← all posts
// economics · pricing

The ten-cent tier: Haiku 5.5, Muse Ember, Hy4 flash, and why cheap models now come with price cliffs

The most useful price movement of the last 45 days happened at the bottom of the rate cards. On October 7, Anthropic priced Claude Haiku 5.5 at $0.10 input and $0.50 output. On November 9, Meta priced Muse Ember at the same numbers. On November 16, Tencent listed Hy4 flash at $0.07 and $0.281. For the high-volume, low-difficulty part of a pipeline, that is a different cost floor: Haiku 4.5 listed at $1 and $5, ten times the new rate.

The cheap and middle tiers after November 19

ModelDateInputOutputThe catch
Hy4 flashNov 160.070.281Tencent API, 1M context
Claude Haiku 5.5Oct 70.100.500.50 / 2.50 above 100k tokens
Muse EmberNov 90.100.500.30 / 1.50 above 200K tokens
Mistral Large 4Oct 60.702.10Open weights also available
Hy4Nov 40.832.50Rate seen on OpenRouter
GPT-6.1 TerraOct 261.005.00Tool use and computer interaction
Qwen4 MaxNov 161.504.50Benchmarks self-reported
Gemini 4 HydrogenNov 121.507.50Flat across 1M, has a free tier

Prices are dollars per million tokens. The first three rows are the ten-cent tier. Everything below them is a middle tier that keeps getting more crowded, and Hydrogen at the bottom of the table costs fifteen times what Haiku 5.5 does on input.

The cliffs

Haiku 5.5 charges $0.10 and $0.50 for prompts up to 100k tokens and $0.50 and $2.50 above that. Muse Ember charges $0.10 and $0.50 up to 200K input tokens and $0.30 and $1.50 above. Anthropic describes the average saving against Haiku 4.5 as about 75%, made up of 90% below the threshold and 50% above it.

Those thresholds change how I would design a pipeline. A summarisation job that feeds 120k tokens into Haiku 5.5 pays five times the headline rate on every token of the request. The same job split into two 60k calls plus a merge step pays the ten-cent rate three times and usually comes out well ahead. I went through the tokenizer side of this in the Haiku 5.5 migration piece: the new tokenizer uses slightly more tokens per task, so a prompt that measured 95k on the old model can cross the line on the new one.

A cheap model with a context threshold is two models with two prices, and your prompt length picks which one you bought.

Below ten cents: models that only decide

For routing, classification, policy checks and grading, the cheapest option is no longer a small general model. It is a decision model, which reads the input once and returns a probability for each option in a closed list without generating text. OpenAI's Decisions API on GPT-6 Luna is $0.10 per million input tokens. Microsoft-Decision-1, a post-trained Qwen3.5-9B, is $0.042 with a 32K context. Google's Gemini Decision 1 arrived on November 11 at $0.10 with a 256K input window and up to 1,000 questions per request.

At one billion input tokens a month of pure classification, that is about $42 on Microsoft's model, $70 on Hy4 flash and $100 on Haiku 5.5 or the OpenAI and Google decision endpoints, before output. The differences are small in absolute terms. The reason to prefer a decision model is latency and a calibrated probability you can threshold, which I covered in the decision models overview.

The middle tier, and one claim I would not budget on

Microsoft announced MAI Code 2 and MAI Code 2 Flash at Ignite on November 17. Its post says MAI Code 2 reaches 98% of GPT-6.1 Sol's performance at 40% of the cost and 96% of Claude Opus 5.5's at 5% of the cost, averaged over four coding benchmarks at max reasoning effort. It also says both models are deliberately not tuned for fundamental science or very long-horizon agentic work. What the post does not contain is a price in dollars. The models are in GitHub Copilot at a 50% promotional rate, and until there is a rate card I cannot turn a percentage of someone else's cost into a number for a spreadsheet.

Qwen4 Max at $1.50 and $4.50 with a 1M context is the most aggressive output price in the middle tier, less than a third of Neon's. Its benchmarks are Alibaba's own. GPT-6.1 Terra at $1 and $5 is OpenAI's entry for tool use and computer interaction. Mistral Large 4 at $0.70 and $2.10 is the cheapest of the group and the only one you can also download.

How I would use this

My rule is to push work down until quality breaks and stop one step above that. Closed-set decisions go to a decision model. Extraction, summaries, compaction and subagent chores go to the ten-cent tier with prompts kept under the threshold on purpose. The middle tier takes interactive coding. The $8 models take whatever failed. The live numbers are on the price board, which reads list rates hourly, so I check there before I trust a table like the one above, including my own.

The limitation is that I am comparing list prices, and list price per token is a poor predictor of cost per finished task once reasoning effort is adjustable. Haiku 5.5 is the first Haiku with an effort setting, and a cheap model at maximum effort can out-spend a mid-tier model at low effort. For the Chinese-hosted endpoints, also check where the service runs and what your contracts allow before a price difference of three cents makes the decision for you.

#pricing#economics#small-models#routing