Why AI API prices fell up to 80% in four months — cache reads are 96–99% of agent spending
According to a timeline by data scientist Max Woolf published October 25, 2026, OpenAI and Anthropic cut API prices by up to 80% between late July and late September. The main pressure comes from cheap open-weight models such as DeepSeek and Kimi served by dozens of providers. In coding agents, 96–99% of token spending is cache reads, so cuts to the cache-read price lower real bills more than cuts to headline rates
The three lines
- Cuts — GPT-5.6 Luna −80% on Jul 30; Opus 5.5 40–60% cheaper in Sept; GPT-6.1 Sol cache halved Sept 29
- Why — open-weight competition, easy switching, cheaper serving, tit-for-tat
- Key — cache reads are 96–99% of agent spend; a 75% cache cut lowers total cost ~70–75%
Key questions
- ai api price cuts 2026
- **Eight cuts between late July and late September (per million tokens: input / output / cache read).** | Date | Model | Before | After | |---|---|---|---| | Jul 30 | GPT-5.6 Luna | $1 / $6 / $0.10 | $0.20 / $1.20 / $0.02 | | Aug 10 | Claude Sonnet 5 | Launch price $2 / $10 / $0.20 | Made permanent | | Aug 21 | GPT-5.6 Sol | $5 / $30 / $0.50 | $4 / $20 / $0.40 (3 months) | | Sep 1 | Fable 5.1 | $10 / $50 / $1 | $10 / $50 / $0.25 | | Sep 22 | GPT-6 Sol | $4 / $20 / $0.40 | $2 / $10 / $0.20 | | Sep 22 | GPT-6 Luna | $0.20 / $1.20 / $0.02 | $0.10 / $0.50 / $0.01 | | Sep 22 | Claude Opus 5.5 | Opus 5: $5 / $25 / $0.50 | $4 / $20 / $0.20 | | Sep 29 | GPT-6.1 Sol | $2 / $10 / $0.20 | $2 / $10 / $0.10 |
- what is cache read pricing
- **The price for reusing a prompt prefix the model has already processed.** | Item | Detail | |---|---| | Applies to | Identical leading context (system prompt, code, history) | | Price | Usually 5–10% of the input price | | Share of agent spend | 96–99% | | Cache hit rate | About 90–100% |
- why are llm prices dropping
- **Cheap open-weight models cap what closed labs can charge, Woolf argues.** | Cause | Detail | |---|---| | Open-weight competition | DeepSeek V4.1 Flash at $0.15 / $0.60, offered by 29 providers | | Switching costs | One config line via OpenRouter or agent harnesses | | Serving efficiency | Anthropic: Opus 5.5 needs less compute to serve than Opus 5 | | Tit-for-tat | One lab cuts, the other follows |
AI model prices fell by up to 80% in four months. On October 25, 2026, data scientist Max Woolf laid out OpenAI's and Anthropic's price cuts by date and asked why. From the 80% cut to GPT-5.6 Luna on July 30 to the halving of GPT-6.1 Sol's cache-read price on September 29, there were eight cuts in two months. His conclusions: the force pushing prices down is cheap open-weight models, and what actually shrank bills was not the headline rate but the cache-read price.
This piece covers what fell, why, and how much your bill actually changes.
1. What fell, and by how much
Prices are per million tokens: input / output / cache read.
