Claude Fable 5.1 and GPT-6 Astra both land, plus Gemini 3.8 Flash

The two biggest labs both shipped a flagship within three days of each other, and I spent more time on the price tables than on the benchmarks.

Claude Fable 5.1 and Mythos 5.1

Anthropic released both on 1 September. Fable 5.1 is generally available. Mythos 5.1 is the same model with more permissive safeguards, limited to vetted cybersecurity and life sciences users through trusted-access programs. Anthropic says the safeguards got more precise: cyber interventions in Claude Code drop by about 60 percent per session, and biology safeguards fire 85 percent less often on harmless requests.

The commercial change is in caching. Cache reads fall 75 percent to $0.25 per million tokens, which Anthropic estimates trims typical costs about 25 percent and highly agentic ones by up to about 45 percent. Base rates stay at $10 per million input tokens and $50 per million output. Reported results against Fable 5 include Terminal-Bench 4.0 at 55.8 percent versus 42.0, and Humanity's Last Exam without tools at 60.9 versus 57.8.

GPT-6 Astra

OpenAI announced GPT-6 Astra on 3 September. The docs describe it as the most capable model in the lineup for complex reasoning, coding, computer use, research and document creation. It has a 1,050,000-token context window, up to 128,000 output tokens and an April 2026 knowledge cutoff, and costs $10 per million input tokens and $50 per million output. Batch and Flex run at half the standard rate.

Gemini 3.8 Flash and Muse Spark 1.3

Google released Gemini 3.8 Flash and a Cyber variant on 2 September, aimed at long-horizon software engineering and agents, and launched the Fairwind Program for trusted cyber defenders the same day. Meta released Muse Spark 1.3 on 2 September, tuned for multi-step tool and browser work.

Regional and open models

On 3 September the Institute of Foundation Models at MBZUAI released the K2 Horizon family, from 0.9B to 375B parameters, and the 375B flagship has Apache 2.0 weights and 512K context. HUMAIN also introduced M3, an Arabic-focused 428B multimodal model in limited preview on HUMAIN Node.

Science and forecasting

DeepMind released WeatherNext 3 on 3 September, and Google Research released TimesFM-3 on 31 August, a 330M-parameter time-series model for several targets at once. A paper asking whether an LLM can rewrite its own agent scaffold topped the Hugging Face list.

What I'd watch next

Whether other labs match the cache-read price cut, since that is the line item that decides what long agent runs cost.