vercel’s gateway data, cited in Max Woolf’s chronology, puts cost per token at 43.3 percent of its April level while aggregate spend climbs. For spend to stand still at that price, volume would have to be 2.31 times higher. It is not: cheaper tokens buy more attempts, longer contexts and more agents, not a smaller invoice. A 62 percent unit-price cut and a rising bill are perfectly compatible facts.
The yen invoice
At ¥156.84 to the dollar (ECB reference rate, 23 October), the session that cost ¥784 on Opus 5 in July costs ¥243 on GPT-6.1 Sol, and the document job falls from ¥188 to ¥75. A Tokyo team running 2,000 sessions a month moves from roughly ¥1.57 million to ¥486,000 without changing a line of code. Japanese buyers carry one extra exposure the dollar sheets hide: resale and gateway pricing is quoted in yen with a fixed spread, so a currency that weakens past 160 erodes part of the saving before it reaches the invoice.
When the cheap model is the wrong call
The answer flips in four situations. First, if the work is output-heavy and uncached, the frontier tier has stopped competing on price — GPT-6.1 Sol, Sonnet 5 and Kimi K3.1 Pro all sit at $0.48 — so pay for quality, because there is nothing left to save. Second, if you cannot hit the cache, the cuts that generated most of this year’s savings give you nothing, and your real discount is 61.9 percent, not 80. Third, if you resell tokens or run a gateway, falling unit prices shrink the revenue line while compliance, latency and support costs stay fixed. Fourth, and most expensive: do not move a task to the cheapest tier when that tier fails it, because an 80 percent saving on work you must redo is a total loss.
What to watch
- Anthropic’s IPO prospectus and the inference margin it discloses — the reported 80 percent adjusted operating margin is what the next price cut has to survive.
- Whether Fable 5.2 or GPT-6.1 Astra hold the $10 / $50 tier.
- Kimi K3.1 Pro’s weights, due 4 November: more than 20 providers serve the previous K3 build, and that is what turns a list price into a market price.
- Cache-read ratios converging on 2 percent of the input price; DeepSeek is already there.
Sources: Max Woolf, “OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?” (25 October 2026), which links the primary announcements from OpenAI, Anthropic, OpenAI and OpenRouter; USD/JPY reference rate from the European Central Bank via Frankfurter. All workload costs were computed for this article from published per-million-token prices.