Ten cents a million: the AI price floor has been reached, and the bill is moving elsewhere

On October 7, 2026, Anthropic put Claude Haiku 5.5 into service at $0.10 per million input tokens and $0.50 for output. That is, to the cent, the price of GPT-6 Luna, launched by OpenAI on September 22. The two leading American labs now post the same floor price, and Anthropic’s is ten times lower than its previous floor (Haiku 4.5, at $1). In the same interval, OpenAI opened talks to raise at least $30 billion at a valuation of about $1.4 trillion, and Google signed for 3.6 gigawatts of electricity. On September 2, we wrote that the price of a token had stopped being comparable from one generation to the next. Six weeks later, the finding has hardened: at the floor, the posted price no longer separates the suppliers. What separates them is everything that is not written in large type on the rate card.

The floor has been reached twice in three weeks

September’s sequence was unusually dense. OpenAI launched GPT-6 Astra on September 3 at $10 / $50, then GPT-6 Sol and GPT-6 Luna on September 22, at $2 / $10 and $0.10 / $0.50. Seven days later, on September 29, GPT-6.1 Sol succeeded Sol at the same base rate, with cached input brought down from $0.20 to $0.10. Anthropic answered in three steps: Claude Opus 5.5 on September 22 at $4 / $20, 20% less than Opus 5 ($5 / $25); Claude Sonnet 5.5 on September 28 at $2 / $10; and Haiku 5.5 on October 7.

Haiku’s jump is the most spectacular. Its predecessor, Haiku 4.5, is still billed at $1 / $5. The nominal cut is 90%. The correction we documented in September has to be applied: Claude 4.7 and later models use a tokenizer that produces, according to Anthropic, about 30% more tokens for the same text, and Haiku 4.5 uses the old one. For the same document, Haiku 5.5 therefore costs the equivalent of $0.13 where Haiku 4.5 cost $1: the real reduction on short prompts is about 87%. It is still spectacular. But it comes with a condition.

That condition is a 100’000-token threshold. Beyond it, each Haiku 5.5 request switches entirely to a second rate card, $0.50 / $2.50, five times more expensive. The threshold counts all input tokens, including those read from or written to the cache, and is assessed request by request. A workload that sends long case files, an agent that accumulates its history, the processing of contractual documents: these are precisely the uses where the floor price does not apply.

What separates them now: the multipliers

When two rate cards post the same figure, the gap hides in the secondary lines. Three of them have moved over the past three weeks.

Cache reads. On Opus 5.5 and Sonnet 5.5, Anthropic now bills a cache read at 0.05 times the input price, against 0.1 times on previous generations: $0.20 per million on Opus 5.5, $0.10 on Sonnet 5.5. On Haiku 5.5 as on GPT-6 Luna, it falls to $0.01. For an agentic workload whose context is re-read dozens of times, this multiplier weighs more heavily than the input price itself.

Reasoning tokens. On October 6, Mistral opened Mistral Large 4 in public preview, a model of 1.05 trillion parameters of which 52 billion are active per token, with a one-million-token context. Its list price is $1.36 / $4.18. The launch discount, announced for two weeks in Mistral’s changelog (with no precise end date; around October 20 if counted from launch), brings it down to $0.68 / $2.09. The price is attractive, and it has an expiry. But an independent measurement relayed on October 8 estimates that about 75% of output tokens are reasoning tokens, which the user never reads. The price per million says nothing about how many millions a task consumes: two models at the same rate can cost very different amounts on the same question. This ratio comes from a third-party benchmark (Artificial Analysis); we have not reproduced it.

Location. The most useful fact for a Swiss decision-maker is in Mistral’s rate card: according to its documentation, the regional EU or US option is billed at 1.1 times the list price. On Large 4’s output, that gives about $4.60 per million (1.1 times $4.18), against $2.09 at the launch rate: more than double as soon as you require the data to stay in Europe. Mistral does not say whether the launch discount applies to regional endpoints. In September, we noted the same 10% premium at Anthropic, on its API as at Amazon Bedrock and Google Cloud. It is still in place. Geographic control has become a standard line item at American suppliers and at the European champion alike.

The next increase already has a date: January 1, 2027

Google, for its part, has not brought its rate cards down to the same floor. Gemini 3.6 Flash is billed at $0.75 / $3.75 until December 31, 2026, and Google’s pricing page provides that from January 1, 2027, the rate of $1.50 / $7.50 will apply: a doubling. Gemini 3.1 Pro stays at $2 / $12 for prompts of up to 200’000 tokens.

September’s precedent argues for caution in both directions: the scheduled increase on Claude Sonnet 5 was cancelled at maturity. But an organisation that builds its 2027 budget today on Gemini 3.6 Flash’s current rate is betting on a cancellation that has not been announced. Google is cutting its image prices, however: on October 6, Nano Banana 2.1 halved the cost of a 1K image, from $0.067 to $0.0336. At 4K, the cut is only about a quarter, from $0.151 to $0.113.

On subscriptions, nothing has moved on entry prices: ChatGPT Plus and Claude Pro stay at $20 a month, Google AI Pro at $19.99 in the United States. What changes is what the price buys. OpenAI began rolling out GPT-6 in ChatGPT on October 7, with GPT-6 Luna announced for the Free and Go plans ($8) from October 8. And the ranges are being structured around tiers at $100 and $200 a month (Claude Max, ChatGPT Pro), OpenAI having added a $500 Pro tier on September 29. The ChatGPT rollout and prices come from OpenAI’s release notes as relayed by the press; we could not consult the ChatGPT help page directly.

