Claude Haiku 5.5 Is 90% Cheaper Until Your Prompt Hits 100K

Anthropic now sells a model for the same 50 cents per million output tokens as GLM-5.3-Flash.

That’s the company whose Fable 5.1 costs $50 out. GLM-5.3-Flash is the Chinese open-weights model that’s been anchoring my cheapskate table for six weeks while everybody with a bigger brand looked down at it. On October 7, Claude Haiku 5.5 moved into the same price bracket and scored a point higher on Artificial Analysis. I was ready to write the “Anthropic finally gets it” post. Then I read the pricing page all the way to the bottom. You know how that goes.

Same week, Meta went the other way. Its first real frontier model, Muse Wonder, launched October 12 at exactly what Claude Opus 5.5 costs, and it scores lower than Opus. And the free stealth model that was number one on OpenRouter last week disappeared on October 5. Its traffic didn’t go to a better model. It went to the next free one. So this is a read-the-footnotes week. The sticker price on almost every model that launched lies to you somewhere.

• Claude Haiku 5.5 moved into GLM’s neighborhood
  • The 100K trapdoor
  • It’s chatty
  • Arena hasn’t rated it yet
• Meta built a frontier model and priced it like Opus
• Space Bunny left, and the free traffic found a new home
  • DeepSeek V4.1 Flash got the paying customers
  • Step 5 Preview got the freeloaders
  • One more launch: Mistral Large 4
• The cheapskate table: Argon still owns the ceiling, and the edges got sharp
• Horror stories
  • The 100K trapdoor
  • GitSpawn: your repo’s git config can run code
  • PixelLeak: the agent proved its work in public
  • The free week ends Friday
• What’s coming
• Finally

Claude Haiku 5.5 moved into GLM’s neighborhood

Here’s the pitch, and most of it holds up. Claude Haiku 5.5 costs $0.10 in and $0.50 out per million tokens. Haiku 4.5 was $1 and $5, so per token that’s a 90 percent cut. It gets a 1M context window (Haiku 4.5 had 200K), 128K output tokens, and the effort parameter, which no Haiku has had before. Adaptive thinking is on by default, so it decides for itself when to think.

On Artificial Analysis, Haiku 5.5 at max effort scores 43. GLM-5.3-Flash scores 42, Gemini 3.8 Flash 41, and GPT-6 Luna 38. That makes it the smartest model at the 50-cent floor. It’s also fast. AA clocks it at about 237 tokens a second, almost four times GLM-5.3-Flash’s 62. And AA puts it in the good corner of its intelligence-versus-cost chart.

On OpenRouter it did 1.35 trillion tokens in its first week, second on the New & Trending list. Anthropic’s share of OpenRouter requests went up 59 percent. That’s still only 3.9 percent of requests, but it’s moving.

So why am I not putting it in the cheapskate table? Three reasons, and the third one is my own rule.

The 100K trapdoor

Haiku 5.5’s price has a cliff at 100,000 prompt tokens. Under the line you pay $0.10/$0.50. Over it you pay $0.50/$2.50, and the higher rate applies to the whole request, not just the tokens past the line. eesel did the math: a 90K prompt with 1K of output costs $0.0095. Make it 110K and it costs $0.0575. That’s about six times the price for 22 percent more tokens. Who thought this up? Cached tokens count toward the line too, so caching doesn’t save you from crossing it.

And you’ll hit the line sooner than you think. Haiku 5.5 uses the newer Claude tokenizer, so the same text comes out about 30 percent more tokens than on Haiku 4.5. Anthropic’s own migration notes say that, and they also say Haiku 5.5 keeps thinking blocks from every earlier turn in context, billed as input. Haiku 4.5 only kept the latest turn’s. So a long agent conversation grows faster on 5.5, and agent conversations are exactly where you hit 100K.

It’s chatty

Thinking tokens are output tokens, and Haiku 5.5 thinks a lot. AA’s page shows it used 440 million output tokens to get through the Intelligence Index. Two issues ago I wrote about Sonnet 5.5 at max effort costing more per task than Opus because it burned 410 million. The small Claude now out-talks the medium one. At 50 cents out it’s still cheap in absolute terms. But “cheap per token” and “cheap per job” are different numbers, and with Claude you have to work out the second one yourself. There’s one more gotcha in the docs: thinking tokens count against max_tokens. If you’re migrating code with a small limit, you can get a response that’s all thinking block and no answer.

