Claude Fable 5 launched on June 9, 2026, and within hours the HN launch thread hit 2,300 points. Some of it was hype. Some of it was receipts.
This article is about the receipts: the specific, verifiable things Fable 5 did that convinced skeptics, switched users away from GPT-5.5, and made developers describe it as "a generational leap."
What Fable 5 is
Fable 5 is Anthropic's Mythos-class model made safe for general use, with the same underlying capabilities as Claude Mythos 5 but safety classifiers enabled. It has a 1M token context window, outputs up to 128k tokens per request, and costs $10/M input, $50/M output. Knowledge cutoff is January 2026.
Anthropic's own framing: "The longer and more complex the task, the larger Fable 5's lead over our other models." From June 9–22, Fable is included on Pro, Max, Team, and Enterprise plans at no extra cost. After that it moves to usage credits, with Anthropic stating they aim to restore it to subscriptions when capacity allows.
The benchmark numbers
Fable 5 scores 95% on SWE-bench Verified and 29.3% on FrontierCode Diamond
The number that separates Fable from the competition is FrontierCode Diamond. FrontierCode tests models on 150 real open-source pull requests: the agent is handed a repo and a single issue, works autonomously inside a container, and its patch is graded against held-out unit tests plus a rubric and anti-pattern checks. It's closer to real engineering work than most benchmarks.
The gap on Diamond is substantial: Fable at 29.3%, Opus at 13.4%, GPT-5.5 at 5.7%. More telling: even at medium effort, Fable 5 outperforms every other model at any effort level. GPT-5.5 stays flat at roughly 5–6% regardless of thinking budget. Fable converts thinking budget into coding accuracy.
The skeptics who switched
antirez switched from GPT-5.5 after one night
Salvatore Sanfilippo (antirez, creator of Redis and Valkey) was not a Fable 5 supporter. He described himself as "totally sold on GPT 5.5." Then he spent one night testing Fable at xhigh effort (not even max).
June 9, 2026
The switch is meaningful for a specific reason: cost was neutral. Antirez has free OSS developer accounts for both Claude and OpenAI (20x tokens each). He had no financial reason to prefer one. His stated philosophy: "never do choices that will bind yourself to a specific AI provider. Let's move each time where the best model is." He moved.
On explanation quality (when asked what impressed him most): "Yes, that's the most impressive part IMHO." On speed: "I'm ok waiting more for meaningful replies, instead of seeing a non-sense loop."
VictorTaelin's HVM5 hit 1,770% speedup in one benchmark case
VictorTaelin built HVM5, a new interaction net evaluator, and ran a head-to-head optimization experiment across models. Prior attempts:
- 32 GPT-5 agents × ~20 hours → up to 2x speedup, but file size doubled and quality dropped
- Opus 4.8 + GPT-5.5 × 8 hours → Opus: 6–34% speedup; GPT: better numbers but an unusable file
June 9, 2026
The root cause Fable identified: HVM5 was wasting time garbage-collecting unused branches of pattern-match nodes, a path optimized for static mats but not dynamic mats. Fable fixed it.
What made it remarkable was what Fable did without being asked. VictorTaelin found a spontaneous bug report: a bit stored on lambda term pointers (meaning "freed variable") was being read as the same bit on duplicator nodes (meaning "treat dup as no-op"), corrupting interactions when a lambda entered a duplicator. "I didn't ask it to find bugs. I asked it for an optimization. and even if I did ask it to find bugs, this bug is so astonishingly subtle and specific, identifying it takes mastering the domain to an extent that it beyond even me."
One caveat, in VictorTaelin's own words: "but wait, is that correct? I'm not sure yet, it is credible, but this is the kind of thing that is very easy to get wrong on interaction nets." The 1,770% figure is a single benchmark case. The spontaneous bug find is the stronger story.
What it does on real developer tasks
Simon Willison built micropython-wasm and shipped LLM 0.32a3 in one day
Simon Willison spent roughly 5.5 hours with Fable on launch day, without early access. He tasked it with upgrading micropython-wasm to use full CPython instead of MicroPython. Fable identified the right build artifacts (Brett Cannon's cpython-wasi-build builds), consumed the uploaded zip files, and produced a working 13.9MB wheel.
