Lay three price tags side by side. On September 22 Anthropic released Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Six days later came Sonnet 5.5, at $2 and $10. Today, October 7, the family is complete with Haiku 5.5, at $0.10 and $0.50 for any prompt up to 100,000 tokens.321 Three models in fifteen days, with a fortyfold spread in price from the top tag to the bottom one. The smallest tag carries a footnote worth reading: Haiku 5.5 is priced 90 percent below Haiku 4.5 on prompts up to 100,000 tokens, and 50 percent below on longer ones.1 A year ago the small model cost $1 and $5.17
A common story about prices like these runs as follows. Open-weights models, the ones anyone can download and run, cannot quite catch the closed frontier, so they compete on price instead, and that pressure drags the closed labs' prices down. The story blends two different gaps. One is how far behind open weights are: the distance, on a shared yardstick, between an open model and the closed model it is weighed against. The other is how much less they charge: the difference in list price per token. The story assumes the second gap is wide wherever the first one is. Today's price lists let you check.
My argument is that the story holds in the middle of the market and fails at the bottom. In the middle, the strongest open model trails the closed mid-tier by about a dozen points and charges roughly half as much per output token. At the bottom, the score gap has shrunk to a point or a few, and the price gap has closed altogether: Haiku 5.5 and OpenAI's GPT-6 Luna list at or below the cheapest open models. The prices that set the pace are being set by closed labs pricing against each other. This is a claim about list prices per token on October 7. It is not a claim about what a finished task costs, which can come out quite differently, and it is not a claim about why anyone chose any price.
The open model is one point behind on the index.
And it charges no less on any line.
What the ladder is for
Why would one company sell three models whose prices span a factor of forty? Because buyers do not send every job to the same place. Anthropic describes Sonnet 5.5 as a faster, lower-cost complement to Opus, strongest at well-scoped everyday work, and it describes Haiku 5.5 as "designed for high-volume, cost-sensitive tasks," suited to running as a helper under the larger models.21 The ladder exists so that each request can be sent to the cheapest rung that will do, which is the point an earlier dispatch here made about routing rules: choosing a model is a spending decision.
To compare rungs you need a yardstick. Artificial Analysis, an independent benchmarking firm, publishes an Intelligence Index that rolls many evaluations into one score. On the version in use today it puts Opus 5.5, at its maximum effort setting, at 58, the highest score it has measured. Sonnet 5.5 at maximum effort scores 56: two points behind for half the price per token.65 But the same firm found that Sonnet 5.5 produced about 193,000 output tokens per index task, the most it had ever measured, which came to $7.60 a task and left it off the frontier of score against cost per task. Opus 5.5 used about 119,000, and four of its five effort settings sit on that frontier.56 Half the price per token did not mean half the bill. Keep that in mind; it returns at the bottom of the ladder.
Who sets the price
The two ends of Anthropic's line have moved differently. Opus has only fallen: $15 and $75 for Claude 3 Opus in March 2024, $5 and $25 for Opus 4.5 last November, $4 and $20 now.15163 Haiku went up before it came down: $0.25 and $1.25 for Claude 3 Haiku, $1 and $5 for Haiku 4.5, and now a tenth of that.15171 Today's release also halved Sonnet 5.5's price for cached input, to $0.10 per million tokens, which Anthropic says cuts the cost of most agentic tasks on Sonnet by about 20 percent.1
Anthropic's own explanations are about its costs, not its rivals. Opus 5.5, it says, "requires less compute to serve than Opus 5, and its pricing reflects that."3 The Haiku page names no competitor in its pricing section. Its only remark on value is that prompts under 100,000 tokens make up about 90 percent of the requests sent to the previous Haiku.1
The nearest match is closed. On September 22, the day Opus 5.5 appeared, OpenAI released GPT-6 Sol at $2 and $10 per million tokens and GPT-6 Luna at $0.10 and $0.50, roughly half what their GPT-5.6 predecessors cost.78 Sonnet 5.5 lists at Sol's price; Sonnet 5 had been there since June. Haiku 5.5 now lists at Luna's, to the cent. Artificial Analysis noticed: the new Haiku costs "the same as GPT-6 Luna."4 The sequence goes back further. In August Anthropic said it would keep Sonnet 5's introductory price rather than move to $3 and $15, less than two weeks after OpenAI cut its own prices on July 30. The Stack, reporting it on August 11, noted that Anthropic did not explain the decision.10 I am not supplying a motive either. The timing is what is on the record.
A wider study points the same way. Luke Emberson and David Roodman of Epoch AI, in a paper published September 22, estimate that the cost of reaching a given level of performance has fallen about 47 percent a quarter since 2023, or about thirteenfold a year. Tracing who undercut whom near the state of the art, they found "the most intense price competition near SOTA has been between closed-weight models." In only one of the eight cases they traced was the model that undercut the leader an open one. They add a fair caution: the closed race may still be helped along by pressure from open models.9
Nor are the open labs a one-way floor. In August DeepSeek moved its V4-Pro model from a flat $0.435 and $0.87 per million tokens to $0.66 and $1.98 off-peak, and double that at peak hours. Mozilla's State of Open Source AI report called it "the first list-price rise from a major open-weights lab."1113
Where open weights still win
None of this means open weights have stopped competing on price. In the middle of the ladder they plainly do. Z.ai's GLM-5.3, an open-weights model, lists at $1.40 per million input tokens and $4.40 per million output on Z.ai's own service, and scores 45 on the index.1214 Against Sonnet 5.5 it is 11 points behind and 56 percent cheaper per output token, and 30 percent cheaper on input. Against Opus 5.5 it is 13 points behind and 78 percent cheaper on output. That is the common story, and here it is true: a real gap in score, bought back with a real gap in price.