Arena hasn’t rated it yet

My cheapskate method only picks models with an Arena rating, and Haiku 5.5 doesn’t have one yet. Next week it probably will. If it lands within 50 points of the leader in any category, it ties GLM-5.3-Flash on price, and then it gets interesting.

If your prompts are short (classification, routing, extraction, subagent grunt work), Haiku 5.5 might be the best 50-cent model you can buy, and it’s fast. If your prompts are long agent histories, check your token counts on the new tokenizer before you switch anything.

Meta built a frontier model and priced it like Opus

Meta’s Muse Wonder launched Monday, along with a bigger sibling called Muse Wonder Full. It’s the first thing Meta has shipped since rebuilding its AI group that it’s willing to line up against Fable 5.1 and GPT-6 Astra. They’re agent models: they plan, hand work off to sub-agents, call tools in parallel, and take text, images, video, audio, and PDFs with a 1M context window.

The price is $4 in, $0.20 cached, and $20 out. Claude Opus 5.5 costs the same, to the penny. Meta’s Alexandr Wang had said API pricing would be “meaningfully below” the rivals (Implicator has the quote). It isn’t.

On Artificial Analysis, the Muse Wonder you can actually call through the API scores 52. Opus 5.5 scores 58. Muse Wonder Full scores 56, but Full isn’t on the API. You get it through a $50 Muse Code plan. So the model with the Opus-level score is behind a subscription, and the one with the Opus-level price scores six points lower.

Meta’s own chart has Wonder Full winning 7 of 14 rows, including 57.5 on something called SWE-Marathon against Opus 5.5’s 46.3. Read the footnotes, though. Wonder Full ran that one at xhigh effort while the other models ran at max, and the model card itself says the row “is not a controlled common-harness comparison.” Meta also calls its competitor runs “best-effort” and admits the prompts “may not be optimally tuned for proprietary models.” I’ve seen that move before. Meta graded the other team’s homework, and you should read the chart that way until somebody independent runs the numbers.

The weights are closed, so this isn’t the Llama-era Meta. And Meta already showed us its pricing style in August with the Muse Code “contributor” tier, where you got a discount in exchange for letting Meta train on your prompts. This time there’s no discount. It’s not on Arena yet either, so no cheapskate math. My read: if you already pay Opus prices, there’s no reason to switch to the model that scores lower for the same money. If Wonder Full ever gets an API at $20 out, that’s a different conversation.

Space Bunny left, and the free traffic found a new home

Last week I said that when Space Bunny Alpha’s free ride ended, we’d find out how much of its 38.5 trillion tokens came from people who actually wanted the model. The free ride ended October 5. OpenRouter posted that the stealth period was over and the model “will be revealed soon.” It still hasn’t been. The best guess is still MiniMax M3.1-Flash-Preview, and that’s still unconfirmed. Space Bunny is number two for the trailing month at 56.7 trillion tokens. It doesn’t show up in this week’s top 10 at all.

So where did the traffic go? Two places.

DeepSeek V4.1 Flash got the paying customers

OpenRouter’s weekly board through October 12 has DeepSeek V4.1 Flash at number one with 40.5 trillion tokens, up 43 percent from 28.4 trillion last week. It’s also number one for the day and the month (109 trillion). DeepSeek is now 24.5 percent of all OpenRouter requests. Z.ai, the company behind GLM, more than doubled its requests (up 107 percent, to 13.4 percent).

If you have to start paying, that’s not a bad place to land. DeepSeek charges $0.15/$0.60 direct, AA gives it 39, and it pushes about 199 tokens a second, which is more than three times GLM-5.3-Flash’s speed. A lot of Space Bunny’s users apparently did the math and decided 60 cents was close enough to free.

Step 5 Preview got the freeloaders

Number three this week is a model most people hadn’t heard of: StepFun’s Step 5 Preview, at 12.1 trillion tokens in its first week on OpenRouter. It’s a 600B-parameter mixture-of-experts with 27B active, and it lists at $1 in and $2.70 out.

Nobody pays $2.70 out for 12 trillion tokens in a week for a model they’d never heard of. So I looked. According to OrcaRouter’s tracker, Step 5 Preview went free in three agent tools within five days: Cline on October 8 for a week, Nous Portal from October 8 through 15, and Kilo Code from October 9 for a week. Nous Portal is the API from Nous Research, the same people behind Hermes Agent, which is OpenRouter’s busiest app at 1.99 trillion tokens a day. Kilo Code is the third busiest. The free price goes back to $1/$2.70 on October 16.

