| |
|
Best-AI.org
The Briefing
| NO. 06 |
28 OCT 2026 / 6 MIN |
|
|
● INSIDE A SWAP TABLE for the six retired Copilot models, and a 20-minute audit of your own bill.
|
|
The cold open
Ten cents a million. Now read the second line.
In October, Anthropic and OpenAI each put a small, fast model at $0.10 per million input tokens. Haiku 5.5 on the 7th. The Decisions API, on gpt-6-luna, on the 22nd. The headline price matches. The similarity stops there.
Haiku 5.5 charges five times more once a prompt passes 100k tokens. The Decisions API charges nothing for output, and its page says long-context multipliers apply without saying how big they are. Both pages tell you. Neither puts it in the headline.
Meanwhile GitHub retired six Copilot models on the 19th, which will not trend and which will page someone. Cheap is a pricing question. Retired is a calendar question. This issue answers both.
|
The signals
|
| 01 |
The cheap tier got a job description. Haiku 5.5 is $0.10 in and $0.50 out per million tokens, up to 100k. |
|
| 02 |
Classification became an endpoint. OpenAI's Decisions API is generally available: up to 400 questions per request, output not charged. |
|
| 03 |
Six Copilot models are gone. GitHub retired them on October 19. The swap table is below. |
|
|
IF YOU DO ONE THING →
Count how many of your calls cross 100k tokens, and price them both ways.
|
|
|
|
| 01 |
Anthropic
Haiku 5.5 is the cheap tier with a job description
Released October 7. Anthropic aims it at high-volume, cost-sensitive work: summaries, compactions, database queries and classification, live customer support, browser use, and as a subagent beside Opus 5.5 and Sonnet 5.5. Price is $0.10 in and $0.50 out per million tokens for prompts up to 100k, and $0.50 in and $2.50 out above that. For complex agentic coding Anthropic still points to Sonnet 5.5 and Opus 5.5.
THE TAKE
The pricing headline is the $0.10. The number that decides your bill sits after the words "over 100k": five times more, on input and on output. A long retrieval prompt can cross that line without anyone deciding it should. Route by prompt length, not only by task type.
SOURCE Anthropic, Claude Haiku 5.5
|
|
|
| 02 |
OpenAI
Classification became an endpoint
Generally available since October 22 on gpt-6-luna, the only supported model. One request takes 1 to 400 questions (200 in the public beta). Each question comes back as a predicate (a probability), a choice (one of your values, with probabilities and a confidence field) or a score (2 to 10 levels). Images and structured JSON can go into a question, and a shared prefix of at least 1,024 tokens can be cached.
|
|
|
Output price / 1M tokens
Inputs: Decisions API (gpt-6-luna) $0.10 uncached, Haiku 5.5 $0.10 up to 100k and $0.50 above. Decisions API output tokens are not charged, cache writes are, and long-context multipliers apply (size not published).
|
|
| |
THE TAKE
This is a classifier with a typed response and a discount, which is what most production AI calls already are. The confidence field is the useful part: it gives you something to route on. The trap is the usual one, a confident wrong answer is still wrong. Keep a person, or a stronger model, behind the low-confidence band, and measure that band before you trust it.
SOURCE OpenAI, Decisions guide · OpenAI, changelog
|
|
|
| 03 |
GitHub
Six Copilot models are gone. Here is the swap table
GitHub retired six models from Copilot on October 19. There is nothing to do to remove them. If a team picked one by name, use the table below, and note that admins may need to enable the replacement in the Copilot model policies before anyone can select it.
|
|
| Retired Oct 19 |
Suggested replacement |
| Gemini 3.7 Flash |
→ Gemini 3.8 Flash |
| GPT-5.5 |
→ GPT-6.1 Sol |
| GPT-5.4 |
→ GPT-6.1 Sol |
| GPT-5.4 mini |
→ GPT-6 Luna |
| GPT-5 mini |
→ GPT-6 Luna |
| Grok 4.5 |
→ Grok 4.7 |
|
|
| |
THE TAKE
A retirement notice is a calendar problem dressed as a news item. The model was never the dependency, the name was. If one of these six identifiers sits in a config file, an agent definition or an eval baseline in your repo, that is your ticket. Search for all six today.
SOURCE GitHub, Copilot deprecations
|
|
Hype check |
OUR OPINION, LOUDLY |
|
| ▲ |
Underrated
The "over 100k" line in Haiku 5.5 pricing
Everyone quoted $0.10. The sentence that changes your invoice is the one that starts with "over 100k".
|
| ▼ |
Overrated
"Decisions API" as a new capability
It is classification with a typed response and a discount. Useful, and not new. If you already run a bounded prompt, you already run this.
|
| ● |
Rated correctly
GitHub's retirement notice
Nobody needs to cheer for a notice like this. It is also the one item this month that can break a build on Monday.
|
|
Receipts |
GRADING OUR OWN HOMEWORK |
"Low-cost or mid-tier model. Cache stable context."
US, ROUTING MATRIX, ROW "ROUTINE SUPPORT AND SUMMARIES", ISSUE NO. 05
Three months on, Anthropic releases a low-cost model aimed at summaries and classification, and OpenAI releases an endpoint with a discount for a cached shared prefix. The direction held. What the matrix left out: one of those low-cost models costs five times more past 100k tokens, and it said nothing about prompt length. Verdict: right, and incomplete.
|
|
The number
5x
What Haiku 5.5 costs per token above 100k compared with up to 100k: $0.10 becomes $0.50 in, $0.50 becomes $2.50 out. The cheap tier is cheap for short prompts.
SOURCE Anthropic, Claude Haiku 5.5
|
|
One thing to try
The 100k audit, 20 minutes
| 1 |
Export one week of production calls with their input token counts, from your provider dashboard or your gateway logs. |
| 2 |
Count the calls over 100k tokens and the share of your bill they carry. |
| 3 |
Take the 20 most routine calls under 100k and run them on Haiku 5.5 beside your current default. |
| 4 |
Set the pass mark before you read the answers, then score them yourself, not with another model. |
| 5 |
Move model names into one config file with an owner. The next retirement notice becomes a one-line change. |
|
|
If you read one thing
The decision types in OpenAI's guide
Predicate, choice and score each return different fields, and the confidence field is where your routing rule lives. Read that section before you design a schema around it.
READ THE GUIDE →
|
|
|
|
From the directory |
NOBODY PAID FOR THIS |
|
Three listed tools, each tied to a story above. |
|
Code assistance
GitHub Copilot
Open the model picker and see which of your team's favourite names survived the October 19 cut.
|
VIEW →
|
|
|
Conversational AI
Claude
Haiku 5.5 sits at the low end of this family. It is the obvious candidate for your 20-call test.
|
VIEW →
|
|
|
Data analytics
Helicone
Logs and routes LLM calls, which is where you would look for the ones that cross 100k tokens.
|
VIEW →
|
|
|
|
Try this · our app
Find AI with AI
Search the directory from your phone and save the tools you want to test.
|
|
|
Built something worth finding?
Submit it for free review. Add our badge with a normal followed link and we gift you Fast-Track: a decision within 48 hours after we confirm the link.
|
|
|
That is October. Go search your repo for the six names.
Think we called one wrong? Excellent. Reply and tell us which, with a source. We will grade ourselves in public.
THE BRIEFING NO. 06 / BEST-AI.ORG
You subscribed to Best-AI.org updates. Unsubscribe anytime.
|