Kimi K3 vs Claude Fable 5: two frontier bets, one decision

Kimi K3 vs Claude Fable 5 is the frontier matchup of July 2026: an open-weight-committed 2.8-trillion-parameter challenger from Moonshot AI against Anthropic’s closed, agentic flagship. Across the 14 benchmarks both vendors report head-to-head, Fable 5 wins eight and K3 wins six — every number vendor-reported, every model at maximum thinking effort.

The stakes are unusually concrete this week. K3 launched July 17 with flat $3/$15 per-million-token pricing and a promise of full open weights by July 27. Fable 5’s discounted plan access ends July 19 after two extensions, with metered credits starting July 20. Whichever model you pick, its access story changes within ten days of this post going live.

This guide recomputes the head-to-head scorecard from the launch data, translates the margins into workload language, prices both models honestly, and closes with a decision framework by workload — not a winner-take-all verdict. Benchmark figures come from Moonshot’s K3 technical blog and launch charts; pricing and policy facts from Anthropic’s and Moonshot’s own pages.

Kimi K3 launched on July 17, 2026 — live same-day on Kimi.com, Kimi Work, Kimi Code, the Kimi API, and mobile apps. (Some US press datelined the launch July 16; that’s a Pacific-vs-China-time artifact. Moonshot’s own materials use July 17, and so does our full Kimi K3 release coverage.) Under the hood: 2.8 trillion total parameters in a Stable LatentMoE design with 16 of 896 experts active per token, a native 1M-token context window, and native multimodal input for images and video. Moonshot credits Kimi Delta Attention with up to 6.3x faster decoding at 1M-token contexts and claims roughly 2.5x overall scaling efficiency versus K2 — vendor figures, like every architecture claim at launch.

Claude Fable 5 launched June 9, 2026 as Anthropic’s Mythos-class flagship: closed weights, 1M context, and a launch framing built around long-horizon autonomy — positioned as the company’s most capable model for ambitious coding projects, able to plan across stages, delegate to sub-agents, and check its own work over multi-day sessions. Stripe reported it compressed months of engineering into days — a customer anecdote relayed in Anthropic’s own launch post, not an independent audit. We compared it against OpenAI’s flagship in our Fable 5 vs GPT-5.5 comparison — this post applies the same discipline to the open-weight challenger.

The comparison below is deliberately a two-model cut. Moonshot’s launch chart is a multi-model grid that includes GPT-5.6 Sol — a model that undercuts Fable 5 at $5/$30 and tops a few of these same benchmarks — but folding it in buries the direct question most teams are asking this week. For the three-way agentic picture, see how K3 stacks up against GPT-5.6 Sol.

The table below takes the 14 benchmarks both vendors report on the same chart and scores them strictly head-to-head: each row goes to whichever of K3 or Fable 5 posts the higher number, ignoring every other model in the grid. Counted that way, Fable 5 wins 8 of 14 and K3 wins 6 of 14. One row — SpreadsheetBench 2 at 34.8 vs 34.7 — is effectively a statistical tie that we credit to K3 only because the number is nominally higher; treat it as a dead heat in practice.

An 8–6 split sounds like a coin flip. The margins say otherwise — they sort the two models into distinct workload territories. Fable 5’s widest win is FrontierSWE at +5.4 points, the frontier-difficulty software-engineering suite, backed by wins on DeepSWE and even Moonshot’s own internal coding benchmark. K3’s widest win is SWE Marathon at +7.0 points — the long-horizon agentic coding suite — plus BrowseComp for agentic web research and Terminal Bench for terminal-driven work.

There’s a genuinely interesting tension in these two headline margins. Anthropic’s launch framing claims that the longer and more complex the task, the larger Fable 5’s lead — yet the long-horizon SWE Marathon suite is K3’s single biggest win, while Fable 5 dominates the hardest individual engineering tasks on FrontierSWE. One vendor-reported reading: difficulty and duration are different axes, and each model currently owns one of them. That’s exactly the kind of split you can only resolve by running your own multi-hour agent harness against both.

The knowledge-work rows are less ambiguous. On GDPval-AA v2 — Elo rankings across real professional tasks — Fable 5’s 1760 sits 92 Elo points above K3’s 1668. For scale, that gap is larger than the 12 Elo separating Fable 5 from GPT-5.6 Sol (1748) on the same metric, and independent coverage placed K3 third on this leaderboard behind both closed flagships. Add JobBench (+4.5) and AA-Briefcase (+35 Elo) and the pattern holds: for professional-deliverable work at max effort, the vendor-reported data consistently favors Fable 5.

K3 lists at $3.00 per million input tokens (cache miss), $0.30 on cache hits, and $15.00 per million output tokens — flat across the entire 1M-token window, with no context-length tiering. Fable 5 lists at $10 input, $1 cache reads, and $50 output, with Batch API rates of $5/$25. Divide any line item and the ratio lands at roughly 3.3x in K3’s favor: 10÷3, 50÷15, and 1÷0.30 all round to 3.3.

Two caveats keep the 3.3x from being the whole story. First, benchmark parity per dollar only matters if the cheaper model completes your task — a model that needs two extra retries erases its own discount, and K3’s max-only thinking effort means you can’t dial spend down per-request the way Fable 5’s full effort range allows. Second, Fable 5’s effective cost depends on which side of July 20 you’re on: plan-bundled access at up to 50% of weekly limits until July 19, then metered credits. For the full post-promo economics, see what Fable 5 costs once metered access begins. 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 Kimi Code plan tiers (Moderato at a listed $19/mo for 256K context, Allegretto at $39/mo for the full 1M) are listed launch pricing, not a guarantee.

The deeper fork between these models isn’t on the scorecard at all — it’s who controls access. K3’s story: Moonshot has promised full model weights by July 27, 2026, ten days after launch, with an accompanying technical report. As of this post, those weights are promised, not shipped, and the license is unconfirmed (the K2 line’s precedent is Modified MIT). Nor will self-hosting be casual: weights ship in MXFP4 and Moonshot recommends 64+ accelerator supernodes for serving — this is datacenter-scale open weights, not a laptop download. If you’re planning for the drop, work through prepping for K3’s open-weight release on July 27.

Fable 5’s story is the mirror image: fully closed, and the access terms have moved twice in twelve days. The promotional plan inclusion was originally set to end July 7, was extended to July 12 hours before the deadline after user backlash, then extended again on July 13 — this time to July 19 at 11:59:59 PM PT. We unpacked the whole saga in the second promo extension through July 19. And closed weights carry a tail risk open advocates rarely have to argue anymore: US Commerce export controls suspended all Fable 5 access worldwide on June 12, three days after launch, and access wasn’t fully restored until July 1.

The honest synthesis: neither access story is stable this month. K3’s openness is a dated promise with an unconfirmed license; Fable 5 is a closed model whose commercial terms have shifted twice in two weeks and which spent nearly three weeks of its first month unavailable by government order. Teams building on either should write down their assumptions — and a fallback — before July 27.

For enterprise buyers, the sharpest single difference is data handling. Fable 5, as a Mythos-class model, carries mandatory 30-day data retention for safety monitoring — and standard Zero Data Retention agreements do not cover Mythos-class models. Anthropic states the retained data is not used for training or non-safety purposes, that human access is logged, and that deletion happens after 30 days in nearly all cases. That’s a well-documented constraint, but a constraint nonetheless; the full picture is in Fable 5’s mandatory 30-day retention policy. Fable 5 also routes cybersecurity- and biology-related queries to Opus 4.8 automatically via a safety classifier — relevant if your workload touches those domains.

K3’s side of this comparison is a blank, and we’re keeping it that way deliberately: Moonshot has not yet published a standing data-retention policy for K3 at launch. A Kimi Enterprise option with data-privacy features is offered for team deployment, but there is no public retention commitment to cite. Depending on your compliance posture, an unpublished policy can be worse than a documented 30-day one — you can’t take an unknown to a security review. For regulated workloads, the practical answer in July 2026 is that neither model is fully comfortable, and the self-hosted path (K3’s weights, after July 27, on your own hardware) is the only route to full data control on this matchup.

The most useful paragraph in the K3 launch blog is the one most launch posts never write. Moonshot lists its own model’s limitations, by name, against named competitors — starting with user experience.

That is vendor self-critique of a rare kind: conceding UX inferiority to two specific named competitors in the same document as your own benchmark wins. The blog names two further launch limitations worth engineering around. First, thinking-history sensitivity: K3’s quality degrades if your harness doesn’t preserve all thinking content, or when a session switches over from another model — agent frameworks that trim reasoning traces to save context will hit this. Second, what Moonshot calls excessive proactiveness: K3 may make unexpected decisions on the user’s behalf when it hits minor issues or ambiguous intent mid-task — helpful in a demo, hazardous in an unattended pipeline.

Add the launch-configuration constraints — thinking effort is max-only for now (lower tiers are promised later), sampling is fixed at temperature 1.0 / top_p 0.95 with no user override, and output defaults to 131,072 tokens — and a fair summary is that K3 ships as a powerful but opinionated instrument, while Fable 5 exposes the more tunable production surface (full effort range, adaptive effort, standard sampling controls). Projecting forward: if Moonshot lands the promised effort tiers and the July 27 weights with a permissive license, the tunability gap narrows fast, and the ecosystem pressure on closed pricing gets real. Until those ship, they’re roadmap, not product.

An 8–6 vendor-reported split with a 3.3x price gap doesn’t produce a winner; it produces a routing table. Here’s ours, tied to the specific rows and facts above.

If your team wants this decision made against your actual workloads rather than vendor charts — benchmark harness, cost model, and a routing policy across K3, Fable 5, and the rest of the frontier — that comparative eval is exactly where our AI transformation engagements start.