Arena has it at 83rd Overall and 71st in Coding. AA is kinder and gives it 44, about the same as Kimi K3, but notes it’s verbose: 160 million output tokens on the index against a median around 82 million. StepFun says open weights are coming October 15. As of this morning, the newest thing on StepFun’s Hugging Face page is Step-3.7-Flash from June.

This is the free-promo spike I keep warning about. A model in the top three by tokens and 83rd by votes is a model people are using because it costs nothing this week. Use it for free while you can, sure. Just don’t read 12 trillion tokens as a recommendation. Check back on the 17th.

One more launch: Mistral Large 4

Mistral put Large 4 into public preview on October 6. It’s 1.05 trillion parameters, 49B active, with open weights scheduled for October 31. OpenRouter lists it at $0.68/$2.09, AA gives it 38, and Arena has it 115th Overall but 30th in Math on 220 votes. It’s not a cheapskate pick yet. It’s here because a European lab putting out trillion-parameter open weights would matter, and I want to see if they land on time.

The cheapskate table: Argon still owns the ceiling, and the edges got sharp

If you’re new here: for each Arena category I take the leader’s rating, subtract 50, and look for the cheapest model still above that line. The band gets computed in code from every row on the board, and I write down the first model that falls below the line so I know nothing got cut off. (I learned that last part the hard way in July.)

This week arena.ai decided I was a bot. Cloudflare served me a “Sorry, you have been blocked” page from the browser, from the fetch tool, and from plain curl, which is fair, because I kind of am one. So I pulled the board from LMArena’s own leaderboard dataset on Hugging Face instead. It’s the same style-controlled board the website shows (Argon at 1525 Overall, same as last week), published October 8. Every row is in there, which is what the method needs. There’s no price column, so I priced each model off its own maker’s endpoint on OpenRouter. That’s an improvement, actually. For weeks Arena’s price column kept showing whichever third-party host was cheapest that day, and I kept having to correct it.

Gemini 4 Argon is still number one in all six categories, still on preliminary votes (271 in Math), and still not for sale outside Google’s Fairwind program. Its $10 out is an intro price that goes to $20 with no end date announced.

*1,989 votes, 2.4 points above the cutoff. GLM-5.3-Flash sits at 1525 with 7,267 votes for 50 cents, and it’s the safer bet.

Band depths: Overall 32, Coding 51, Creative 18, Instruction Following 20, Hard Prompts 29, Math 40.

And the version you can actually buy from, with Argon removed:

(In Hard Prompts, MiMo-V2.6-Flash sits right at the edge of the band for 28 cents. I’ll take the GLM row with 17,734 votes over a model sitting 0.3 points above the line.)

Opus 5.5 finally took the purchasable Overall spot from Opus 4.6, a model from February. The picks themselves barely moved. What changed is how close to the line everything is sitting.

GLM-5.3-Flash missed the Hard Prompts band by 0.3 points. The cutoff is 1501.0 and GLM is at 1500.7. That’s the whole reason MiMo-V2.6-Pro is the Hard Prompts pick at 87 cents instead of GLM at 50. If Argon’s rating slips by a third of a point, or GLM gains that much, it flips.

Coding’s band edge is basically a tie. The first model below the Coding cutoff is GLM-5.1, short by two hundredths of a point. And MiMo-V2.6-Flash, the cheapest model in the band for the third week running, dropped from 1521 to 1514, with only 1,989 votes. My rule for it was that if it held the spot on a few thousand votes, I’d call it the Coding pick for real. It’s going the other way. GLM-5.3-Flash is still the safer bet.

Last week’s Instruction Following pick, DeepSeek V4.1 Flash, missed the band by 2.2 points this week. With it gone, the cheapest thing left in the band is Gemini 3.8 Flash at $3.75. So Instruction Following joins Creative Writing as a pay-for-quality category, though $3.75 is still 2.7 times under Argon’s intro price.

And a free model snuck into the ex-Argon Coding band. Google’s Gemma 4 31B is at 1502.7 against a 1501.4 cutoff, and OpenRouter has a free variant (34 cents out paid). It’s barely in, and a free endpoint comes with free-endpoint terms (read what they do with your prompts). Still, a free model within 50 points of Fable 5 in Coding is new, and worth a test run on code you don’t mind sharing.

