In June, OpenAI listed its flagship at $5 per million input tokens and $30 per million output tokens. By the end of September, its replacement was listed at $2 and $10. Anthropic cut too. A widely shared analysis published on October 25, 2026 by Max Woolf lays the whole timeline side by side, and the pattern is stranger than a simple price war.
The most important finding is not the headline percentage. It is that for coding agents, one line of the price sheet decides almost everything: the price of cached input. This guide walks through the dated numbers, what is vendor-listed versus the author's inference, and how to use the cuts.
Quick summary: Per Max Woolf's October 25, 2026 analysis, GPT-5.6 Luna was listed with an 80% cut on July 30 (Woolf flags that entry with "???"), Claude Opus 5.5 is 40 to 60% cheaper in total cost than Opus 5, and the GPT-6.1 Sol family is about 75% cheaper than the GPT-5.6 family from June (his estimate). In agentic coding, cache reads are 96 to 99% of token spend. DeepSeek V4.1 Flash 1017 is hosted by 29 providers on OpenRouter. Prices are per million tokens as listed by the vendors on the dates shown.
The timeline, price by price
Woolf lists prices as input / output / cache read per million tokens. The table below reproduces his dated entries. (Source: Max Woolf)
| Date (2026) | Model and event | Input / output / cache read |
|---|---|---|
| June 26 | OpenAI GPT-5.6 Sol launches | $5 / $30 / $0.50 |
| June 26 | OpenAI GPT-5.6 Luna (small model) launches | $1 / $6 / $0.10 |
| June 30 | Anthropic Claude Sonnet 5 launch pricing | $2 / $10 / $0.20 |
| July 24 | Anthropic Claude Opus 5 launches | $5 / $25 / $0.50 |
| July 30 | GPT-5.6 Luna cut by 80% (Woolf marks this entry "???") | $0.20 / $1.20 / $0.02 |
| August 21 | GPT-5.6 Sol cut about 20% (for three months) | $4 / $20 / $0.40 |
| September 1 | Anthropic Fable 5.1, cache read 75% below Fable 5 | $10 / $50 / $0.25 |
| September 3 | OpenAI GPT-6 Astra | $10 / $50 / $1 |
| September 22 | OpenAI GPT-6 Sol and GPT-6 Luna | $2 / $10 / $0.20 and $0.10 / $0.50 / $0.01 |
| September 22 | Anthropic Claude Opus 5.5 | $4 / $20 / $0.20 |
| September 29 | OpenAI GPT-6.1 Sol, cache price halved | $2 / $10 / $0.10 |
Reading the first and last rows together: GPT-5.6 Sol at $5 / $30 / $0.50 became GPT-6.1 Sol at $2 / $10 / $0.10. Woolf calls the GPT-6.1 Sol family about 75% cheaper than the GPT-5.6 family from June, which is his cumulative estimate and not an OpenAI statement. (Source: Max Woolf)
Why cached input is the number that matters
Woolf points out that in agentic coding, cache reads make up 96 to 99% of total token expenditure, with cache hit rates typically 90 to 100% depending on the inference provider. That is because an agent re-sends the same growing conversation on every step, and only the newest slice is fresh text. The consequence is a rule of thumb from his analysis: a 75% cut in cache-read cost correlates with a 70 to 75% cut in overall model cost. (Source: Max Woolf)
That explains why a headline like "Fable 5.1 cache read cut 75%" matters more to a developer than a cut to the output price. It also explains an odd detail: GPT-6 Astra on September 3 launched with no lower cache-read cost, at $1, and the cheaper Sol and Luna families were the ones that carried the discounts.
Not just OpenAI and Anthropic
- Opus 5.5 versus Opus 5: Woolf describes a 40 to 60% total cost reduction, combining a 20% input and output cut with a larger cache-read cut.
- Vercel gateway data cited by Woolf: cost per token fell to 43.3% of April 2026 levels, though aggregate spend rose because models improved and usage grew.
