OpenAI's new model runs tasks about nine times cheaper than Claude Opus 5.5, the model in Anthropic's S-1
OpenAI published Terra's price sheet the same day Anthropic's S-1 went public: frontier capability at a tenth of the flagship cost, against a prospectus priced for a $1.5 trillion valuation.
Richard Tang · 3 min read
The price of frontier intelligence fell again, and it fell on the one day this year when the market's only public price list for an AI lab was open for reading. On Monday, October 26, Anthropic$350B — Anthropic, private, latest valuation $350B filed its S-1 for a Nasdaq debut under the ticker ANT 4; the same day, OpenAI$500B — OpenAI, private, latest valuation $500B shipped GPT-6.1 Terra, a model that reads screens and acts across applications, priced 50 percent below GPT-6.1 Sol 1.
Investors deciding whether to buy Anthropic at its mid-November debut, or to hold Microsoft+0.15% — Microsoft, up 0.15 percent today into its October 28 results, are now pricing a revenue line their rival reprices weekly.
A tenth of the flagship price is the number that matters
Not the headline cut. OpenAI's own evaluations say Terra matched its flagship GPT-6 Astra on Agents' Last Exam, a benchmark of professional tasks, at roughly one-tenth of Astra's price per token, and ran tasks roughly nine times cheaper than Claude Opus 5.5, the model Anthropic's prospectus sells 2.
Terra also scored 15 percentage points above GPT-5.6 Terra at about a third of the cost per task 1, and cut its factual error rate from 53.6 percent at low reasoning effort to 16.7 percent, falling to 4.3 percent at maximum effort, per OpenAI's reported evaluations 2.
Terra cut factual errors from over half of prompts to under 5
Data
| Value | |
|---|---|
| GPT-5.6 Terra, low effort | 53.6 |
| Terra, low effort | 16.7 |
| Terra, max effort | 4.3 |
What changed for buyers is concrete: a model that can see a screen and operate software makes computer-use agents practical, but practicality depends on the cost per task, and Terra drops that cost by an order of magnitude. Capability stopped being the constraint; the price of running it now is.
The prospectus is priced for a market Terra is repricing
For Anthropic, the arithmetic is uncomfortable. Its filing reports third-quarter revenue of $16.8 billion, up 45 percent, against an $80.7 billion first-half loss and valuation talk that has slipped toward $1.5 trillion from $2 trillion 3.
A price war already cost it most of Replit's reported $100-plus million a year in coding spend, with discounts reported up to 90 percent in individual cases 3. Terra extends that discounting from contracts to the published price sheet, and nothing in the filing shows how Anthropic's revenue survives a rival selling comparable capability at a tenth of the cost.
Both sides hold real evidence
Falling prices do expand usage; Anthropic's own prospectus reports that cheaper tasks drove adoption, with net revenue retention near 900 percent 5. But a price sheet that halves in a quarter turns a valuation into a volume bet nobody has published numbers for.
The decision rule before buying Anthropic or holding Microsoft: check the prospectus revisions through mid-November for pricing and margin language, and read Microsoft's October 28 cloud results for what enterprises actually pay per token. If margins hold while usage rises, Anthropic's line survives; if revenue per token keeps falling faster than usage grows, the valuation is the thing repriced, not the intelligence.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
