Anthropic's Opus 6 Ignites a Terrifying AI Theory

A 17-minute news roundup, uploaded on 16 October 2026. Despite the title, most of it is sorting rumours, more carefully in some places than others.

The first is “Claude Fable 5.5”. It turned up in LiteLLM’s model list on 13 October, after Azure’s East US 2 model directory listed it as “generally available”. Everything else in the entry, including the 1M-token context and 128K output, was copied from Fable 5.1, because no source gave figures for 5.5. The narrator notes the precedent (Fable 5.1 surfaced on a cloud listing a day before release) and the skeptics’ case: a config sync is not an independent second source, and a directory entry is not an announcement. Two pull requests to remove the entry followed a day later, and the “leaked” $10/$50 price was copied from Fable 5.1 too. Anthropic hasn’t announced it. The video then repeats a run of social-media anecdotes credited to the unreleased model, from an MMORPG built in five hours to a nine-minute symphony, and a theory that Opus 5.5 quietly routes hard prompts to it, without the same scrutiny.

The second rumour, Opus 6, it debunks. The rumour traces to a single Geeky Gadgets article (14 October) crediting another YouTube channel, with no model card, date or benchmarks, and it spread through aggregators. The narrator’s argument against a rushed release is that Anthropic’s IPO prospectus reportedly runs to about 80 pages of risk factors. The prospectus was filed in confidence and isn’t public; the figure comes from press reports of it. It also covers OpenAI’s reported “Bell” model, where claims of 10T and 100T parameters differ tenfold.

The real news is fast mode. On the API it is generally available for Opus 5.5: the same weights, up to 2.5 times the output speed, at $8/$40 per million tokens, on the Claude API and Microsoft Foundry but not Bedrock. The narrator says it is GA for Opus 5 and 4.8 too, but those are still in research preview. In Claude Code, /fast defaults to Opus 5.5 from v2.1.280. Idle compaction summarises a long, idle session just before the hour-long prompt cache expires. Users complained it lost detail, and it can now be switched off. The last item is Meta’s Muse Spark Open: 816B parameters with 42B active, 1M context and open weights under Meta’s own licence, built with Apple to run across two M5 Ultra Mac Studios through MLX. The narrator says it isn’t the strongest open model on coding but does well on agent tasks, and warns that harness mixes make those comparisons shaky.