- DeepSeek V4 Flash 0420 (April): $0.14 / $0.28 / $0.0028. DeepSeek V4.1 Flash 1017: $0.15 / $0.60 / $0.003, with its Artificial Analysis index score rising from 24 to 45.
- Provider count on OpenRouter: 13 for V4 Flash in May, 29 for V4.1 Flash 1017. Cloudflare and DeepInfra are within 10% of DeepSeek's effective cost, though they hold only a minority of token share.
- Kimi K3.1 Pro: $2 / $10 / $0.20, index score 55 versus 58 for Opus 5.5, at half the price of Opus 5.5.
Why are they cutting? The author's hypotheses
Woolf is explicit that these are hypotheses. His list: models are commoditizing and switching is easy through tools like OpenCode and OpenRouter; the July Luna cut was likely aimed at small open-weight models such as DeepSeek Flash; efficiency gains in compute may explain part of it; and inference margins may be healthier than the headline losses suggest, with reports citing up to 80% margins, with caveats. He rejects the theory that prices will spike later, arguing providers will keep cutting until prices approach marginal cost. (Source: Max Woolf)
He also finds the timing odd, since leaked financials reportedly show both companies losing billions while both are rumored to be heading for IPOs. The analysis is mostly qualitative, built on published list prices, Artificial Analysis indices, OpenRouter provider counts and Vercel data. It is not an audit of any company's margins.
The subscription side: cheaper tokens, smaller plans
One of the most useful personal data points in the piece: Woolf says he reduced his Codex plan's usage to one-fifth after efficiency gains while his spending stayed roughly flat, and that OpenAI cut the $200 per month tier's allowance by 50%, citing more efficient models. The lesson is that cheaper tokens do not always show up as bigger subscriptions; sometimes the same price buys a smaller allowance of a more efficient model. (Source: Max Woolf)
A daily news digest on October 26 summarized the same piece and added that the cuts coincide with price pressure from cheap open-weight rivals such as DeepSeek and Kimi. (Sources: dentro.de, Bitdoze)
What to do with this
If you pay per token, audit your cache hit rate before you shop for a cheaper model. Keeping a stable system prompt and tool list at the start of every request can move more money than switching vendors.
- Reprice your workloads quarterly. A model that was too costly in June may be cheap now.
- Compare effective cost, not list price: DeepSeek's lower cache share means its real cost is even lower than the table suggests, per Woolf.
- Do not lock into annual commitments on a single model family while list prices keep falling.
- Treat predictions of a coming price hike as unproven; the author argues against them.
01How much have OpenAI and Anthropic cut prices?
Per Max Woolf, GPT-5.6 Luna was listed with an 80% cut on July 30, 2026 (Woolf marks it "???"), and Claude Opus 5.5 reduced total cost 40 to 60% versus Opus 5. He estimates the GPT-6.1 Sol family is about 75% cheaper than the GPT-5.6 family from June.
02What is the GPT-6.1 Sol price?
Woolf lists GPT-6.1 Sol, launched September 29, 2026, at $2 input, $10 output and $0.10 cache read per million tokens.
03Why does cache-read price matter so much?
In agentic coding, cache reads are 96 to 99% of token spend, according to Woolf, so a 75% cut to cache reads correlates to a 70 to 75% cut in overall cost.
04Will AI prices go back up?
Woolf argues a sharp increase is unlikely, since competition and cheap open-weight models push prices toward marginal cost. That is his opinion, not a guarantee.
05Are cheaper tokens making subscriptions cheaper?
Not necessarily. Woolf notes OpenAI cut the $200 per month tier's allowance by 50% citing more efficient models, so some plans kept their price but changed the allowance.
With prices moving this fast, the sensible habit is to keep options open. LumiChats offers 40+ premium AI models side by side in one subscription, so you can test a newly discounted model against your current favorite without opening another account.
