Anthropic released Claude Fable 5.1 and Mythos 5.1 yesterday morning, calling Fable 5.1 its most capable generally available model. Pricing stays at $10/MTok input and $50/MTok output, with a 1M-token context window and a June 2026 knowledge cutoff. Max output is 128K. The change that saves money is cache reads, down from $1.00/MTok to $0.25/MTok, which Anthropic puts at roughly 25% less than Fable 5 for typical workloads, up to 45% for heavily agentic work.
Three API changes will break existing code, all on the what’s new page. Forced tool use (tool_choice: {"type": "any"}) now returns a 400. Thinking blocks are bound to the model that produced them, so earlier Claude models can no longer read Fable 5.1’s. And for accounts created on or after August 31, editing an earlier turn invalidates the thinking blocks after it.
TL;DR
Fable 5.1 and Mythos 5.1 are the same model with different safeguard levels. Mythos is gated to approved customers in Project Glasswing.
Anthropic’s docs still tell most users to start with Opus 5 and turn to Fable 5.1 only for demanding reasoning and long-horizon agentic work.
Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 against Opus 5’s 29.0%, on Anthropic’s own numbers.
On a separate third-party composite, Artificial Analysis’s Intelligence Index, it scores 66 to Opus 5’s 63.
Cache reads cost 75% less. Nothing else in the price table moved.
It wrote an ISO 8601 duration parser that passed 19 of its own tests first try. A reviewer on another model found a real bug anyway. Its unedited prose scored 100% AI on Pangram.
The Fable/Mythos Split
Fable 5.1 and Mythos 5.1 are, in Anthropic’s words, “the same model, but with different levels of safeguards”. Mythos is offered only to approved customers in Project Glasswing, reached through an Anthropic, AWS, or Google Cloud account team, at identical specs and pricing.
The restriction shows up inside Fable 5.1 too. It routes dual-use cybersecurity work (penetration testing, exploit generation, binary vulnerability scanning) and life-sciences R&D queries to the Opus models. Anthropic reports that the new safeguards block 60% fewer false positives than before, and that Claude Code users should see roughly 60% fewer cyber-safeguard interventions per session than on Fable 5. That’s the number I’d want if I ran a security team that kept getting refused mid-task.
Then the docs decline to recommend it. “For most workloads, start with Claude Opus 5 ... Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short.” Anthropic put its new flagship at the top of the comparison table and told most of its customers to keep using the model below it. Price is the likeliest reason: that table puts Opus 5 at $5/$25 per MTok, half Fable 5.1’s rate.
What the Benchmarks Say
These are Anthropic’s own figures from the announcement, several corroborated independently by VentureBeat.
Benchmark Fable 5.1 Fable 5 Opus 5 GPT-5.6 Sol Terminal-Bench-Science 0.1 52.6% 24.7% 29.0% 22.4% Terminal-Bench 4.0 55.8% 42.0% 52.3% 37.3% AutomationBench 31.4% 17.1% 26.9% 19.6% CursorBench 3.2.0 73.4% 70.5% 70.0% 67.2% Humanity’s Last Exam (no tools) 60.9% 57.8% 56.6% —
The Terminal-Bench-Science jump is the one people are talking about, and it earns the attention. Doubling a science-agent score is a different kind of result than a three-point coding gain. Everywhere else the margins are ordinary: 2.9 points over Fable 5 on CursorBench, 3.5 over Opus 5 on Terminal-Bench 4.0.
Artificial Analysis’s Intelligence Index has Fable 5.1 at 66 at max effort, ahead of Opus 5 at 63 and GPT-5.6 Sol at 61. CodeRabbit ran its 105-point internal code-review benchmark on launch day and got 61.0% recall for Fable 5.1 against Fable 5’s 61.9%, precision up from 32.8% to 37.3%. Their summary: “Fable 5.1 found one fewer known-issue point than Fable 5. The improvement came from reducing output.” Fable 5.1 is not on the LMArena leaderboard yet.
A Quick Code Test
I ran both tests on release day from a session in trusty-mpm, the open-source multi-agent harness I maintain for Claude Code, dispatching subagents pinned to claude-fable-5-1.
The spec was an ISO 8601 duration parser with the traps that make that format annoying: a fraction only in the last component present, no combining the week form with anything else, strict ordering, no repeats, and a leading sign that negates the whole duration. Fable 5.1 wrote the module and a unittest suite, ran it, and got 19 of 19 first try.
Then a reviewer agent on a different model probed the module with 31 inputs and found what the suite missed. The component regex is re.compile(r"(\d+)(?:([.,])(\d+))?([A-Z])"), and \d in Python’s re matches any Unicode decimal digit. So PT١H, with an Arabic-Indic numeral, parses silently to 3600.0, identical to PT1H. The designator class is locked to ASCII [A-Z], so a fullwidth H gets rejected. The numeral class isn’t. For a parser whose whole premise is loud rejection of malformed input, a silent accept is the wrong failure mode, and the fix is one character class.
The verdict was PASS WITH NOTES. The quality notes were good ones: docstrings on every function, a typed constant table instead of magic numbers, and error messages naming the offending input and the reason (invalid ISO 8601 duration 'PT1.H': unexpected text at '1.H'). Nineteen real tests, none tautological, and one gap the model wrote its tests around.
A Quick Writing Test
The second test was a 350-word opinion prompt with no style guidance, on why the pull request is the wrong review artifact for agent-written code. Fable 5.1 returned 369 words:
Every review process is a bet about where mistakes hide. The pull request bets they hide in the diff: read the lines that changed, and you will find the bug. For code written by a person, that bet mostly pays.
That is a good paragraph. The argument holds up, the closing analogy (”Nobody read raw packet captures either until someone built Wireshark”) lands, and I’d have been happy to find it in a newsletter.
Scored raw by Pangram 3.3.2, it came back at 100% AI across a single 369-word window, assistance score 99.3%, confidence high. The structural fingerprint is the giveaway, not the vocabulary. All five paragraphs end on a quotable clincher. Three consecutive sentences open with “It shows.” And the piece announces its own rhetorical choreography before executing it:
The obvious objection is that transcripts are enormous and nobody will read them. [...] I think this objection is right about the volume and wrong about the conclusion.
Nobody arguing a point in real time sets up their own counterargument that tidily. I spent most of last month counting Opus 5’s verbal tics in my session logs, so I went in expecting Fable 5.1’s prose to read better. It does, and it still sits at the ceiling of the detector’s range.
One Day In
Two small tests on launch day tell you what two small tests tell you. The parser result says Fable 5.1 can hold a fiddly spec in its head and write tests covering the parts it thought about, and that a reviewer on a second model is still worth the tokens. The Pangram result says the raw output is unmistakably model prose, which matters if you publish and not at all if you ship code.
The pricing is the part I’d act on this week. Cache reads at a quarter of the old rate change the arithmetic on long agentic sessions more than three points of Terminal-Bench do. The rest can wait for the eval you run against your own workload, which is what the docs tell you to do anyway.
I use Fable exclusively as my PM runner now, as long as I have the credits.
Bob Matsuoka is CTO of Duetto, a hospitality revenue-management platform, and writes about AI-augmented engineering practice.
Related reading:
The Word Problem — Counting Opus 5’s verbal tics in a month of session logs.
It’s the Harness, Stupid — Why the tooling around the model has mattered more than the model since April.
Claude Sonnet 5 Takes the Default — The last time a Claude launch changed which model I reach for first.
AI Power Ranking — Tool comparisons and benchmarks for AI practitioners.
LinkedIn Newsletter — Strategic AI insights for CTOs and engineering leaders.