The middle is not uniform. Moonshot's Kimi K3, an open model with 2.8 trillion parameters, scores 44, about the same as the new Haiku, and lists at $3 and $15: half as much again as Sonnet 5.5 on both lines.418 It is 12 points behind Sonnet and charges more. Open weights is a licence, not a price.
The bottom rung
Now go down to where Figure 1 was printed. Artificial Analysis scored the small models today: Haiku 5.5 at maximum effort 43, Z.ai's open GLM-5.3 Flash 42, Google's Gemini 3.8 Flash 41, and GPT-6 Luna 38.4 DeepSeek's open V4.1 Flash scores 39.14 So the open models here are one to four points behind Haiku, which on a composite index is close to a tie.
The price gap runs the other way. GLM-5.3 Flash lists at $0.15 per million input tokens and $0.50 output. DeepSeek V4.1 Flash lists at $0.15 and $0.60 off-peak, and $0.30 and $1.20 at peak.1213 Haiku 5.5 is $0.10 and $0.50. On input both open models charge half as much again; on output GLM-5.3 Flash matches Haiku and DeepSeek charges 20 percent more off-peak. How much less they charge, on the main lines: nothing. DeepSeek does undercut Haiku on cached input, at $0.003 off-peak against $0.01, which matters for workloads that resend the same long prompt.
This is where the story turns over. At the bottom, open models are not trading a score gap for a price gap. They are about as good and no cheaper, and the two closed models at that price, Haiku and Luna, are each other's nearest rivals.
A token price is not a task price
Three facts keep the bottom rung from being a verdict. First, Haiku 5.5 is hungry. At maximum effort it used about 162,000 output tokens per index task, roughly three times Luna's 50,000. At its high setting it scores 38, level with Luna at maximum, on about 55,000.4 A model that spends three times the tokens can lose a per-token lead by the time the bill arrives. That is the argument of an earlier dispatch here, The cheaper query and the larger power bill, and it applies with full force.
Second, the price has a step in it. Above 100,000 tokens Haiku 5.5 costs $0.50 and $2.50, five times as much, which puts it above GLM-5.3 Flash and above DeepSeek's peak rate.11213 Anthropic's figure that nine requests in ten fall below the line describes the previous Haiku's traffic, not yours.
Third, Artificial Analysis has not yet published a cost per task for Haiku 5.5, and says its site does not yet reflect the tiered pricing.4 So nobody can say from public data today which small model is cheapest per finished job. What anyone can say is which is cheapest per token, and today that is a closed one.
Objections, considered
Self-hosting makes open weights cheaper than any list price. An open model's list price is not its only price; anyone with the hardware can run it, and at steady, high volume that can beat a rate card. But hardware, staff and idle capacity are real costs, and I have found no public measure of them, published by today, that I would trust to put a number on it. So the claim stays where the evidence is, with list prices. Running a model yourself for control over data or uptime is a good reason that this piece does not touch.
The closed labs cut prices because of open weights, whatever they say. Perhaps. Epoch itself allows that pressure from open models may help drive the closed race, and nothing here rules that out. The observable facts are narrower: one company's stated reason is the cost of serving, and its middle and bottom prices match a closed rival's to the cent. I would rather report that than guess at anyone's intentions.
Benchmarks do not measure your work. True. The index is a composite of other people's tests, and Artificial Analysis itself flagged one of Haiku's component scores as likely understated.4 Use it for direction, not for ranking. A gap of one to four points says "about the same"; a gap of eleven to thirteen says "not the same." The argument rests on that difference, not on any single score.
Where open models are a dozen points behind, they are half the price; where they are one point behind, they are no cheaper.
Before you pick the cheap model
- Compare on cost per finished task. A token price is an input; the bill is tokens times price, and token use varies threefold between models with similar scores.
- Know how many of your requests cross a tier. Haiku 5.5's price rises fivefold above 100,000 tokens. Your own share of traffic over that line decides which rate you actually pay.
- Put the effort setting in the cost model. The same model at a lower setting can match a rival's score on a third of the tokens.
- Re-check the bottom rung. A month ago an open model was the obvious cheap choice on list price. Today it is not.
- If you chose open weights for price, ask whether price is still the reason. Control, privacy and the option to run it yourself are good reasons too, but they are different ones.
Back to the three tags: $20, $10 and fifty cents per million output tokens. The middle tag has an open rival eleven points behind at less than half its price. The bottom tag has open rivals a point or a few behind and no cheaper, and a closed rival at exactly its price. At that end of the ladder, the frontier is pricing against itself.