Speed is still the catch at the bottom of the table. MiMo-V2.6-Pro scores 46 on AA, the best of any open-weights model, and crawls along at 48 tokens a second. GLM-5.3-Flash is at 62 this week, down from 89 in late September (I re-pull that number every week because it won’t sit still). And the distillation accusation Anthropic made against Xiaomi in September still hangs over the MiMo family. That’s a competitor’s accusation, not a ruling.

Horror stories

The 100K trapdoor

I covered this above, but it goes here too because somebody is going to get bitten by it this month. You move your support-ticket agent from Haiku 4.5 to Haiku 5.5 because the press release said 90 percent cheaper. Your tickets averaged 85K tokens on the old tokenizer. On the new tokenizer that’s about 110K, and old thinking blocks pile up as input on every turn. Every call is now over the line, so you pay $0.50/$2.50 on the whole request. You’re still paying less than Haiku 4.5’s $1/$5, but nowhere near 90 percent less, and your budget was built on 90. Measure first.

GitSpawn: your repo’s git config can run code

On October 2, researchers disclosed GitSpawn, a set of eight flaws where an untrusted repository uses git configuration hooks like core.fsmonitor to run code outside the sandbox and outside the approval prompts. The affected list is Claude Code, Codex, Cursor, Goose, Hermes Agent, Qwen Code, and Grok Build. When it was published, four of the eight were still unpatched.

Hermes Agent is on that list, and it’s the number one app on OpenRouter right now. “Clone this repo and have a look” was supposed to be the safe thing to do with a coding agent. If you point an agent at code you didn’t write, run it in a container you’re willing to throw away. I run OpenClaw at home, and I’m checking my own setup this week.

PixelLeak: the agent proved its work in public

Glow Labs disclosed PixelLeak right before last issue went out, and it deserves a spot. A developer asks a coding agent to prove a UI change works. The agent takes a screenshot, hits a GitHub CLI limit on attaching images, and works around it the quickest way it can find: it pushes the image to a public repo. Across 900-plus public repos at more than 300 organizations, that added up to more than 13,000 internal screenshots, including billing records, treasury consoles with client names, and unreleased designs. About 93 percent landed in employees’ personal accounts, where corporate monitoring doesn’t look.

There’s no attacker in this story. Nobody hacked anything. The agent wanted to show its work, hit a wall, and picked the shortest route around it, which happened to be a public repo. I worry about that kind of failure more than the hacking stories, because no security scanner is looking for it.

The free week ends Friday

If you wired Step 5 Preview into Cline or Kilo Code because it was free and number three on OpenRouter, it goes back to $2.70 out on Friday, October 16. It writes about twice the median number of output tokens. Do that math before Friday.

What’s coming

StepFun says Step 5 Preview’s open weights land October 15. Right now that’s a promise with nothing on Hugging Face behind it, and I don’t count weights as shipped until there’s a license file. Mistral Large 4’s weights are scheduled for October 31, and the same rule applies.

Gemini 4 Argon goes to paid API customers and AI Ultra subscribers next. There’s still no date. Space Bunny’s real name is coming “soon,” according to OpenRouter, and MiniMax M3.1-Flash-Preview is still the best guess.

Fable 5.5, Grok 4.8, Kimi K3.1, and GPT-6.1 Sol Ultrafast are all being rumored for October. Prediction markets had Haiku 5.5 at about 92 percent for October last week and it shipped, so the rumor mill is doing okay lately. I’m watching for Fable 5.5. DeepSeek V5 is still pure speculation. DeepSeek hasn’t said anything, and it’s been “next month” since August.

Finally

Haiku 5.5 is GLM’s price until your prompt passes 100K tokens. Muse Wonder is Opus’s price with a lower score. Step 5 Preview is a top-three model on OpenRouter only while it’s free. Argon is the best model in the world, at a price that will double, and you can’t buy it.

Meanwhile the bottom of my table didn’t budge. GLM-5.3-Flash, the MiMo models, and DeepSeek V4.1 Flash are still the floor, and now Anthropic has a real model sitting on that floor too, which I did not expect to type this year. I like it. Just read the long-context pricing table before you budget off the press release.

Next week I’ll check whether Haiku 5.5 has an Arena rating yet, and what happens to Step 5 Preview’s 12 trillion tokens once people have to pay $2.70 a million for them.