| Date | Lab | Model | Before | After | Change |
|---|---|---|---|---|---|
| Jul 30 | OpenAI | GPT-5.6 Luna | $1 / $6 / $0.10 | $0.20 / $1.20 / $0.02 | ~80% cut |
| Aug 10 | Anthropic | Claude Sonnet 5 | $2 / $10 / $0.20 (temporary launch price) | Same, made permanent | Planned increase cancelled |
| Aug 21 | OpenAI | GPT-5.6 Sol | $5 / $30 / $0.50 | $4 / $20 / $0.40 | ~20%, for three months |
| Sep 1 | Anthropic | Fable 5.1 | $10 / $50 / $1 (Fable 5) | $10 / $50 / $0.25 | Cache read −75% |
| Sep 22 | OpenAI | GPT-6 Sol | $4 / $20 / $0.40 | $2 / $10 / $0.20 | Halved |
| Sep 22 | OpenAI | GPT-6 Luna | $0.20 / $1.20 / $0.02 | $0.10 / $0.50 / $0.01 | Halved |
| Sep 22 | Anthropic | Claude Opus 5.5 | $5 / $25 / $0.50 (Opus 5) | $4 / $20 / $0.20 | 40–60% lower total cost |
| Sep 29 | OpenAI | GPT-6.1 Sol | $2 / $10 / $0.20 | $2 / $10 / $0.10 | Cache halved; family ~75% cheaper |
2. Cache reads — the real line item
A coding agent calls the model dozens or hundreds of times per task, resending the system prompt, open files and conversation history each time. When the leading context is identical to the previous request, the provider reuses the computation and charges only a cache read, typically 5–10% of the input price.
| Item | Woolf's measurement or estimate |
|---|---|
| Cache reads as share of coding-agent token spend | 96–99% |
| Cache hit rate | About 90–100% |
| OpenAI/Anthropic cache-read price | About 5–10% of input price |
| DeepSeek V4 Flash cache-read price | 2% of input price |
| Effect of a 75% cache-read cut | ~70–75% lower total agent cost |
So Fable 5.1 on September 1 — same $10 / $50 headline, cache read from $1 to $0.25 — amounted to roughly a three-quarters cut in total cost for heavy agent users. Conversely, a headline "20% off input" can mean less than it sounds for agentic workloads.
A quick way to estimate your own exposure: check what share of your tokens are cache reads. Above 90%, compare models on cache-read price before input and output prices.
3. Why prices fell — Woolf's five reasons
| Cause | Explanation | Evidence |
|---|---|---|
| Open-weight price pressure | Downloadable models are served very cheaply | DeepSeek V4.1 Flash $0.15 / $0.60 / $0.003, offered by 29 providers on OpenRouter |
| Easy switching | Agent harnesses and OpenRouter swap models with one line | Little moat |
| Tit-for-tat | One lab cuts, the other follows | OpenAI and Anthropic both cut on Sep 22 |
| Serving efficiency | Same capability with less compute | Anthropic: Opus 5.5 "requires less compute to serve than Opus 5" |
| Obsolescence | Cheaper, better models push out pricier ones | Also a strategic risk for the labs |
Woolf thinks the July Luna cut targeted small open-weight models, especially DeepSeek. There is capability pressure too: Moonshot AI's Kimi K3.1 Pro, at $2 / $10 / $0.20, performs close to Opus 5.5 at half its price.
Woolf also finds it odd that both labs cut their main revenue source while preparing IPOs, when investors scrutinize losses. Anthropic's S-1, made public October 26, shows compute and infrastructure costs at 48% of revenue in Q2 2026 — evidence that serving costs really did fall.
4. Will prices go back up?
| Scenario | Woolf's view |
|---|---|
| Big price hikes after IPOs | Unlikely — open-weight models and providers cap prices |
| Next flagship launches expensive | Open question — watch Fable 5.2 and GPT-6.1 Astra |
| Failed IPOs kill the industry | No — models and agents survive as an industry |
5. FAQ
My headline price didn't change but my bill dropped — why? Probably a cache-read cut. In agentic use, cache reads dominate spend.
Is a cheaper model always cheaper? No. Total tokens to finish a task differ by model, and the same sentence tokenizes differently across models.
Aren't open-weight models free? The weights are; the servers are not. That is why inference providers charge to run them.
6. What remains and what is unconfirmed
- Official explanations: neither lab has explained the cuts the way this analysis does.
- Margins: high inference margins are reported but not confirmed in public figures.
- Next prices: pricing for the next flagships is unknown.
Sources
- Max Woolf — OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why? (2026-10-25)
- ProPakistani — OpenAI and Anthropic Slashed AI Model Prices by Up to 80% (2026-10-26)
- ExplainX — LLM Price Cuts 2026: 80% Off in 4 Months, Explained
- Byteiota — Why OpenAI and Anthropic Cut API Prices 80%