Claude Haiku 5.5 Price Drop and Google's SynthID Detector

Anthropic cut the price of its small model by ten times, and Google handed the general public a tool that checks if an image was made by AI. Claude Haiku 5.5 now runs at 10 cents per million input tokens. Google's SynthID Detector came off the waitlist and is open to everyone.

One of these changes your AI bill. The other changes what you can trust in your inbox. Both are worth ten minutes of your week.

Claude Haiku 5.5 pricing: 10 cents per million tokens

Anthropic released Claude Haiku 5.5, the small and fast member of the Claude family. The old Haiku 4.5 cost $1 per million input tokens and $5 per million output. The new one is 10 cents in and 50 cents out.

That is a 10x cut in one generation. Prices step up for very long sessions, once you push past 100,000 tokens of context. For normal short tasks, you are paying pennies.

It is already live on the API, in Claude Code, and through Amazon Bedrock, Google Vertex, Microsoft and OpenRouter. So you do not have to change where you buy it. You change one model name in a config file.

Anthropic also says Haiku 5.5 beats the competing small model on a benchmark that measures real economic work, scoring 1620 against 1437. Keep in mind who published that number. Every lab reports the benchmark where it wins, which is why I treat these charts like a restaurant's own review of itself.

The benchmark barely matters here anyway. The price does. A model this cheap turns a whole category of "not worth automating" tasks into "sure, why not".

Think about the boring work inside a workflow. Summarising a long email thread. Pulling five fields out of a PDF.

Tagging a support ticket. Deciding which of three branches to take next. None of that needs a frontier brain. It needs something fast that does not get bored.

If you are running those small steps on a big model today, you are overpaying by a lot. The practical move: look at your AI spend, find the dumbest tasks, move them to the cheapest model that still passes your tests. Keep the expensive model for the parts where being wrong actually costs you money.

This is exactly how we structure AI agents for clients. A capable model runs the plan, cheap helpers do the fetching and sorting.

The cheap helpers used to be genuinely dim. They are not anymore.

One warning. Cheap models invite waste. When each call costs almost nothing, people stop thinking and start calling the model forty times for something a database query would answer.

Your bill stays flat and your system gets slower and harder to debug. Cheaper tokens are not a reason to be sloppy with them.

Also worth saying out loud: Haiku 5.5 would have been close to the best model on earth not long ago. Now it is the thing you use to clean up formatting.

That is the actual story of this year. The floor keeps rising and nobody throws a party for the floor.

Google's SynthID Detector is now open to everyone

Google opened SynthID Detector to the public in English. Until now it was a waitlist tool, mostly for journalists and researchers. Now anyone can upload an image, a video or an audio file and ask whether AI made it.

It works by looking for an invisible watermark baked into the file when it was generated. Google says it covers media from its own image, video and audio models, plus content from OpenAI, Nvidia and Kakao models. The same underlying generators sit inside a lot of other apps, so the real coverage is wider than those four names.

Here is the part that matters, and I want to be blunt about it. The tool can tell you something is AI. It cannot tell you something is human.

If the watermark is there, you have an answer. If it is missing, you have nothing.

The file might come from a model that does not watermark. Someone might have screenshotted it, cropped it or run it through an editor. "No watermark found" is not a certificate of authenticity, and people are absolutely going to treat it like one.

Still useful. I expect this to show up fast in places where fake images cost real money. Hiring, where the portfolio is suspiciously clean. Insurance claims, marketplace listings, and dating apps, where the watch in the photo somehow has no reflection.

For a small business, there are two sides to this. First, the incoming side.

If you receive documents, product photos or voice messages that lead to a payment, a check like this is now free. Add it to your process for anything above a certain value. It takes seconds.

Second, the outgoing side. Your own marketing images probably carry these watermarks now. Light AI editing is normal and fine.

A tidied background, a cleaned-up headshot, nobody cares. But if your website claims a real photo of your real workshop and the detector says otherwise, that is a bad afternoon. Know what you published.

I also think the detection arms race is not winnable in the long run. Watermarks survive until someone has a reason to strip them.

The people most motivated to lie to you are the ones most motivated to figure that out. A tool like this catches the lazy fakes, which honestly is most of them.

What I'd do this week

Two small jobs. Open your AI billing dashboard and find the tasks that do not need a smart model. Move them to Haiku 5.5 and compare the output on twenty real examples before you commit.

Then bookmark SynthID Detector and run three images through it. One of yours, one from a supplier, one you already suspect. You will learn more in five minutes than from any article about deepfakes, including this one.

My prediction: within a year, "was this checked?" becomes a normal question in procurement and hiring, and some vendor will sell it back to you as an enterprise dashboard for €400 a month. The tool is free today. Use it while it still is.