Claude Fable 5 vs. MiniMax M3: Real-World Agentic Benchmarks Compared

Claude Fable 5 and MiniMax M3 shipped nine days apart in June 2026, and they represent two different bets about where agentic AI is headed.

Fable 5 is Anthropic's Mythos-class flagship, built to plan, delegate, and grind on a task for days inside a harness like Claude Code. MiniMax M3 is a 428B-parameter open-weight mixture-of-experts model built around a sparse-attention architecture that makes million-token context practically affordable.

The short version: Fable 5 leads on every coding and terminal-agentic benchmark either vendor has published, sometimes by a wide margin. MiniMax M3 costs about 3% as much per token, ships its weights, and beats Anthropic's previous flagship on a couple of benchmarks. Neither fact cancels the other out.

A note on methodology

This isn't a live task-execution study. Both models weren't wired into the same set of tools and run through identical agent tasks side by side, the way some benchmark write-ups do it. Neither model's API was available to run that kind of test for this piece.

Instead, every number below comes from vendor benchmark pages, Artificial Analysis, OpenRouter, and independent evaluators, cross-checked across sources and cited at the bottom. Where sources disagree, which happens more than benchmark leaderboards like to admit, it's flagged instead of averaged away.

The two models, side by side

The Intelligence Index rankings aren't directly comparable since Artificial Analysis tests each model against a different field size. The raw 60 vs. 44 gap is the more honest number to look at.

Gap 1: coding and terminal-agentic benchmarks

This is where the two models separate most clearly. Fable 5 leads on every reported coding and terminal benchmark, and the gap widens on tasks that require holding more state: chaining tool calls, recovering from a failed command, planning across a longer session.

* Anthropic-reported. At least one independent evaluator publicly disputed this figure and reported a lower score on re-run, so treat 80.3% as the upper bound rather than a settled number.

** Also reported as 83.4% in other sources. Terminal-Bench 2.1 has the same disagreement problem, just a smaller one.

MiniMax hasn't been benchmarked in public against Fable 5 on MCP Atlas, SWE-fficiency, or KernelBench Hard, so those rows only tell you about M3 in isolation. Still, 74.2% on MCP Atlas is a strong number for a model priced at a few cents per million tokens, and it matters if your agent's real job is mostly MCP-style tool orchestration rather than open-ended software engineering.

Gap 2: where MiniMax M3 actually wins

M3 doesn't beat Fable 5 head-to-head on any published benchmark. It does beat Claude Opus 4.7, Anthropic's previous-generation flagship and still a widely deployed model, on two of MiniMax's own reported benchmarks:

These are MiniMax's own comparisons, and they're against Opus 4.7, not Fable 5. Anthropic hasn't published a head-to-head against M3 on either benchmark. So this is "M3 vs. the previous Claude generation," not "M3 vs. Fable 5," and the distinction matters more than it looks like it should.

Gap 3: speed, and a genuinely strange number

On raw throughput, M3 is faster: roughly 84 to 86 output tokens per second against Fable 5's 66.1 (the reasoning-model median sits around 71). But the number that actually matters for an interactive agent is time-to-first-token, and there it isn't close.

• MiniMax M3: 1.8 to 2.1 seconds to first token
• Claude Fable 5: 149.8 seconds to first token, against a median of 2.6 seconds for comparable models

That 150-second figure isn't a typo. It's the cost of Fable 5's extended-thinking mode doing substantial reasoning before it emits anything visible. For an agent planning across dozens of tool calls, that upfront thinking is the point. For a chat product where a user is staring at a loading spinner, it's close to disqualifying.

Gap 4: cost, compounded

Here's where the story tips the other way.

A single agent task rarely costs enough for that gap to register. Ten thousand agent tasks a day, running continuously, is a different conversation. At volume, the cost curve stops being a footnote and starts being the deciding factor, especially for tasks where M3's lower ceiling on raw accuracy still clears the bar.

The part that isn't in any benchmark: availability risk

On June 12, 2026, three days after Fable 5 launched, the U.S. government applied export controls to Fable 5 and its sibling Mythos 5. Anthropic had to restrict access for foreign nationals. Access was fully restored globally on July 1.

For three weeks, any product built on Fable 5 outside the U.S. was running on a dependency that could legally disappear again without much notice. MiniMax M3, being open weight, has no equivalent failure mode. You can pin a checkpoint and keep running it no matter what any government decides next month. That's not a benchmark score, but for procurement purposes it belongs on the same spreadsheet as SWE-Bench numbers.

What this actually tells us

Fable 5 is the more capable model on every coding and terminal-agentic benchmark either vendor has published, and sometimes the margin is wide. That's the whole story if raw accuracy on hard software engineering tasks is what you're optimizing for.

MiniMax M3 isn't trying to win those benchmarks. It's competing on a different axis: cost per correct task, deployment freedom, and a growing list of areas (BrowseComp, MCP tool-use) where "good enough, at 3% of the price, with weights you control" beats "best possible score" as the relevant comparison.

Which one to use mostly comes down to what you're optimizing for:

Both numbers are true at once. Fable 5 wins on the leaderboard. MiniMax M3 wins on the invoice. Which one matters more depends entirely on what you're building.

Benchmark scores are self-reported by vendors unless otherwise noted. Methodologies differ between test suites, and figures should be treated as directional rather than exact, especially where sources disagree, as flagged inline above.