Tokens Got 80% Cheaper. Why Your AI Bill Didn't Drop.
· AI Hot Editorial
OpenAI and Anthropic have been cutting model prices hard — Luna's unit price is down 80%. Yet total spend went up. Here's why: per-token price, cache-read price, and per-task cost are three completely different bills.
A lower per-token price does not mean a lower bill. OpenAI and Anthropic cut prices by up to 80% in a few months, but per-token price, cache-read price, and per-task cost are three different bills. Cache reads dominate in the agent era while task volume climbs — Vercel's data shows September spend at 13× December's as per-token price fell 47%.
Confusing numbers: OpenAI cut GPT-5.6 Luna by 80% and the GPT-6 series in half; Anthropic cut Fable 5.1's cache-read price 75% and Opus 5.5 to 40% below the prior generation. Yet spend rose.
1. What happened: how much prices fell in four months
Prices per 1M tokens.
Jul 30 — OpenAI: GPT-5.6 Luna -80%, Terra -20% — Luna $0.20 in / $1.20 out.
Aug 10 — Anthropic: Sonnet 5 discount permanent — $2 in / $10 out.
Sep 1 — Anthropic: Fable 5.1, input/output same — cache reads -75% to $0.25.
Sep 22 — OpenAI: GPT-6 Sol + Luna, -50% vs promo — Anthropic: Opus 5.5 $4 in / $20 out.
Sep 29 — OpenAI: GPT-6.1 Sol — $2 in / $10 out / $0.10 cache.
Late Oct — Moonshot: Kimi K3.1 Pro, same as GPT-6.1 — $2 in / $10 out; Artificial Analysis 55, Opus 5.5 58.
Opus 5.5 is billed as serving more efficiently than Opus 5.
2. The three bills, and why you shouldn't mix them
First, the per-token price everyone watches. The 80% cut refers to this. It only answers "what does one token cost."
Second, the cache-read price — the real driver in the agent era. Every turn the model re-reads prior context at cache rates, typically 5% to 10% of normal input. A 75% cache cut moves a bill far more than a 50% input cut.
Third, per-task cost: unit price times tokens burned. If unit price falls 80% but each task burns twice as many tokens, the bill does not move. September spend was 13× December's while per-token price fell 47% — agents did more work.
3. Vendors cut prices and quietly collect more
On the same Sep 29, OpenAI adjusted the allowance on the $200/month Pro 200 tier, so new subscriptions get less. The help center's official line: this "reflects our models getting more efficient." Technically not wrong. But when "models got cheaper" and "your plan contains less" land together, the bill curve stops being simple.
4. Why prices kept falling
Competition first: DeepSeek Flash V4.1 has 29 inference providers on OpenRouter at lower prices, cache reads only 2% of input cost — better value than OpenAI's and Anthropic's 5% to 10%. Second, vendors say inference efficiency is improving; Anthropic's line that "Opus 5.5 requires less compute to serve than Opus 5, so the pricing reflects that" is the round's most honest rationale. Third, both companies are reportedly preparing for IPOs, so undercutting pricing on their main revenue line looks odd. For users, open-source models set a floor.
5. What users and builders should do
Regular users: stop judging cost by "price per million tokens." Log what you ran this month — if unit price fell and you ran more work, you saved nothing. For subscriptions, watch two numbers, the monthly fee and the included allowance, since price and allowance cuts come separately.
Builders: your biggest cost line is probably not input, it is cache reads and retries — pull those out separately. Switching models is easier than switching API endpoints, so evaluate while models are cheap: same tasks on old and new models, recording latency, cost, error.
FAQ
Token prices fell — why didn't my bill go down?
Three bills. Unit price falls 80% but each task burns twice as many tokens, so the bill holds.
What are the three bills?
Per-token price, cache-read price, per-task cost = unit price × tokens burned. Cache reads run at 5%-10% of input.
Did vendors quietly collect more while cutting prices?
Yes. On Sep 29 OpenAI cut the $200/month Pro 200 tier allowance.
How should a builder figure out their real cost?
Split out cache reads and retries, and evaluate while models are cheap: same tasks on old and new models, recording latency, cost, error.
Prices fell 80%; the bill may not have dropped by a cent. The per-token price is a number someone else told you. Per-task cost is the number in your wallet.