Claude Opus 5.5 and GPT-6 Sol land, Claude finds a new enzyme

What shipped

Anthropic

Claude Opus 5.5

The first Claude 5.5 model matches Fable 5.1 on most tasks at $4 in and $20 out per million tokens, with typical workloads about 40% cheaper than Opus 5.

OpenAI

GPT-6 Sol and GPT-6 Luna

Two cheaper tiers below Astra, with a 1.05M window, 128K output and six reasoning-effort levels, for coding and high-volume work.

xAI

Grok 4.7

xAI's top coding and knowledge model, served at the same price and speed as Grok 4.6 and working longer on hard tasks.

Xiaomi

MiMo-V2.6-Pro

A 1.02-trillion-parameter open model with 42B active parameters, MIT license and a 1M context, trained for long-horizon agent work.

Anthropic

Claude discovers a novel enzyme system with CRISPR-like repeats

About 950 Claude agents searched a DNA database for 21 hours and flagged an uncharacterized array-associated reverse transcriptase system that lab work then confirmed.

Liquid AI

d1 decision model

A model that answers typed yes/no, choice and score questions in one pass and returns calibrated probabilities without writing any text.

Google DeepMind

Gemini 3.8 Flash TTS and Flash-Lite TTS

Google's text-to-speech models move to the 3.8 generation, in two sizes.

Perceptron

Perceptron Mk1.5

An embodied vision-language model with native audio input and video object tracking, about two to five times faster than Mk1.

arXiv

Training Object Permanence in World Models

Tests whether video-generation world models hold onto objects that go out of view and trains them to do so.

Three labs updated their lineups in the same two days, and an Anthropic science result gave me more to think about than any of them.

Claude Opus 5.5

Anthropic released Opus 5.5 on 22 September as the first model in the Claude 5.5 family. The company says it matches Fable 5.1 on most work and calls it the new leading model. Pricing is $4 per million input tokens and $20 per million output, with cache reads at $0.20, all below Opus 5. Anthropic estimates typical workloads cost about 40% less than on Opus 5 and that output is more than 30% faster. It also reports better resistance to prompt injection.

For someone running many agents, this one is about routing. A model at Fable-class capability for less money changes which jobs deserve the top tier.

GPT-6 Sol, Luna and Grok 4.7

OpenAI announced GPT-6 Sol and GPT-6 Luna on 22 September, the tiers below the flagship Astra. Both have a 1.05M-token window, 128K output and reasoning effort from none to max. xAI released Grok 4.7 on 21 September, which it calls its most capable model for coding and knowledge work, served at the same price and speed as 4.6.

An enzyme found by agents

On 23 September Anthropic described how Claude found a previously uncharacterized system called array-associated reverse transcriptase, which sits next to a repeating DNA array that resembles CRISPR. About 950 agents searched a large sequence database for roughly 21 hours, with high-level direction from scientists. One agent spotted the repeat beside an unusual reverse transcriptase gene. Anthropic's new life sciences group did the lab validation, and the CRISPR pioneer Feng Zhang reviewed the preprint and commented positively. The function is still unknown.

Smaller releases

Xiaomi opened MiMo-V2.6-Pro on 21 September: 1.02T parameters, 42B active, MIT license. Liquid AI released d1, a decision model that reads input plus typed questions and returns calibrated probabilities with no generated text. Google released Gemini 3.8 Flash TTS models, and Perceptron released Mk1.5, an embodied vision-language model. A paper tests object permanence in video world models.

What I'd watch next

Whether decision-style models that return only probabilities catch on, since they are cheap, quick and easier to check than a chat answer.

What shipped

Claude Opus 5.5

GPT-6 Sol, Luna and Grok 4.7

An enzyme found by agents

Smaller releases

What I'd watch next

Sources