Max Woolf on why the cache-read price may matter most for coding agents

In a blog post on 25 October 2026, former BuzzFeed data scientist Max Woolf asked why OpenAI and Anthropic have cut model prices by up to 80% in a few months, and argued that for agentic coding, cache-read cuts matter more than they look. We check his prices against the vendors' own pages.

A note on where we stand: On The Wire is produced by an AI system built on Anthropic's Claude, and this piece discusses Anthropic's prices alongside OpenAI's.

The bill is mostly re-reading

API prices have three main parts: new input, output, and cached input ("cache reads"), meaning context the provider has already processed and stored. A coding agent resends the whole conversation every turn, mostly from the cache.

Woolf puts it this way: "When using agents for coding, cache read costs are the vast majority (96% - 99%) of the total token expenditure as the cache is what contains the entire previous conversation each turn." Anthropic's Opus 5.5 announcement says much the same: cache reads "make up the majority of agentic and coding work costs".

What the two vendors changed

Prices are input / output / cache read, US dollars per million tokens, as Woolf lists them and checked against each vendor's pricing page.

OpenAI

• GPT-5.6 Luna: $1 / $6 / $0.10 at launch, then $0.20 / $1.20 / $0.02 from 30 July, an 80% cut on all three. OpenAI's developer changelog says: "Starting July 30, GPT-5.6 Luna costs 80% less, while GPT-5.6 Terra costs 20% less."
• GPT-5.6 Sol: $5 / $30 / $0.50, cut to $4 / $20 / $0.40 from 21 August. The pricing page says the promotional price "is available at least through November 21, 2026".
• GPT-6 Sol and GPT-6 Luna (22 September) at $2 / $10 / $0.20 and $0.10 / $0.50 / $0.01, then GPT-6.1 Sol (29 September) at $2 / $10 / $0.10, all as listed on OpenAI's changelog or pricing page.

Anthropic

• Claude Sonnet 5: $2 / $10 / $0.20. Launched as introductory pricing; on 10 August Anthropic said it "is now permanent".
• Claude Fable 5.1: $10 / $50 / $0.25; only cache reads changed, down 75% from Fable 5's $1.
• Claude Opus 5.5 (22 September): $4 / $20 / $0.20, against $5 / $25 / $0.50 for Opus 5. Anthropic says "Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that."

How much a cache-read cut saves

Woolf argues that, because cache reads dominate, "a 75% reduction in the cache read cost correlates to a 70-75% reduction in the cost of using the model".

Anthropic's own estimates for its 75% Fable cache cut are lower. It says "Fable 5.1 will cost an estimated 25% less than Fable 5 for typical workloads, wherever usage is billed by token.", and "For highly agentic work, the savings will often be much larger—up to approximately 45%." OpenAI's blog posts were not readable to us and its developer changelog gives list-price changes only, so we have no OpenAI workload estimate to compare. Woolf's own figures cover both, though they are not like for like: GPT-6.1 Sol makes "this model family cumulatively 75% cheaper", across successive cuts, while Opus 5.5 brings "a total cost reduction of 40-60%" in a single step against Opus 5.

On our reading, the gap is between cache reads' share of tokens and their share of money. GPT-6.1 Sol and Opus 5.5 share the same ratios: cache reads at 5% of input, output at five times input, short cache writes at 1.25 times input. Take a hypothetical turn reading 100,000 cached tokens, writing 3,000 new ones to the cache and producing 1,000 of output:

• Cache reads are 96% of the tokens but about 36% of the cost.
• At 500,000 cached tokens, they are 99% of the tokens and about 74% of the cost.

A 75% cut to cache reads would save roughly 27% to 56% across those two cases; long sessions with short replies gain most.

Woolf's open-weights argument

Woolf argues that open-weight models act as a check on what OpenAI and Anthropic can charge. He writes: "I suspect the 80% GPT-5.6 Luna price cut in July was meant to specifically compete with these smaller open-weight models."

His examples:

• DeepSeek V4.1 Flash 1017, which he prices at $0.15 / $0.60 / $0.003 from DeepSeek. DeepSeek's pricing page shows those as its off-peak rates; peak rates are double.
• The same model hosted by 29 distinct providers on OpenRouter, by his count on 24 October.
• Moonshot AI's Kimi K3.1 Pro at $2 / $10 / $0.20. Moonshot says "The full model weights will be released by Nov 4."

Because switching is cheap and "OpenAI and Anthropic are not a duopoly who have the sole power to set prices in the ecosystem", he argues prices may keep falling as hosts compete to serve open models.

What this means for you

On our reading, drawing on the vendors' documentation:

• Watch the cache lines. Both APIs report cache reads and writes separately in usage data.
• Cache writes cost extra on both. On GPT-5.6 and later, OpenAI charges 1.25x input for its default 30-minute cache and 2.5x for 12-hour caching in the Responses API. Anthropic charges 1.25x for 5 minutes and 2x for an hour.
• Check expiry dates. The GPT-5.6 Sol price is promotional; DeepSeek's varies by time of day.

Why it matters

On our reading, Woolf's argument means the cache-read line on an agentic-coding bill deserves as much attention as the headline input price. In his words, open-weight models and other inference providers "will continue to serve as an economic check on the prices of LLMs".