What happened to model prices in October?

On 20 October 2026, Stork AI Daily's price monitoring caught Alibaba raising its list price for DeepSeek V4 Flash overnight, with no announcement. Input went from $0.20 to $0.44 per million tokens. Output went from $0.40 to $1.32, which is 3.3 times what it had been the day before.

The same month moved the other way at OpenAI. GPT-5.6 Terra launched in July at $2.50 per million input tokens and $15 per million output tokens; by 27 October apidog listed it at $2 and $12, and the newer GPT-6.1 Terra had arrived at $1 and $5.

• 3.3x

  DeepSeek V4 Flash output price on Alibaba's price list, $0.40 to $1.32 per million tokens, changed overnight with no announcement, reported 20 October.

  Source: Stork AI Daily, 20 Oct 2026
• $15 to $12

  GPT-5.6 Terra output price per million tokens, July launch to October. GPT-6.1 Terra launched at $5.

  Source: apidog, 27 Oct 2026

For a CFO the direction matters less than the variance. A company running an agent feature on any of these models had its gross margin move in October, and nobody in its finance team changed a number to make that happen.

How much of what you pay is the supplier's margin?

The Anthropic S-1, made public on 26 October 2026, lets anyone work this out. Tomasz Tunguz pulled the lines that matter. In the first half of 2026, revenue was $16,286 million and cost of revenue was $2,380 million, which gives a gross margin of about 85%. In fiscal 2025 the same lines were $4,585 million and $2,773 million, about 40%. Forrester's own calculation also lands at about 85% for the first half.

Both margins are derived from the filed revenue and cost of revenue. The filing does not disclose a gross margin, and the figures cover all of Anthropic's revenue, the API and everything else together.

Of each dollar Anthropic billed, what it cost to serve

Cost of revenue per $1 of revenue, US cents. Derived from S-1 figures as reported by Tomasz Tunguz.

• FY2025Revenue $4,585M
  60.5¢
• H1 2026Revenue $16,286M
  14.6¢
• 025¢50¢75¢100¢

An 85% gross margin is what a successful software supplier looks like. Read from your side of the invoice, it means that on these derived figures about 85 cents of each dollar Anthropic billed in the first half was margin rather than cost to serve. Every token in your agent run carries a share of that margin into your own cost of revenue, at a level the supplier sets. When its pricing changes, so does your gross margin.

What does one agent run cost you?

Suppliers bill per token, per month. You sell per seat, per customer or per contract. Between the two sits the unit that drives your cost: the run, meaning one agent completing one task for one customer. A run is several model calls and at least one tool call, plus the retries when a check fails. Its cost is the sum over its model calls of tokens in times the input price plus tokens out times the output price, times retries, plus the tool calls.

Forrester's first advice to buyers after the S-1, published on 27 October, was to budget on the cost of a finished task, expect that cost to fall, and retest it with each model release. It also told buyers to watch usage limits, which can change while the list price stays the same.

The figure below splits one run in a claims administration product into its steps. Four of the five call a model, and only the drafting step does the kind of reasoning that needs a frontier model. The other three read a document, sort it or check a draft. The policy lookup calls your own systems and uses no model tokens. When the check fails, the run goes back to drafting, and you pay for the most expensive step a second time.

One agent run, step by step, and who sets each step's price

What a run costs depends on your prompts, your traffic and your retry rate, so the number has to come from your own product.

Does every step need a frontier model?

Once a run is split into steps, extraction, routing and checking turn out to be work a small model can do. Cloudflare tested exactly this on a claims pipeline of its own, as Help Net Security reported on 20 October 2026. A cascade of small extraction and decision models cost $6,032 per million claims, against $11,740 for a pipeline that sent every step to a large language model, and each claim took 19.3 seconds instead of 61.7.

Those are Cloudflare's figures for its own Clairic models, from a vendor testing its own product, and no independent test has been reported. The shape is still worth testing on your own runs.

Cloudflare's own test: small models in the pipeline, half the cost and a third of the time

Cost per million claims, USD

• All-LLM pipeline
  $11,740
• CascadeSmall extraction and decision models
  $6,032
• 03,0006,0009,00012,000

Time per claim, seconds

• All-LLM pipeline
  61.7 s
• CascadeSmall extraction and decision models
  19.3 s
• 020406080

Vercel's AI Gateway Production Index, published on 15 October 2026, shows a similar split in live traffic. In September 2026, open-weight models carried 71.5% of the tokens on its gateway and 23% of the spend, and DeepSeek models alone carried 58% of the tokens. Vercel estimates spend at list prices, counts only traffic through its own gateway, and defines open-weight more broadly than earlier reports did. Read it as a direction across many companies, and run the numbers for your own product before you rely on it.

Which steps can nobody else reprice?

A step served by an open-weight model on your own infrastructure has a cost you set: the compute you contract for, running weights you hold. If a supplier changes its price list overnight, that step costs what your compute contract says it costs.

Steps that stay on a frontier API keep their exposure, and after the exercise above you know how large it is, step by step. Some steps will stay there because the reasoning needs a frontier model and the price is worth paying. For the steps that read, sort and check, the question becomes which model to run and where to run it. Many vertical domains already have open-weight models trained for them, and we compared them in our survey of domain-trained open-weight models.

The steps you run on weights you hold are the only steps in the run that nobody else can reprice overnight.

The choice is made one step at a time, with the cost of the run in front of the CEO and the CFO together.

What should the CEO and CFO ask for this week?

1. Price one run. The cost of one finished run of your most used AI feature, by model and by provider. Owner: the CFO.
2. Split it into steps. That run broken into its steps, each marked as frontier reasoning or as extraction, routing or classification. Owner: the CEO, with the CPO.
3. Move one step. One extraction or classification step moved to an open-weight model on your own infrastructure, with the cost of the run measured before and after. Owners: the CPO and the CFO.

SynOS runs inside your product and prices every run, per model and per provider, so the AI line in your gross margin has a number per feature and per customer. The runs it records stay on your own infrastructure in the shape a training set is built from, ready for the steps you decide to move to a model you own.

Questions