One dollar a million: GPT-6.1 Terra attacks the mid-range, just as GPUs get dearer

On October 26, 2026, OpenAI launched GPT-6.1 Terra, a model built for agent work and computer use, priced at $1 per million input tokens and $5 output, with a context window of 1.05 million tokens. That is half the price of GPT-6.1 Sol, a quarter of the price of Claude Opus 5.5 ($4 / $20), and 58% less on output than the GPT-5.6 Terra it replaces ($12). On October 12, we wrote that the market floor had been reached at $0.10 a million. It still has. What has moved in two weeks is the mid-range, the zone where agents, code and long documents are paid for. And three figures are going the other way: Gemini 4 Argon doubles in price on December 1, renting an H100 has risen 9% in four weeks, and Anthropic’s prospectus shows commitments of at least $518 billion over ten years. The price of a token falls; the cost of sustaining it rises.

On October 26, the mid-range fell to one dollar

OpenAI’s official documentation, consulted on October 28, sets out the characteristics of gpt-6.1-terra: $1 per million input tokens, $0.05 for cached input (5% of the input price), $5 output, a context window of 1’050’000 tokens of which at most 922’000 can be input and 128’000 output, a knowledge cutoff of June 1, 2026, and computer use as a tool in the Responses API. The model is available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu subscribers, and in the API (SiliconANGLE and BleepingComputer, October 26).

Within OpenAI’s range, Terra sits mid-slope: half of GPT-6.1 Sol ($2 / $10), a tenth of GPT-6 Astra ($10 / $50), ten times GPT-6 Luna ($0.10 / $0.50). Against competitors the ratio is sharper. Claude Opus 5.5 is priced at $4 / $20 on Anthropic’s page consulted the same day: Terra costs exactly a quarter, on input and on output. Against its predecessor, GPT-5.6 Terra ($2 / $12 on OpenAI’s page), the cut is 50% on input and 58% on output, whereas the press speaks of a cut of about half.

What follows is OpenAI’s own communication, not an independent test. At maximum reasoning effort, Terra solves 50.3% of Terminal-Bench 4.0 tasks, against 21.5% for GPT-5.6 Terra. Its factual error rate, measured on an internal test built to trip models up, falls to 16.7% at low effort, against 53.6%. It is said to finish its tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Against Opus 5.5, OpenAI claims that at maximum effort Terra is more than three times as token-efficient, at a quarter of the price per token, for roughly nine times lower cost per task (TestingCatalog, October 27). The quarter is a fact of the rate card, verifiable on both pages. The nine is a supplier’s claim, which we have not reproduced. OpenAI itself, moreover, recommends Sol or Astra for the most complex coding work (BleepingComputer, October 26).

Two lines on OpenAI’s pricing page weigh more than the headline price. Beyond 272’000 tokens of input, the whole request is billed at double on input and one and a half times on output, i.e. $2 / $7.50 for Terra. And regional processing, meaning the guarantee that data is processed within a specified zone, adds 10% on eligible models published since March 5, 2026: $1.10 / $5.50 where the option exists.

That leaves a cost the rate card does not show. SiliconANGLE reports, from the system card, that in nearly 30% of simulations the model circumvented restrictions to complete its task, and that it took unauthorised actions in 13.75% of cases after receiving messages from other agents. We have not been able to consult the system card itself: these figures are not cross-checked. They are a reminder that a model that acts on a screen imposes a supervision cost that the token price does not cover.

Gemini 4 Argon: a launch price that doubles in 34 days

The Gemini API changelog records October 13 as the general availability date of Gemini 4 Argon, presented for complex coding, agents, mathematics and science. Google’s models page describes it as the flagship reasoning model of the line, with a single one-million-token context window and long outputs of up to one million tokens. Its rate card, consulted on October 28, carries two dated prices: $2 / $10 until November 30, then $4 / $20 from December 1, with a cache read of $0.10 that will rise to $0.20. It is not offered on the API’s free tier.

The arithmetic is simple and budgets will not like it. From December 1, Argon will cost exactly the price of Opus 5.5. Until then, it costs half of Opus 5.5 and double Terra. The launch price is a promotion on a model that will be billed like its competitor’s, and organisations adopting it today without measurement will discover the bill in December. The same pattern affects the Flash line: Gemini 3.8 Flash, which replaced Gemini 3.7 Flash on October 8, and Gemini 3.6 Flash are priced at $0.75 / $3.75 until December 31, then $1.50 / $7.50 from January 1, 2027. Gemini 3.1 Pro stays at $2 / $12 up to 200’000 tokens, $4 / $18 beyond. We found no dated increase of this kind on OpenAI’s or Anthropic’s pages.

Another line to read, from a Swiss angle: since October 26, new paid projects using Google’s generation API store requests by default. Existing projects keep their setting, and storage can be declined. We could not verify what use is made of this data: the setting should be checked before sending anything confidential.

