Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

Sonnet 5 is not worthy

Why Sonnet 5 is not worth of its name and Fable's return

Read in

News

Anthropic roller coaster

Fable is back


Anthropic has been able to negotiate with the US government and has gotten Fable unrestricted for everyone, not just US citizens. Hopefully this will also mean that GPT 5.6 will be released soon as well, since it is a (on paper) slightly weaker model than Fable is.

It does have even more aggressive guardrails now, which can be triggered with routine tasks like coding or debugging, which will cause it to fall back to Opus.

This is a bit worrying for future releases, as now we can expect the US government to step in and force a “review period” for all frontier models that get released from here on out. I would expect the release cadence of top models from OpenAI and Anthropic to slow because of this going forward.

It will also be interesting to see how they react to Chinese models that start reaching similar levels of capability (which we would expect to happen by the end of the year).

They are spying on you

It was discovered this week that in Claude Code, Anthropic has been secretly fingerprinting users to track them.

Specifically they check your timezone and proxy URL to see if you are using the model in China or in affiliation with Chinese labs.

Once they have identified locally if you are from China, they then add in special hidden characters to the prompt that are invisible to the user to fingerprint them.

This comes on the heels of Anthropic announcing that they caught the Alibaba Qwen team distilling over 28 million conversations from Claude.

Releases

Sonnet 5

Anthropic could not stop doing things this week, as they already released the first model in the Claude 5 series, being Sonnet 5.

bench

Why are the Opus 4.8 numbers explicitly “for reference”, isn’t that always the case for benchmarks?

This model is not worthy of the “5” moniker. With a new generation I expect a leap in performance, but with Sonnet 5 we get no new meaningful capability change from Sonnet 4.6. You can tell its not that different or exciting of a model because the US government hasn’t stopped them from releasing it at all.

It scores a bit better on benchmarks, but for real world use I would doubt if it even reaches Opus 4.6 levels of quality. This disappointing performance is going to cost you as well, as it costs more money to run than Opus and Fable due to its massive increase in token use.

yikes
Cost to run the Artificial Analysis benchmark suite

This model would be underwhelming if it was called Sonnet 4.7, and definitely does not deserve the Sonnet 5 name. It also fails to be meaningfully better than GLM 5.2, which is a model that is half the cost and fully open source.

This could be due to fears of regulation from the US government, so they are purposely releasing models that don’t have large capabilities increases, or they could just be hitting a wall with how far they can push these “smaller” models. If the latter is the case, then they would have known this for a while, which may be why they started training Fable, a much larger model, because they are running out of juice that they can squeeze from these smaller models.

We will find out in the coming month if this is true or not, as the Chinese open source models are roughly the same size (estimated) size as Sonnet, so we will see if they can continue to progress or if they will hit a wall as well and need to scale model size even further.

Quick Hits

Nano Banana 2 Lite

Despite the decline of the Gemini team at Google, the image and video generation teams are still very much on the frontier still.

This week they released Nano Banana 2 Lite, a model comparable to Nano Banana 2 that is half the price and 5 times faster.

It is still not at the same level at GPT Image 2, but it is 6 times cheaper at ~$33 per thousand images, which makes it far more useful at scale if you are willing to take the hit in quality.

image example

Prompt: a minimal flat animated scene of something neutral but interesting, a single character, pick an unusual animation style and palette — From fofr on Twitter

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

Highest resolution 3D images of the brain from Aleph on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS