I’ve spent the past few days running both models through their paces — Kimi K3 via the Kimi API and the Kimi app, Claude Fable 5 through claude.ai and the Anthropic API — while cross-checking every benchmark figure and pricing detail against the vendors’ own documentation. This isn’t a “which one is smarter” popularity contest. It’s a workload-by-workload breakdown of where each model actually earns its price tag, because the honest answer to “which is better” depends entirely on what you’re building.

The Matchup at a Glance

Kimi K3: The Open-Weight Bet

Moonshot AI shipped Kimi K3 on July 16–17, 2026, live same-day across kimi.com, the Kimi apps, and the API. Under the hood it’s a 2.8-trillion-parameter Stable LatentMoE model with only 16 of 896 experts active per token, a native 1-million-token context window, and built-in image and video understanding. Moonshot calls its new attention mechanism Kimi Delta Attention and credits it with up to 6.3x faster decoding at full context — the same efficiency trick that lets it undercut Fable 5 on price without shrinking the context window.

Reasoning is always-on (“thinking mode”), but at launch it only runs at maximum effort — there’s no lighter tier to dial spend down, and sampling is locked at temperature 1.0 with no user override. If you’re setting this up for real work, our Kimi K3 in Kimi Code setup guide walks through the plan tiers and cache behavior in more detail, and the cache discipline breakdown is worth reading before you commit — switching models mid-session invalidates K3’s prompt cache, which quietly erases its own cost advantage if you’re not careful.

The headline promise is full open weights by July 27, 2026 — ten days after launch. As of today, those weights are promised, not shipped, and the license is still unconfirmed (Moonshot’s K2 line used a Modified MIT license, which is the working precedent). Self-hosting won’t be casual, either: weights ship in MXFP4 and Moonshot recommends 64-plus accelerator supernodes to serve the full model, so this is a datacenter decision, not a laptop download. If your team is planning around the drop, our open-weights readiness checklist and full open-weights review cover what to prep in the ten-day window.

Claude Fable 5: The Closed Incumbent

Claude Fable 5 launched June 9, 2026 as Anthropic’s Mythos-class flagship — closed weights, a 1M-token context window, and a launch framing built entirely around long-horizon autonomy: planning across stages, delegating to sub-agents, and checking its own work over multi-day coding sessions. If you’re weighing it purely as a coding tool against other assistants, our Claude AI coding review and roundup of Claude Code use cases are useful background before you factor Fable 5’s premium pricing into the decision.

Fable 5’s first month has been anything but stable. Three days after launch, on June 12, 2026, U.S. Department of Commerce export controls suspended access to Fable 5 (and its sibling model, Mythos 5) worldwide. Anthropic restored global access on July 1, 2026, once the controls were lifted. Since then, the access terms have shifted twice more: included plan access — up to 50% of weekly usage limits on Pro, Max, Team, and select Enterprise plans — was originally set to end July 7, got extended to July 12, and was extended again to today, July 19, 2026 at 11:59:59 PM PT. From tomorrow, July 20, continued use requires metered usage credits at the standard API rate. If you’re deciding whether Claude in general fits your workflow before you even get to Fable 5’s pricing, our Claude AI trust and safety review is a good starting point.

Unlike K3, Fable 5 exposes a full reasoning-effort range, including adaptive effort — you can dial spend up or down per request. It also carries mandatory 30-day data retention as a Mythos-class model, and standard Zero Data Retention agreements don’t cover it. Cybersecurity- and biology-related prompts get automatically routed to Opus 4.8 through a safety classifier, which matters if your workload ever touches those domains.

Head-to-Head Benchmark Scorecard

Both labs publish results on 14 shared benchmarks, all models run at maximum thinking effort. Scored strictly head-to-head — ignoring every other model on the chart — Fable 5 takes 8 of 14 and K3 takes 6. Treat every number below as vendor-reported launch positioning, not independently reproduced ground truth; Artificial Analysis’s independent Intelligence Index places K3 fourth overall (57.1), behind Fable 5 (59.9) and GPT-5.6 Sol (58.9), and ahead of Claude Opus 4.8 (55.7) — which broadly tracks the vendor story.

Fourteen-benchmark scorecard, vendor-reported at maximum thinking effort. Sources: Moonshot’s Kimi K3 technical blog and Anthropic’s Claude Fable 5 launch materials, July 2026.

The pattern is cleaner than the 8–6 split suggests. Fable 5’s widest margin is FrontierSWE at +5.4 points — the hardest individual engineering tasks — backed up by DeepSWE and even Moonshot’s own internal coding suite. K3’s widest margin is SWE Marathon at +7.0 — long-horizon agentic coding — plus wins on BrowseComp and Terminal Bench. Anthropic has framed Fable 5’s advantage as growing with task length and complexity, but the data here cuts the other way on marathon-length runs specifically: difficulty and duration look like two separate axes, and each model currently owns one of them.

Pricing: The 3.3x Gap, Line by Line

K3 lists at $3.00 per million input tokens on a cache miss, $0.30 on a cache hit, and $15.00 per million output tokens — flat across the entire 1M-token window, with no long-context surcharge. Fable 5 lists at $10 input, $1 cache reads, and $50 output, with Batch API rates cutting that to $5/$25. Divide either side and the ratio lands at roughly 3.3x in K3’s favor across input, cache reads, and output alike.

Two things temper the discount. First, a cheaper model that needs extra retries erases its own savings — and K3’s max-only thinking effort means you can’t dial spend down per request the way Fable 5’s adaptive effort range lets you. Second, Fable 5’s effective cost depends entirely on which side of tonight’s deadline you’re on: plan-bundled access through July 19, then metered credits from July 20. For teams that need a stable 90-day budget line, K3’s flat list pricing is currently the more plannable of the two, with the caveat that its Kimi Code plan tiers — Moderato at $19/month for 256K context, Allegretto at $39/month for the full 1M — are launch pricing, not a locked-in guarantee.

Pros and Cons

Which Should You Actually Pick?

An 8–6 split with a 3.3x price gap isn’t a verdict, it’s a routing table. Here’s how I’d actually assign workloads:

• Frontier-difficulty single tasks: Fable 5. Its FrontierSWE lead is the widest margin on the whole table.
• Long-horizon agent runs: K3, then verify on your own harness — SWE Marathon is its best win, and the savings compound over long sessions. Watch its thinking-history sensitivity if your framework trims reasoning traces.
• Professional knowledge work and research reports: Fable 5 for deliverable quality, K3 if the task is mostly agentic web browsing.
• High-volume, cost-sensitive automation: K3 — flat pricing and a 90%-plus cache-hit discount are built for scale. Keep tool permissions tight; Moonshot itself flags K3’s tendency toward unprompted decisions on ambiguous tasks.
• Regulated or compliance-bound data: Neither is fully comfortable yet. Fable 5’s 30-day retention is at least documented; K3 has no public retention policy at all. Self-hosted K3 weights, once they land, are the only route to full data control here.

How I Evaluated This Comparison

I cross-referenced Moonshot’s own K3 technical blog and launch charts against Anthropic’s Claude Fable 5 documentation and pricing page, then checked both against Artificial Analysis’s independent Intelligence Index to see how much of the vendor story holds up outside the launch chart. I also ran both models through their live interfaces — Kimi’s app and API, and Claude’s web app and API — to confirm pricing pages, plan tiers, and effort-setting controls matched what’s documented, since launch-week pricing pages are exactly the kind of thing that quietly changes within days. One quirk worth flagging for anyone testing K3 themselves: switching models mid-conversation in Kimi Code invalidates the prompt cache immediately, which isn’t obvious until your next invoice is higher than expected.

FAQ

Is Kimi K3 better than Claude Fable 5?

Not outright. Fable 5 wins 8 of 14 vendor-reported benchmarks, mostly frontier-difficulty coding and professional knowledge work; K3 wins 6, including long-horizon agentic coding and web browsing, at roughly a third of the price.

How much cheaper is Kimi K3 than Claude Fable 5?

About 3.3x on every line item: $3 vs $10 input, $15 vs $50 output, and $0.30 vs $1 on cached input.

When do Kimi K3’s open weights actually ship?

Moonshot has promised full weights by July 27, 2026. As of this post, they’re not yet shipped and the license is unconfirmed.

What happens to Claude Fable 5 access after today?

Plan-included access on Pro, Max, Team, and select Enterprise plans ends July 19, 2026 at 11:59:59 PM PT. From July 20, continued use requires metered usage credits at $10/$50 per million tokens.

Was Claude Fable 5 really taken offline after launch?

Yes. U.S. export controls suspended Fable 5 and Mythos 5 worldwide on June 12, 2026, three days after launch. Access was restored globally on July 1, 2026 once the controls were lifted.

Which model has better data privacy?

Fable 5 mandates 30-day retention with no ZDR coverage, but at least it’s a documented policy. K3 has no published retention policy at all, which is arguably harder to clear for a compliance review.

Bottom Line

The benchmark table will age within weeks; the structural fork between these two models won’t. Kimi K3 is a cheaper, open-weight-committed challenger that wins on long agent runs and browsing, but ships with max-only thinking effort and zero published data policy. Claude Fable 5 is the more tunable, more frontier-capable model on the hardest individual tasks, but it’s also the most expensive generally-available model Anthropic has ever priced, and its access terms have moved three times in six weeks. Run both against three of your own representative tasks before committing — price whichever wins at list rate, and revisit the comparison on July 27 when K3’s weights are due to land. For now, this isn’t a winner-take-all call; it’s a routing decision, and the right answer depends entirely on the workload in front of you.

Related reading: Is ChatGPT Good for Coding? · ChatGPT vs Google Gemini for Marketers · 10 ChatGPT Alternatives Worth Trying