AI this week: GPT-6 Astra, Claude Fable 5.1, and a transcription race

OpenAI and Anthropic trade the top spot, while Google, Meta and Microsoft all release speech-to-text models.

September 11, 2026 · By Bohol AI

The two leading labs released new flagships within a week of each other, and speech-to-text quietly became one of the most competitive areas in AI.

GPT-6 Astra

OpenAI launched GPT-6 Astra, its new flagship. It tops or nearly tops the major leaderboards while using fewer tokens, and costing less, than the few models that beat it. OpenAI says it is the first model to reach the “critical” cybersecurity level in its risk framework, so its most advanced security abilities are limited to selected organisations.

Claude Fable 5.1 and Mythos 5.1

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1, the same model under two names with different safeguards. Fable is open to everyone; Mythos is limited to selected U.S. cybersecurity and life-sciences organisations. It earned top marks from the independent evaluators Artificial Analysis and Vals AI.

Three new transcription models

Google released Gemini 3.5 Transcribe on August 26, Meta launched Muse Voice Transcribe on September 1, and Microsoft released MAI-Transcribe-2. All three turn speech into clean text for use in wider AI workflows.

Why it matters: competition is making accurate transcription cheap, which helps anyone who records meetings, interviews or classes.

Research: agents that compress their memory at the right moment

Researchers at Johns Hopkins and Apple built SelfCompact, which lets an agent decide when to summarise its working memory at natural stopping points instead of cutting the oldest material.

What it means for Bohol

Try these transcription tools on Bisaya and Taglish recordings before relying on them; most are trained mainly on English and major languages.
Top models now change every few weeks. Pick tools you can switch easily rather than locking into one.

Sources

All articles