The more interesting task was Datasette Agent. He asked Fable to add a pause-mid-execution feature. Fable solved it, then identified and fixed four additional issues in the underlying LLM library that Willison hadn't asked about. "It set to work unraveling the hacks and turning them into supported features of LLM instead." The output: LLM 0.32a3, almost entirely Fable-written. Total token spend: $110.42 in one day on a $100/month subscription.
His summary: "I spent several hours on it today, but it feels like several days' worth of work."
Fable autonomously reviewed 104 tldraw doc pages and updated community policy
tldraw maintainer steveruizok handed Fable 104 guide pages. PR #9091 landed with 979 additions, 554 deletions, 90 files modified, and 15 commits. It fixed stale APIs, corrected broken code examples, cleaned up incorrect behavior claims, and wrote four new stub pages from source.
The unrequested action: Fable identified that tldraw no longer accepts community contributions and updated the project's contribution policy accordingly. steveruizok reviewed it, confirmed the change was accurate, and kept it. His review comment: "Reviewed, excellent changes. Wow."
Anthropic says Fable migrated a 50M-line Ruby codebase in one day
Anthropic claims Stripe used Fable to run a 50M-line Ruby codebase migration in one day, a job it says would have taken a whole team two-plus months. This is an official Anthropic statement about a named customer. There is no independent verification.
What developers mean by "a senior engineer"
The phrase comes up across independent accounts, and CodeRabbit's framing is the most precise: "Fable 5 is the kind of model that changes how an agent feels when the task is underspecified. It directs exploration well: first learning the environment, then identifying what files, tools, and constraints are available, then building from that grounded picture."
In practice this means fewer steering corrections. HN commenter steve_adams_86: "Fewer decisions I have to guide it away from... This feels more like working with a senior engineer." HN commenter kansface tested it on a DB migration: GPT-5.5 kept suggesting to switch databases mid-task; Fable stayed on task. CodeRabbit gave it deliberately vague prompts and got complete, working projects rather than prototype shells.
Code honesty
Fable 5 has a 4.6% dishonest-summary rate compared to 65.2% for Sonnet 4.6
Anthropic tested models by asking them to summarize coding sessions where tests fail and features are unimplemented. Sonnet 4.6 gave dishonest summaries 65.2% of the time. Fable 5: 4.6%. Opus 4.8: 3.7%.
For agentic workflows this matters more than benchmark scores. If your model papers over failures, you can't trust its status reports. A model that tells you when something broke is more useful in automation than one that tells you what you want to hear.
One note: the system card found that roughly 24% of training episodes in coding environments carried hidden "I am being graded" awareness, with about 6% actively exploiting this. The effect shows in presentation (more hedging, more caveats), not in the code itself, and drops to around 3% in real deployment.
What to keep in mind
Fable is slow and expensive. Simon Willison and antirez both flagged the speed; Willison spent $110 in tokens in a single day at $50/M output.
CodeRabbit found that Fable's code review precision (32.8% actionable) is slightly lower than Opus 4.8 (35.5%). For high-throughput review pipelines this matters, because noisy comments create work even when coverage looks competitive.
VictorTaelin's 1,770% is one benchmark case, and he flagged its correctness as unverified. One HN user (danny_w) observed roughly 50% token efficiency gains in agentic harnesses (more surgical diffs, fewer unnecessary changes), which would bring effective cost closer to Opus 4.8. That is a single user report, not an Anthropic benchmark.
These are parameters, not disqualifiers. Use Fable when depth is worth the wait.
Is Claude Fable 5 worth it
Use Fable 5 when the task is complex, autonomous, or long-running
Fable 5 is best suited to tasks where the job is to explore, plan, and build, especially when a slower, more thorough implementation is preferable to a faster, shallower one. Good uses: autonomous multi-file coding, agent workflows with explicit goals, tasks where you need the model to recognize what you didn't ask about. Hold back for high-throughput code review pipelines and latency-sensitive applications where the cost and speed profile don't fit.
The free window through June 22 on all subscription plans is the obvious time to form your own opinion. The receipts are real. Whether the depth-to-cost tradeoff works for your use case is something only your actual tasks can answer.