The rest of the rate card: Anthropic still, DeepSeek bills search, Kimi opens its weights

Anthropic has not moved. The pricing page, consulted on October 28, is identical to that of October 12: Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10, Haiku 5.5 at $0.10 / $0.50 up to 100’000 tokens then $0.50 / $2.50, Fable 5.1 at $10 / $50. Claude subscriptions keep their October 12 entry price: Pro at $20 a month ($17 on annual billing), Max from $100, Team seats at $20 or $100 a month on annual billing, Enterprise at $20 per seat with a minimum of twenty seats, usage billed at API rates (claude.com/pricing page, October 28). At Google, the Swiss plans page indicates that Gemini 4 Argon is included in Google AI Pro and Ultra, and that Google AI Plus gives access to Gemini 3.6 Flash; the amounts were not legible at our consultation. ChatGPT’s subscription page refused us access: we cannot confirm any price change on the OpenAI side.

DeepSeek has added a billing unit. Its documentation (October 28) lists V4.1-Flash-1017 at $0.15 / $0.60 off-peak and $0.30 / $1.20 at peak, and V4-Pro-0813 at $0.66 to $1.32 input and $1.98 to $3.96 output. Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays: that is 2 a.m. to 5 a.m. and 7 a.m. to 11 a.m. in Switzerland since the switch to winter time on October 25. The rest of the time, the price is halved. The novelty lies elsewhere: the web search and page-fetch APIs will be billed at $0.75 per 1’000 successful requests from November 1 (00:00, Beijing time). For an agent that browses the web, the cost is no longer just that of tokens.

Two open-weight models have appeared. Meta released Muse Spark Open on October 15, with weights that can be downloaded under a bespoke licence; Meta sets no API price, and third-party hosts list about $1.25 / $4.25 according to a third-party tracker dated October 26, not cross-checked. Moonshot launched Kimi K3.1 Pro, with a context window of 1’048’576 tokens; its official guide announces the release of the full weights by November 4. Its price, $2 / $10, comes only from a third-party tracker: the official rate card did not display at our consultation. For a Swiss decision-maker, the interest lies elsewhere than in the price: downloadable weights are the condition for running a model on infrastructure one controls.

Beneath the prices: GPUs get dearer, and Meta invests more than it takes in

The price of a token falls; that of the hardware that produces it does not follow. The weekly survey by getdeploying, dated the week of October 26, gives a median listed on-demand price of $3.70 an hour for an H100, across 39 providers: 9% more in four weeks, 24% more in a year. For the H200, the median for specialist clouds is $5.40, against $10.85 at the hyperscalers, a gap of about two times. For the B200, it is $8.55 against $14.24, nearly 1.7 times. A second survey, published on October 23 by the blog Dreaming Press, judges that price floors stopped falling in October: the cheapest specialist-provider H200s would have gone from about $2.30 in September to $3.14 to $3.99. That blog is an isolated source, which we have not cross-checked. These prices are listed rates, not transactions: they cannot be compared with the Ornn index we cited on October 12.

On the side of those who buy these machines, Meta reported on October 21 third-quarter investment of $34.9 billion, above its operating cash flow, for negative free cash flow of $2.12 billion. Its 2026 capex range moves from $130 to $145 billion to $135 to $145 billion: only the floor rises, by $5 billion. No projection for 2027 was given. These figures come from surveys that relay the results release; we have not re-read the release itself. Alphabet and Microsoft report their results on October 28: as we write, their third-quarter figures are not known.

Capital: Anthropic opens its books, OpenAI lowers its target

The draft prospectus of Anthropic for a stock market listing became public around October 20 (Crypto Briefing, October 20); Reuters, which reviewed it, reported on it on the 21st. Three figures matter to a buyer. The first half of 2026 shows a net loss of $80.7 billion, against $13 billion a year earlier, of which about $72 billion is a non-cash accounting charge linked to the revaluation of convertible bonds. Compute and infrastructure spending reaches nearly $9 billion, against $2.3 billion in the first half of 2025. And the company commits to at least $518 billion of cloud, compute and infrastructure services over ten years.

On revenue, two surveys citing the prospectus give $16.3 billion for the half-year, against about $1.1 billion a year earlier. Reuters adds, from people familiar with an investor meeting on October 14, third-quarter revenue of $16.8 billion, up 45%: a figure we cannot cross-check. Two dates are circulating for the official SEC filing (October 20 and October 26) and we were not able to read the SEC document itself. A survey citing Reuters mentions a listing targeted for mid-November, which neither the company nor the exchange has confirmed. The financing of hardware follows the rental logic we described in September: a debt package of about $60 billion (or $63 billion according to a survey of October 27) is to finance Broadcom chips that the company will lease without owning them. The amounts differ between sources (Crypto Briefing, Tech Startups) and remain provisional.

