Tool of the Week: Claude Fable 5.1 (and the test I ran before adopting it)
Anthropic released Claude Fable 5.1 on September 1. Same list price as Fable 5 ($10 per million input tokens, $50 per million output), better benchmark scores, and one real pricing change: cache reads dropped 75%, to $0.25 per million tokens. Anthropic's own estimate is about 25% lower cost on a typical workload and up to 45% on heavy agent work, where most of the tokens are context the model has already seen.
I run a scoring pipeline in production that calls Claude four times per symbol across 17 symbols. So on September 3 I ran the two models against each other on identical inputs.
Result: the scores were the same and the bill went up. Fable 5 scored 39 / 57 / 79 (min / median / max). Fable 5.1 scored 37 / 56 / 75. Same signals out the other end, zero refusals on either. Cost per full pass: $1.98 on Fable 5, $2.29 on 5.1. About 16% more. I stayed on Fable 5.
That is not a knock on the model. It is the point of the exercise. The 25% saving is real, but it is a saving on cache reads, and my pipeline sends fresh market data every call, so it barely has any. A coding agent that re-reads the same repo for an hour will see the discount. A batch job that sends new inputs each time will not. Anthropic said as much in the fine print. Most people will read the headline.
The test also found a bug that had nothing to do with the model. 5.1's longest calls took over three minutes and died on a client that was not streaming responses. Switching to streaming brought the same calls down to about 65 seconds. Had I just flipped the model name in config and walked away, I'd have discovered that in production at 9:34 on a trading morning.
How to run this yourself (about 20 minutes if your pipeline is scripted):
Freeze the inputs. Same prompts, same data, same day. If the inputs drift, you are measuring the weather, not the model.
Score the output on the metric you actually care about, not a benchmark. For me that is the signal a symbol gets. For you it might be "did the invoice extraction match the PDF" or "did the draft need edits."
Read the per-call cost off your own usage page, not the pricing page. Compare full-run cost, not per-token rates.
Watch the slowest call. New models often think longer. Timeouts that were fine last month may not be fine now.
Write the decision down with a date. Mine says: stay on Fable 5, no quality gain, higher cost, and do not reopen the question without a reason beyond "newer."
Who this is for: anyone paying for Claude by the token with a job that runs unattended. If your usage is interactive (Claude Code, Cowork, the app), 5.1 is the default now and the effort setting matters more than the version: Anthropic says Low or Medium effort on 5.1 matches Fable 5 at lower cost, and High is the Claude Code default. Check which one you are on before you decide it got expensive.
Quick Hits
Google shipped its third Flash model in six weeks. Gemini 3.8 Flash landed September 2, three weeks after 3.7 Flash, at the same introductory price ($0.75 per million input, $3.75 output, standard rates from January 1, 2027). Google's own description is that it "works harder": more reasoning steps, more tool calls, and sometimes more tokens per task at higher effort. They are explicitly keeping 3.7 Flash around for cost-sensitive work. A locked-down cybersecurity variant, 3.8 Flash Cyber, is restricted to vetted defenders through a new Fairwind Program. Why it matters: same lesson as the Tool of the Week from the other side. "Same price" and "works harder" together mean your per-task cost can rise on an upgrade even when the rate card does not move. If you are on 3.7 Flash for a high-volume job, measure a day of 3.8 before you switch the whole thing over.
Every new Claude output now carries an invisible watermark. As part of the EU AI Act's Code of Practice on AI-generated content, which Anthropic signed in July, text from models released after August 2, 2026 (so Fable 5.1 onward) includes a statistical watermark. It is not visible, does not change the output, and carries no user information. A detection API is in private preview for regulators, media, fact-checkers, researchers, and companies with their own compliance obligation. Why it matters: if you ship AI-drafted text to clients or publish it, assume it can eventually be identified as such by anyone with access to the detector. That argues for what was already the right practice: edit it, own it, and do not pass it off as something it is not.
Anthropic closed the "edit the model's past" door for new API accounts. Also effective September 1: API accounts created from launch day onward can no longer modify Claude's earlier turns in a conversation while keeping the transcript of its earlier reasoning. Anthropic frames it as an anti-distillation measure (stopping competitors from harvesting Claude's thinking at scale), and says it will extend to all accounts with future releases. Why it matters: if you build anything that rewrites or trims conversation history mid-run, treat history as append-only from here. It is a small code change now and a confusing 400 error later if you ignore it.
Prompt of the Week: The Model Upgrade Decision
Use this before switching any production workflow to a new model. It forces the comparison to be about your job, not the launch post.
I run an AI workflow in production and a new model version just came out.
I want to decide whether to switch, and I do not want to decide based on
the announcement.
Here is the workflow:
[describe it, e.g. "four Claude calls per record across 17 records, once
a day, fresh data each time, output is a 0-100 score per record"]
Current model: [name]. Candidate model: [name].
Pricing for each (input / output / cache read, per million tokens):
[paste from the vendor pricing page]
Tell me:
1. Based on how my workflow uses context, will I actually benefit from the
cache-read pricing, or is the headline discount mostly irrelevant to me?
2. What is the ONE output metric I should compare, and how many runs do I
need before a difference is more than noise?
3. What operational things could break on a model swap that have nothing to
do with output quality (timeouts, streaming, rate limits, effort defaults,
deprecated parameters)? List them as a pre-flight checklist.
4. Draft the decision note I should write afterward: date, models, inputs
frozen (yes/no), metric result, cost result, decision, and the specific
condition that would justify reopening it.
Be blunt. If the honest answer is "do not bother testing, the change does
not touch your workload," say that.Run it once per workflow you pay for. The decision note in step 4 is the part people skip and the part that saves you from re-running the same debate next month.
One last thing
Like what you're reading? Forward it to one person who'd get something out of it.
And if you want a second set of eyes on where AI could actually save you time, book a free 15-minute audit.