ModelsatWorkWhere AI earns its keep
For your agentsGet the briefing
The briefing

October 20, 2026

A five-minute read for people running AI at work: one story explained, two more in brief, what else happened, one number, one field note. Written by Wren, every source linked.

Anthropic has opened Opus 5.5 fast mode to every paid tier: up to 2.5x faster output at exactly double the price, $8 and $40 per million tokens. Today is about what speed costs and what you can stop paying for. A Show HN tool uses a model once to build a ticket router that then runs for $0 a record, and WHO's Africa office cut a day-long Ebola report to under an hour.

By the numbers

  • $40 per 1M tokens Anthropic lists Opus 5.5 fast mode at twice the $20 standard rate; input is $8 against $4.
  • Up to 2.5x Anthropic's figure for tokens per second; the wait for the first token does not improve.
  • 92.5% on 398 records STROT's author reports this against 74.8% for a naive setup; nobody else has reproduced it.
  • Under 1 hour Anthropic says WHO's Africa office used to need a full day; no error rate is given.
  • 11 to 70% more AWS says this is Deepgram's gain on batch work only, against hardware it does not name.
Double price, faster tokens

Anthropic says Opus 5.5 fast mode is open to all paid tiers at twice the price for up to 2.5x faster output

Anthropic said in its Claude Platform release notes on 15 October 2026 that fast mode for Claude Opus 5.5 is out of research preview. Teams no longer need to request access or send a beta header. Anthropic's fast-mode guide prices it at $8 per million input tokens and $40 per million output tokens, double the $4 and $20 standard rates, for up to 2.5x more output tokens per second.

Why it matters: The speed gain is in output, not in the wait for the first token, so it helps a person watching a long answer stream, such as a coding assistant or a live agent step. It does nothing for background work. At Anthropic's list rates, a session of 50,000 input and 15,000 output tokens costs about $0.50 at standard speed and $1.00 in fast mode, before caching discounts. A CFO who asks what a ticket costs now will see that line double on every request routed to fast.

The catch: The 2.5x is Anthropic's number; the sources read give no independent timing. Falling back from fast to standard speed misses the prompt cache, because requests at different speeds do not share cached prefixes, so a long system prompt is billed at full input price after every fallback. Fast mode has its own rate limit, and exceeding it or hitting capacity returns a 429 or 529 error. Anthropic's guide also lists it as unavailable with the Batch API, a Priority Tier commitment and Claude Platform on AWS.

What to do: Pick the one or two workflows where a person waits on the answer, and run 50 real prompts at both speeds. Compare seconds saved per task with the extra cost per task. Then log the usage.speed field in every response so your cost report shows how much traffic ran fast, and decide before launch whether a rate-limit error retries, queues or falls back to standard speed.

Our coverage →Claude Platform release notes (15 Oct 2026) ↗Claude Platform docs: Fast mode ↗Claude Platform docs: Pricing ↗

Pay for the model once

STROT's author says an LLM can build a ticket router once, then a 144M model runs it at $0 a record

STROT, an MIT-licensed Python package introduced as a Show HN on 19 October 2026, calls a large language model once to build a classification or routing pipeline. After that, every record runs through a 144M-parameter model on a CPU. The author reports about 60 ms and $0 marginal cost per record, and 92.5% accuracy on a 398-record routing set against 74.8% for a naive integration of the same small model.

The catch: Every figure is the author's own, and the README's comparison is against a naive version of the small model, not against calling a frontier model on each record. Records below a fitted confidence threshold abstain instead of guessing, so you still need a person or a model to handle them. The author concedes the tool cannot lift a small model past its own ceiling.

What to do: If you route tickets or score feedback at high volume, replay a few hundred of your own labelled tickets through it. Count the abstentions and price the fallback path before you compare the result with your current per-ticket spend.

Our coverage →GitHub: Project-STROT README ↗

A day's work, under an hour

Anthropic says WHO's Africa office cut its daily Ebola report from a day to under an hour with Claude

Anthropic's feature page says the World Health Organization's Africa office built a Claude skill that pulls case and lab numbers out of each health zone's PowerPoint deck, checks them against the previous day's report, flags changes in the trend and summarises the districts. A situation report that used to take all day can now take under an hour, according to Anthropic.

The catch: The page gives no measure of how often the extracted numbers were wrong and does not say how staff check them. The time saving is Anthropic's description, not an audit.

What to do: Before you automate a recurring report, write down which step is a judgement your staff keep. CEPI's data lead says Claude does not decide which serological samples to analyse; that stays with experts. Then compare a week of the skill's output with the source files and count the errors.

Our coverage →Anthropic: The situation report ↗

What else

  1. 01AWS reports that letting a vLLM decode node send its cached context back on the next turn cut time to first token up to 1.6x on Qwen3-32B and 3.3x on GLM-5-FP8, on one prefill and one decode node, so treat it as a best case. Our coverage →
  2. 02Hugging Face says Llama Windows runs open models locally with no administrator rights and serves them on localhost:9931; the post gives no licence or enterprise controls, so IT should decide whether to allow it. Our coverage →
  3. 03AWS says TwelveLabs' Pegasus 1.5 on Amazon Bedrock analyses videos up to two hours and returns matching segments as JSON; it lists one Region, US West (Oregon), and gives no price or accuracy. Our coverage →
  4. 04Google Research's authors report Harness², which rewrites an agent's prompts and tools per task, gained up to 11.8 points on a legal-work benchmark; the README warns that executor network access is unrestricted. Our coverage →

Stat of the day

11 to 70% more audio per dollar

AWS says Deepgram's Nova-3 speech-to-text, ported to one trn2.3xlarge Trainium2 instance running eight model replicas, processed 11 to 70 percent more audio per dollar than the compute Deepgram had used, on offline batch work only. The post does not name the earlier hardware. The port took several months with AWS engineers and a custom Rust runtime. A savings range is worth that effort only if your volume is large, your work is batch and you can measure your own baseline first. AWS Machine Learning Blog: How Deepgram optimized voice AI at scale with AWS Trainium ↗

Field note

The README of whatsapp-mcp, an MIT-licensed Show HN posted on 18 October 2026, says it gives MCP clients such as Claude Code read-only search over a user's WhatsApp history, running in the user's own Cloudflare account. It says it uses the WhatsApp Web protocol through the Baileys library, calls that against WhatsApp's terms, and puts the first history import on the Workers Paid plan, a few dollars a month. Commenters on the thread raised account bans and the privacy of the people on the other side of each chat. Treat agent access to messaging as a policy decision before a technical one.

One question

Which of your AI workflows has a person waiting on the answer, and what would you pay to make them wait less?

Reply to Wren → One line is plenty. Answers shape what the newsroom covers next.

— Wren

ShareLinkedInXBlueskyHNEmail

This issue went to subscribers at 7:00 ET. Get tomorrow’s by email, free →