Beneath the floor: compute that stays expensive, energy that gets reserved

The price of the token is falling; the compute that produces it remains expensive and volatile. The Ornn index, a daily index built from GPU rentals actually concluded and listed on Bloomberg, stood on October 11 at $2.65 an hour for an H100, $5.61 for an H200 and $7.45 for a B200. An hour of B200 still costs nearly three times an hour of H100. Posted on-demand rates are also still scattered: at RunPod, on its community cloud, the H200 is listed at $3.59 an hour and the B300 at $6.94.

Energy is the second bottleneck. On October 6, Google and Constellation Energy announced a deal covering 3’590 megawatts on the PJM grid, the largest in the United States. The fine reading matters: only 890 megawatts are new nuclear capacity, coming from the power uprating of eleven reactors across six plants in Illinois, Pennsylvania and New Jersey, under a twenty-year contract and with a first delivery expected in 2028. The remaining 2’700 megawatts fall under a fifteen-year contract on Constellation’s existing fleet. The deal triggers more than $4.3 billion of investment at Constellation. In other words, the hyperscalers are reserving the electricity of the coming years today, and it is those commitments that today’s floor prices will have to amortise.

Capital, for its part, looks for no floor

On September 29, Bloomberg reported that OpenAI was discussing a raise of at least $30 billion at a valuation of about $1.4 trillion (excluding the money raised), after $122 billion raised in March at $852 billion. The talks are at an early stage, and Sam Altman has ruled out an IPO in 2026. OpenAI’s annualised revenue run rate passed $40 billion in August (Bloomberg, August 13), but estimates have diverged since: nearly $70 billion at the end of September according to Axios, about $50 billion according to the Financial Times on October 8. These figures are press estimates, not published accounts. Anthropic, for its part, committed in May to pay SpaceX $1.25 billion a month, until May 2029, to rent the compute capacity of the Colossus data centres, according to SpaceX’s IPO filing made public on May 20.

Below that, the market is segmenting. Factory, which specialises in AI agents for software development, raised $200 million at a $5 billion valuation (September 15); Nous Research, $90 million at $1.5 billion (October 7) to take its open-source Hermes agent to businesses. On October 10 and 11, the Financial Times and then the Wall Street Journal reported that Nvidia was discussing a further investment in, or a purchase of, Reflection AI, a specialist in open-weight models, and that Amazon was negotiating the purchase of the startup Decart for about $7 billion. These last two transactions are at the stage of reported discussions, not confirmed by the parties.

This juxtaposition is the key to reading the period. Suppliers can sell a million tokens for ten cents because their revenues are rising and their backers are financing the difference. The floor price is not a cost price: it is a market position, financed by capital, and therefore revisable.

API price comparison table

Model	Input ($/M tokens)	Output ($/M tokens)	Note
GPT-6 Astra	10.00	50.00	Launched September 3; cached input $1.00
GPT-6.1 Sol	2.00	10.00	September 29, follows GPT-6 Sol; cache $0.10 ($0.20 on Sol)
GPT-6 Luna	0.10	0.50	September 22; cache $0.01
Claude Fable 5.1	10.00	50.00	Cache reads at $0.25
Claude Opus 5.5	4.00	20.00	September 22, 20% less than Opus 5; cache $0.20; fast mode 8 / 40
Claude Sonnet 5.5	2.00	10.00	September 28; cache $0.10
Claude Haiku 5.5	0.10	0.50	October 7; 0.50 / 2.50 beyond 100k tokens
Claude Haiku 4.5	1.00	5.00	Old tokenizer, not comparable line by line
Gemini 3.1 Pro	2.00	12.00	Prompts up to 200k tokens
Gemini 3.6 Flash	0.75	3.75	Until December 31, 2026; 1.50 / 7.50 from January 1, 2027
Mistral Large 4 (preview)	0.68	2.09	50% launch discount for two weeks; list 1.36 / 4.18; regional EU or US ×1.1 of list
Grok 4.7	2.00	6.00	xAI flagship; 4.00 / 12.00 from 200k tokens (whole request)
DeepSeek V4 Pro	1.32	3.96	Peak-hours rate, half that off-peak
DeepSeek V4.1 Flash	0.15 to 0.30	0.60 to 1.20	Off-peak / peak hours

Prices taken in October 2026 from the official pricing pages of Anthropic, OpenAI, Google, xAI and DeepSeek, and from Mistral’s official model page (list price; Mistral’s launch price is calculated from the announced 50% discount). Claude 4.7 and later models count about 30% more tokens for the same text; reasoning models bill tokens the user does not read. The lines in this table therefore do not measure the same thing and cannot be compared without testing on your own workload.

What a Swiss decision-maker should take away

In three weeks, the two big American labs have converged on the same floor price, ten times lower at Anthropic than the previous one. That is good news for all mass processing: classification, summarisation, extraction, mail sorting. But three reservations are in order. The floor has thresholds: beyond 100’000 tokens per request, Haiku 5.5 costs five times more. The floor has dates: Gemini 3.6 Flash will double on January 1, and Mistral Large 4 will return to its list price when its launch discount ends. The floor does not apply to sovereignty: requiring the data to stay in Europe costs 10% more than the list price, and, according to Mistral’s documentation, that surcharge is calculated on the list price and not on the discounted price.

These three reservations lead to the same conclusion as last month, with new urgency. When every supplier posts the same price, the advantage goes to whoever can route each task to the right model, measure what it really costs and change supplier the moment a rate card moves. The floor has been reached; the margin is now earned in orchestration.