Anthropic extends Fable 5 again — and won’t talk about what developers found inside Cursor

Anthropic pushed its Claude Fable 5 promotion again Sunday night, hours before the previous deadline expired, giving paid subscribers free access through July 19. That's three extensions in five weeks, and the pattern is starting to look less like generosity and more like a company buying itself time. Once the offer ends, Fable 5 reverts to $10 per million input tokens and $50 per million output tokens — the priciest tier Anthropic sells.

The extensions trace back to June 9, when Fable 5 launched with free access through June 22. Three days later, a U.S. export-control directive forced Anthropic to yank both Fable 5 and its sibling Mythos 5 worldwide. Access came back July 1 after the Commerce Department lifted restrictions on June 30, but with a 50% weekly usage cap attached, and every deadline since has been pushed at the last minute rather than resolved. Anthropic's Claude Code lead, Thariq, has said the plan is to make Fable 5 permanent once compute capacity allows — but there's no date attached to that promise, and subscribers are noticing.

What's actually irritating users isn't the extension itself, it's that Anthropic keeps extending the window without resetting anyone's usage allowance. On Reddit, people who already burned through their 50% weekly cap say another seven days of "free access" is worthless to them. They've pointed out that OpenAI has reset usage caps after outages or demand spikes before, and some say the lack of a reset is nudging them toward rival models entirely.

That competitive pressure got sharper this week when developer @chetaslua spotted an unlisted model called Claude Honeycomb EAP sitting in Cursor's model picker on July 8, described as a research preview with per-turn controls, safety fallbacks, a million-token context window, and an "extra high effort" mode. It vanished within hours, but not before screenshots spread across X and Hacker News. The detail that got people talking: sensitive prompts in Honeycomb reportedly got routed down to Claude Opus 4.8, the same fallback behavior Anthropic already documents for certain cybersecurity requests in Fable 5. If a model routes down to Opus 4.8 for safety, the logic goes, it probably sits above it — fueling theories that Honeycomb is an early build of Opus 5, possibly landing before July ends.

The stakes for that launch aren't small. Fable 5 still leads on repository-scale coding benchmarks like SWE-bench Pro, but OpenAI's GPT-5.6 Sol and xAI's Grok 4.5 are chasing the same high-end developer market at noticeably lower prices. Anthropic has tools to soften Fable 5's cost — prompt caching cuts cached input token prices by 90%, and the Batch API halves both input and output pricing for jobs that can wait — which brings scheduled workloads roughly in line with Opus 4.8's $25-per-million-output rate. But for anyone running long agentic sessions with heavy output, the bill adds up fast, and Claude Sonnet 5, priced at $2 and $10 per million tokens through August 31, remains the more sensible default for most day-to-day coding work.