Claude's New 10-Cent Model Has a Tokenizer Catch

Free ChatGPT got GPT-6. Claude's small model dropped to 10 cents per million input tokens. OpenAI shipped an API that never bills output.

News · No. 34 · Tue Oct 13

Cost Week, Not Capability Week

Free ChatGPT got GPT-6. Claude's small model dropped to 10 cents per million input tokens. OpenAI shipped an API that never bills output.

Nobody shipped a model last week that changes what's possible. Three companies changed what it costs. That's the part that hits your Monday.

In This Email

» Free ChatGPT now runs GPT-6, and its answers can be buttons, charts and mini-tools

» Haiku 5.5 at $0.10 in, and why your real saving is closer to 75% than 90%

» OpenAI's input-only Decisions API, and quick hits on Gemini's agent, textGrain and Nano Banana 2.1

The Big Thing: The Free Tier Just Moved

Free ChatGPT Now Runs GPT-6

In No. 30 we told you GPT-6 Sol skipped Chat. That gap is closed. OpenAI put GPT-6 into the regular Chat tab on October 7 for paid plans (GPT-6 Sol) and on October 8 for Free and Go (GPT-6 Luna). Notebookcheck has the rollout details.

What Intelligent UI means: answers aren't only text anymore. GPT-6 can build a reply out of diagrams, buttons, side-by-side comparisons and small working tools, like a calculator or a bill splitter, right inside the chat.

Two catches: older Mac and Windows desktop apps don't support it (use the web or the latest phone app), and your plan's normal limits still apply.

Why you care: a lot of small "AI tools" are a form, a prompt and a chart. Free ChatGPT now builds that on demand. If your product is a thin wrapper, your competitor is now free.

Try This in 15 Minutes (Free)

1. Open ChatGPT on the web in a private window, on a free account.

2. Ask it to build the exact tool your product, lead magnet or client deliverable offers. For example: "Build me a pricing calculator for a three-tier SaaS with an annual discount."

3. If it gets most of the way there, write down what your version does that this can't: your data, your workflow, your distribution. That's your edge.

My take: this is a pricing story dressed up as a product launch. What people paid for in a chat window last year is now the free tier.

Worth Your Time

Claude Haiku 5.5: The 10-Cent Model Has a Tokenizer Catch

Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens. Haiku 4.5 was $1 and $5. Looks like a 90% cut. It isn't.

The catch, defined: a tokenizer chops your text into the tokens you pay for. Haiku 5.5's newer one uses slightly more tokens per task, and prompts over 100,000 tokens cost $0.50 and $2.50. Anthropic's own footnote nets it out at about 75% cheaper on average. Budget on 75, not 90.

Same day, Anthropic halved Sonnet 5.5 cache reads to $0.10 per million (about 20% cheaper on most agentic work, it says) and added a monthly API credit: $100 on Max 5x, $200 on Max 20x, up to $500 pooled on Team.

Monday use: pull 50 real prompts from a summarize, classify or sub-agent job. Run them on Haiku 4.5 and claude-haiku-5-5, log tokens, multiply by the new prices. That's your real saving, and on Max or Team the credit pays for the test.

OpenAI's Decisions API Bills Input Only

The Decisions API answers typed questions instead of writing text: the probability that something is true, a pick from your list, or a score against your rubric. It runs on gpt-6-luna at $0.10 per million input tokens, with no output, cache-read or cache-write charges. OpenAI says it's about 10x faster than the Responses API. Public beta.

Monday use: any step where a chat model answers "which team gets this?" and you parse the reply is a Decisions call now. Cheaper, faster, and it can't ramble.

Local option: Liquid's open-weight d1-3B does the same job on your own GPU, about 8 ms a decision on an RTX 4090 per Liquid. Its license sets a $10 million annual revenue threshold, so check it first.

Quick Hits

Gemini's Work Agent Is Private Preview Only. Google's new Gemini agent plans work, uses tools and returns finished work in Docs and Gmail, even as a coworker agent with its own email address. The Verge reports it's enterprise private preview for now, with no price.

OpenAI's TextGrain Watermark Is Opt-In for API Builders. OpenAI is adding an invisible watermark to ChatGPT and Codex text for EU users. In the API it's off by default, opt-in for select models. EU-facing builders should decide now. Don't build "proof of human writing" on it; OpenAI says detection isn't guaranteed. The Verge.

Nano Banana 2 Has a New Model ID. Google deprecated gemini-3.1-flash-image with no shutdown date yet. Swap to gemini-nano-banana-2.1 now, while it's a choice and not an outage. Google's release notes.

Skip It

Reflection's 501B Beam, for Now

An open-weight model you can't download. Access is a waitlist and the weights and technical report are promised later this month, so every benchmark is self-reported. Spend the hour on the Haiku test instead.

Step-by-step lessons for tools like these live in the free AI Academy.

Join the free AI Academy →

Free, no card. The lessons behind Friday's tools live here.

That's the Debrief.

See you Friday for the tools.

From Drew and The AI Debrief team

P.S. Run the Haiku 4.5 vs 5.5 test and tell us your real number. Closer to 75% or 90%? Hit reply, we read every one.

P.P.S. New here, or skipped one? Every past edition is in the archive.