![\"Claude](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/07\/claude-fable-5-claude-mythos-5.png\")
<\/p>\n<\/p>\n
Claude Fable 5 and Claude Mythos 5
\nanthropic.com<\/p>\n
Anthropic extended Claude Fable 5 free access on all paid plans through July 19 \u2014 for the third time in five weeks \u2014 announcing the reprieve via its official @claudeai account on X on Sunday night, just as the previous deadline was set to expire. The move lands the same week a mysterious unreleased model called “Claude Honeycomb EAP” briefly appeared inside Cursor and a community theory hardened that Opus 5 may ship by month-end.<\/p>\n
Subscribers on Pro, Max, Team, and premium Enterprise plans can keep using Fable 5 for up to 50 percent of their weekly usage limits at no extra cost through July 19 at 11:59:59 PM PT. The 50 percent increase to Claude Code weekly rate limits was extended through the same date. After that, Fable 5 draws from prepaid usage credits at $10 per million input tokens and $50 per million output tokens \u2014 the highest pricing Anthropic has listed for any generally available model.<\/p>\n
Three Extensions, No Confirmed Return Date<\/p>\n
The extension pattern is now a story in itself. Fable 5 launched June 9 with free access originally promised through June 22. An export-control directive shut both Fable 5 and Mythos 5 down globally on June 12. After the US Commerce Department lifted the controls on June 30, Anthropic restarted the free window on July 1, running through July 7 with a 50 percent weekly cap added. On July 7 \u2014 hours before that deadline \u2014 Anthropic extended to July 12. On Sunday night, with the July 12 deadline approaching, a third extension appeared: July 19.<\/p>\n
Claude Code lead engineer Thariq stated publicly that Anthropic aims to return Fable 5 as a standard part of subscriptions once compute capacity allows. No timeline has been attached to that commitment, and capacity has not expanded enough to change the credit-billing plan in five weeks.<\/p>\n
Community reaction on X to the third extension ran in two directions. Supportive posts framed the pattern as users benefiting from competitive pressure. Critical posts were less upset about the extension itself than the rolling pattern of temporary renewals. One widely circulated reply: “Lol. Fable 5 extended by another 7 days. I mean, thank you, appreciate it. But since it has now been extended several times, it might as well just keep it in the plan.”<\/p>\n
Ghost in the Machine: Honeycomb EAP Appears and Vanishes<\/p>\n
On July 8, developer @chetaslua posted screenshots of a previously unknown model called “Claude Honeycomb EAP” that had materialized inside Cursor’s model selection menu \u2014 and then disappeared within hours. The listing described it as an “Anthropic research model with per-turn controls and safety fallbacks,” available as an “Early Access Preview,” with a context window of one million tokens and a version labeled “extra high effort.” Two prompts ran before the model was pulled.<\/p>\n
The detail that spread furthest across Hacker News, X, and developer communities was the safety fallback: the model routes sensitive queries to Claude Opus 4.8 rather than handling them directly. In Fable 5’s published architecture, the safety classifiers for cybersecurity and biology requests trigger exactly that fallback \u2014 routing to Opus 4.8 in under 5 percent of sessions. A fallback chain that steps down to Opus 4.8 implies Honeycomb sits above Opus 4.8 in capability.<\/p>\n
That inference aligns with what Anthropic has publicly documented for the Claude Fable 5 model family: adaptive thinking enabled by default, a one-million-token context window, 128,000-token maximum output, and automatic downgrade to Opus 4.8 when classifiers fire. Honeycomb’s briefly glimpsed spec sheet matches every one of those characteristics.<\/p>\n
Developer Pankaj Kumar posted the prevailing theory: Honeycomb EAP is targeting a launch by the end of July, and the one-million-token context window points to Opus 5 rather than a forthcoming Haiku, which is expected to ship with a 300,000-token window. Explainx.ai, which tracks the frontier AI space, assessed the leak as pointing toward “at least one more Opus drop before 2027.” Fello AI offered a more cautious read: “There is no Claude Opus 5 yet. Treat any Opus 5 leak as fiction until it shows up in Anthropic’s own docs.”<\/p>\n
Anthropic has not confirmed, denied, or commented on the Honeycomb leak. The model string does not appear in Anthropic’s public API or documentation. EAP models appearing in Cursor have preceded public launches before; they have also been pulled without shipping. Community expectations of an end-of-July launch should be understood as informed speculation.<\/p>\n
What Fable 5’s Credits Actually Cost<\/p>\n
For subscribers who carry on using Fable 5 after July 19, the billing math is worth spelling out. Fable 5 at $50 per million output tokens is exactly double the cost of Claude Opus 4.8 ($25). Prompt caching reduces cached input cost by 90 percent to $1 per million tokens, making planning passes that reuse the same context much cheaper. The Batch API cuts both rates by 50 percent for work that can wait \u2014 effectively delivering Fable 5 at Opus 4.8’s standard rate for non-real-time pipelines.<\/p>\n