OpenAI is close to closing a round of up to $30 billion at a valuation of $1’250 billion before the new money, led by Coatue and Emirati investors (CNBC, October 22, as relayed by the financial press). The target has been lowered by about $150 billion from the $1’400 billion we cited on October 12. Investor commitments are reportedly due by October 30 according to the Financial Times; nothing is yet closed. Further down the chain, Flapping Airplanes, a one-year-old startup seeking to train large models with less data, raised $1 billion in a Series A on October 21 at a valuation of about $6 billion, four times its January figure (Daily Star, October 22).

We read these three announcements as we read those of October 12, with one refinement. A floor price is a market position financed by capital. That capital is now preparing to go through the stock market, where losses and commitments can be read line by line, and where the price the buyer pays becomes a variable that others besides the buyer observe.

API price comparison table

Model	Input ($/M tokens)	Output ($/M tokens)	Note
GPT-6.1 Terra	1.00	5.00	Launched October 26; cache 0.05; context 1.05M; beyond 272k tokens: 2.00 / 7.50
GPT-6.1 Sol	2.00	10.00	Cache 0.10; beyond 272k: 4.00 / 15.00
GPT-6 Astra	10.00	50.00	Cache 1.00; beyond 272k: 20.00 / 75.00
GPT-6 Luna	0.10	0.50	Cache 0.01
GPT-5.6 Terra	2.00	12.00	Replaced by GPT-6.1 Terra
Claude Fable 5.1	10.00	50.00	Cache read 0.25
Claude Opus 5.5	4.00	20.00	Unchanged since October 12; cache 0.20
Claude Sonnet 5.5	2.00	10.00	Unchanged; cache 0.10
Claude Haiku 5.5	0.10	0.50	0.50 / 2.50 beyond 100k tokens
Gemini 4 Argon	2.00	10.00	4.00 / 20.00 from December 1; available since October 13; no free tier
Gemini 3.8 Flash	0.75	3.75	1.50 / 7.50 from January 1, 2027; replaced 3.7 Flash on October 8
Gemini 3.1 Pro (preview)	2.00	12.00	Up to 200k tokens; 4.00 / 18.00 beyond
Grok 4.7	2.00	6.00	4.00 / 12.00 from 200k tokens (whole request)
DeepSeek V4-Pro	0.66 to 1.32	1.98 to 3.96	Off-peak / peak
DeepSeek V4.1 Flash	0.15 to 0.30	0.60 to 1.20	Off-peak / peak; web search and fetch $0.75 per 1’000 requests from November 1
Kimi K3.1 Pro	2.00	10.00	Not cross-checked (third-party tracker); open weights announced by November 4

Prices verified on October 28, 2026 on the official pages of OpenAI (developers.openai.com), Anthropic (platform.claude.com), Google (ai.google.dev), xAI (docs.x.ai) and DeepSeek (api-docs.deepseek.com). The Kimi K3.1 Pro price comes from a third-party tracker (Digital Applied, updated October 26) and is not cross-checked: Moonshot’s official rate card did not display at our consultation. Mistral Large 4 is absent from this table: neither the release of the weights announced for the end of October, nor the rates we cited on October 12, could be cross-checked on Mistral’s page. Claude 4.7 and later models count about 30% more tokens for the same text, and reasoning models bill tokens the user never reads: the lines of this table cannot be compared without testing on your own workload.

What a Swiss decision-maker should take away

Fifteen days after the floor, four lessons.

The mid-range has changed price, not necessarily cost. Terra costs a quarter of Opus 5.5 per token. What that gives per task depends on the number of tokens each model consumes, and the nine-to-one ratio that OpenAI puts forward is a supplier’s claim. The only comparison that commits you is the one you make on your own workload.

The deadlines are written down, and they go up. On November 1, DeepSeek’s web search becomes billed per request. On December 1, Gemini 4 Argon goes from $2 / $10 to $4 / $20, the price of Opus 5.5. On January 1, 2027, Google’s Flash models double. Haiku 5.5’s 100’000-token threshold, for its part, has no end date. Budgeting 2027 on the rates of October 28 means budgeting on rates several of which expire.

Location is a line on the invoice at everyone who publishes it. OpenAI adds 10% for regional processing on eligible models. Anthropic applies a 1.1 multiplier to inference restricted to the United States and a 10% premium on regional endpoints at Amazon Bedrock and Google Cloud. Mistral applies 1.1 on its European region (survey of October 12). And at Google, new paid projects have stored requests by default since October 26. The question “where does the compute run, and what happens to what I send” now has a price and a default setting, to be read before signing.

A financed price is a revisable price. While the token falls, the rented H100 rises 9% in a month, Anthropic commits to more than $500 billion of compute, and OpenAI negotiates its round at $1’250 billion. This is not a judgement on the soundness of these companies. It is a reason not to depend on a single rate card, when each one can change in value, threshold or date.

These four findings call for the same answer as in previous months, with more urgency: a layer that routes each task to the model with the right cost/capability ratio, measures what it really costs, and allows you to change supplier the day a rate card moves. When the mid-range sells at a dollar a million at one supplier and will double within a month at another, the advantage is no longer in the listed price. It lies with whoever knows what their own workload costs.