The Signal — October 8, 2026

Anthropic's new small model now costs the same per token as OpenAI's cheapest one for short prompts, though a new tokenizer means the rate card overstates the savings. Common Sense Media says ChatGPT's teen protections fall short of what OpenAI promised parents, and US agencies and AI labs are contributing funding and existing biological data to Biohub's effort to build "virtual cell" models.

Anthropic's Claude Haiku 5.5 costs a tenth as much per token as its predecessor for short prompts, with a catch in the tokenizer

Anthropic released Claude Haiku 5.5 on October 7, a small model aimed at high-volume jobs like summaries, classification, customer support, browser use and acting as a cheap subagent for bigger models. It's available on Anthropic's own platform and through AWS, Google Cloud and Microsoft Azure. For prompts up to 100,000 tokens it costs $0.10 per million input tokens and $0.50 per million output tokens, a tenth of what Haiku 4.5 charged. Longer prompts cost five times as much, $0.50 and $2.50. Anthropic says about 90% of Haiku 4.5 requests fell under the 100,000-token line.

Haiku 5.5 also uses a new tokenizer that turns the same text into more tokens, so the bill falls by less than the rate card suggests. Simon Willison found that one long prompt came out at about 1.25 times as many tokens as it did on Haiku 4.5. Anthropic's own estimate, which accounts for the tokenizer, is that the model costs about 75% less to run on average. Willison points out that under 100,000 tokens Haiku 5.5 now matches the price of OpenAI's GPT-6 Luna exactly, while above that line Luna is much cheaper, because its price only rises past 272,000 tokens and then only to $0.20 and $0.75. Haiku 5.5 is also the first Haiku with an adjustable effort setting, though Willison notes reasoning can't be switched off and defaults to medium.

Anthropic's benchmark table shows a large jump over the old model. On the offline subset of OSWorld 2.1, a test of operating a computer through multi-step tasks, Haiku 5.5 scored 72.4% against 15.7% for Haiku 4.5 and 48.9% for GPT-6 Luna. On Terminal-Bench 4.0, a command-line agent test, it scored 39.2% where Haiku 4.5 scored zero, though Sonnet 5.5 reached 70.6%. Anthropic itself says its bigger models remain the better choice for complex coding. These are Anthropic's numbers and haven't been reproduced by anyone else yet. Alongside the launch, Anthropic halved Sonnet 5.5's cache-read price to $0.10 per million tokens, which it says makes most agentic work about 20% cheaper. It's also adding monthly API credits for subscribers: $100 on Max 5x, $200 on Max 20x and up to $500 pooled for Team plans. Willison notes the credits don't roll over.

Sources: Anthropic · Simon Willison · The Decoder

Common Sense Media rated ChatGPT for Teens an "Unacceptable Risk," and OpenAI disputes the testing

Common Sense Media's Youth AI Safety Institute tested ChatGPT before and after OpenAI launched ChatGPT for Teens on August 18. Testers ran more than 4,000 prompts on accounts registered to 13- to 17-year-olds, and experts, including child psychiatrists and a pediatrician, reviewed the responses. On October 7 the Institute rated the product an "Unacceptable Risk" and said OpenAI should limit ChatGPT to adults until it fixes the gaps and proves the fixes through independent testing. According to the Institute and Axios, some protections held up, including refusing sexual roleplay and generally declining to give instructions for self-harm.

The failures the Institute reports are mostly about the features OpenAI marketed to parents. On newly created accounts linked to a parent, testers spent up to an hour discussing suicidal thoughts, self-harm or disordered eating and the parent received no alert; they could only trigger alerts on much older accounts with weeks of sensitive conversation history. ChatGPT missed more than one in four crisis referrals the experts judged necessary. A "Show me the answer" option let users get finished homework out of study mode, and parent-set study hours could be turned off by deleting a prefix from the prompt. Adult-registered test accounts that told ChatGPT they were 13 never got moved into the teen experience. The Institute says it is funded partly by industry, including the OpenAI Foundation, and that it keeps editorial control of its results.

OpenAI disputes much of this. A spokesperson told Axios and TechCrunch the testing doesn't reflect how the safeguards work in practice, and said much of the alert testing may have happened before parent and teen accounts finished linking, which OpenAI says can take several hours. The company says its age prediction uses several signals and can take up to two weeks, and that letting teens leave study mode during study hours is a deliberate design choice. It also says its own larger-scale data shows hotline resources being shown more often to under-18 users. This is one group's test on test accounts, so it can't say how many real teens hit these gaps, and some of the disagreement is about what the features were meant to do in the first place.

Sources: Common Sense Media · Axios · TechCrunch

Biohub, US agencies and three AI companies combine funding and existing data in a $1.8 billion effort to model cells

Biohub, the research nonprofit backed by Mark Zuckerberg and Priscilla Chan, announced on October 7 that the US Department of Energy, the National Institutes of Health, Google DeepMind, Isomorphic Labs and Meta are joining its Virtual Biology Initiative. The goal is an open collection of measurements of how cells respond to interventions, across far more cell types and conditions than have been studied so far, large enough to train AI models that can predict those responses before anyone runs the experiment.

The $1.8 billion figure adds up several different kinds of contribution. The DOE says it will spend more than $500 million over five years on lab measurement, modeling and computing at the national labs. The NIH will contribute existing datasets and repositories built with more than $500 million in earlier federal funding, which Biohub will standardize for AI training. Google DeepMind, Isomorphic Labs and Meta are putting in $300 million between them. Biohub's own $500 million, announced when the initiative launched in April, pays for new measurement tools such as cryo-electron tomography and microscopy that can image millions to billions of cells in living tissue. With existing data and an earlier commitment included, the total is well short of $1.8 billion in new cash.

Reuters, as relayed by The Decoder, adds two details that are not in the Biohub release. Companies that fund data collection get a year of exclusive access to what they paid for before it goes public, while the government-funded data carries no such restriction. The first dataset should be ready in about a year. The release describes predictive cell models as a goal of the work, not something the partners have built.

Sources: Biohub · The Decoder

On the Editor's Desk

OpenAI also started bringing GPT-6, with a new "Intelligent UI" that builds charts, buttons and small tools into answers, to free ChatGPT users. The models first reached paying customers last month, and so far the only examples of the new interface are OpenAI's, so we'd rather wait until people have actually used it. We held Representative Lori Trahan's AI liability proposal because it is a discussion draft she has put out for public comment, not a bill introduced in Congress. NVIDIA's results for its Nemotron models on 2026 math and programming olympiad problems were first published in September and weren't official medal results.