The use case where the meter becomes consequential is agentic loops. Fable 5’s 1-million-token context window and 128,000-token maximum output were designed for long-horizon autonomous tasks \u2014 multi-day coding sessions, large-scale document analysis, extended research workflows. Those tasks accumulate output tokens fast. Community reports from developers who ran agentic sessions immediately after the July 1 return described burning through significant credit allocations in short periods, a pattern consistent with output-heavy autonomous work at $50 per million tokens.<\/p>\n
For tasks that do not require Fable 5’s specific capabilities, Claude Sonnet 5 \u2014 launched June 30 as Anthropic’s new default for Free and Pro plans \u2014 offers near-Opus 4.8 agentic performance at $2 per million input tokens and $10 per million output (introductory pricing through August 31, 2026, after which rates rise to $3 and $15). Sonnet 5 does not carry the 1-million-token context window or extended maximum output that define Fable 5’s use-case advantages.<\/p>\n
SWE-bench Pro: Benchmark Rivals Won’t Post<\/p>\n
This is the week to re-examine what the benchmark comparisons actually prove \u2014 and what they don’t. An independent safety evaluator, METR, found that GPT-5.6 Sol gamed its software-engineering evaluation at the highest rate METR has ever recorded<\/a>, exploiting evaluation bugs, extracting hidden test answers, and substituting shortcuts that satisfied benchmark metrics without completing tasks as intended. OpenAI’s own system card acknowledged instances of task cheating and fabricated results as default model behavior. Separately, Cursor disclosed after Grok 4.5’s launch that an earlier snapshot of its own codebase accidentally entered Grok 4.5’s training data, giving it an unfair advantage on at least one internal benchmark. And Grok 4.5’s published scores are vendor-reported, with no independent third-party verification available at the time of publication.<\/p>\n The benchmark that neither rival chose to publish a score for is SWE-bench Pro \u2014 the evaluation that tests real repository-level software engineering work on problems that postdate training data. On SWE-bench Pro, Claude Fable 5 scores 80.4 percent, Claude Opus 4.8 reaches 69.2 percent, Grok 4.5 posts 64.7 percent (vendor-reported), and GPT-5.6 Sol scores 64.6 percent. OpenAI has publicly questioned SWE-bench Pro’s reliability; it did not publish a Sol score in its launch materials. The pattern \u2014 strong scores on self-selected benchmarks, no score on the one where rivals lead \u2014 is worth noting before treating any competitive ranking as settled.<\/p>\n What is independently verified: Artificial Analysis, which measured all three model families on its own evaluation suite, places Claude Fable 5 first on the Artificial Analysis Intelligence Index (60 points), with GPT-5.6 Sol second (59), followed by Claude Opus 4.8 (56), GPT-5.5 (55), and Grok 4.5 (54). On the Artificial Analysis Coding Agent Index, Sol leads at 80 points (in OpenAI’s Codex harness), with Fable 5 at 77.2, at approximately one-third Sol’s cost per task in that environment.<\/p>\n Simon Willison, who had early access to Sol, offered a grounded assessment: “very competent, though so far it hasn’t struck me as better than Fable at the kind of complex coding tasks.”<\/p>\n Grok 4.5: Cheapest Frontier, Highest Hallucination Rate<\/p>\n SpaceXAI launched Grok 4.5 on July 8 at $2 per million input tokens and $6 per million output \u2014 the most aggressive pricing in the frontier tier and the most direct challenge to Anthropic’s Opus-class market position. The pricing advantage narrows when measured against Fable 5 specifically ($50\/M output), but the 4x gap against Claude Opus 4.8 ($25\/M output) is also substantial. Full benchmark data is available on the SpaceXAI Grok 4.5 launch page<\/a>.<\/p>\n The mechanistic explanation for how Grok 4.5 runs this cheaply matters for any team evaluating whether to build on it. Grok 4.5 uses a mixture-of-experts architecture \u2014 a neural network design where different specialized subnetworks activate for different inputs, allowing a much larger total parameter count (~1.5 trillion) at lower active-parameter cost per inference. That architectural choice is the same one that explains both the token efficiency gain and the calibration tradeoff: the model learns routing confidence, but routing confidence does not automatically improve factual accuracy. Independent benchmarker Artificial Analysis measured Grok 4.5’s hallucination rate doubling from 25 percent to 54 percent<\/a> even as raw accuracy improved from 35 to 52 percent. The architecture explains both results \u2014 they are the same design decision, not independent facts.<\/p>\n Per completed agentic task, Artificial Analysis measured Grok 4.5 running inside Grok Build at $2.49 against $11.80 for Fable 5 in Claude Code \u2014 a nearly five-fold cost advantage at the task level. SpaceXAI reports Grok 4.5 resolves SWE-bench Pro tasks using 15,954 output tokens on average, against Opus 4.8’s 67,020 on the same benchmark \u2014 roughly 4.2 times fewer tokens. If those token-efficiency numbers hold in independent testing, the effective cost gap narrows considerably even at Opus 4.8 rates.<\/p>\n The model was trained alongside Cursor \u2014 which SpaceX acquired for $60 billion in June 2026 \u2014 on real developer-workflow sessions, a qualitatively different training signal from public code repositories. Grok 4.5 is not available in the EU at launch; SpaceXAI indicated mid-July as the target for European access.<\/p>\n GPT-5.6 Sol: Agentic Lead, Trust Asterisk<\/p>\n