Kimi K3 vs Claude Fable 5 vs GPT-5.6 Sol: What the Benchmarks Actually Show

Kimi K3 is not merely “good for an open model.” On Moonshot AI’s launch evaluations, it is close enough to OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable 5 to force a more uncomfortable question: how much proprietary-model premium is still justified when an open-weight contender can win meaningful coding, browsing, automation, spreadsheet, and vision tests?

The excitement is deserved. Kimi K3 is a 2.8-trillion-parameter mixture-of-experts model with native vision and a 1-million-token context window. Moonshot says it activates 16 of 896 experts and achieves roughly 2.5 times better scaling efficiency than Kimi K2. It is available now through Kimi products and the API, priced at $3 per million uncached input tokens and $15 per million output tokens.

But one correction matters on July 17, 2026: K3 is not yet downloadable as a fully open-weight release. Moonshot says the full weights will arrive by July 27. Until that happens—and until the license and independent deployment experience can be inspected—the accurate description is an API-available frontier model with a public open-weight commitment, not an already delivered open-source package.

Quick verdict: Kimi K3 belongs in the frontier comparison

Kimi K3 vs Claude Fable 5 vs GPT-5.6 Sol at a glance

The honest conclusion is not that K3 defeats both proprietary leaders everywhere. It does not. The conclusion is more important: K3 now belongs in the same evaluation set. It wins enough practical categories—and gets close enough in others—that dismissing it as a budget alternative is no longer credible.

The coding benchmarks that actually matter

Repository-scale software work is not one skill, so no single coding score settles the comparison. Moonshot’s table gives K3 a 67.5 on DeepSWE, behind GPT-5.6 Sol at 73.0 and Fable 5 at 70.0. On Terminal-Bench 2.1, which tests practical terminal execution, K3 scores 88.3 versus Sol’s 88.8 and Fable’s 84.6. A half-point gap from Sol is effectively a “test it yourself” result, not a reason to reject K3.

K3 pulls ahead on several longer or more varied engineering tests. It scores 77.8 on Program Bench, narrowly above Sol at 77.6 and Fable at 76.8. On SWE Marathon it reaches 42.0, ahead of Sol at 39.0 and Fable at 35.0. FrontierSWE flips the ordering: Fable leads at 86.6, K3 follows at 81.2, and Sol records 71.3 under its Codex harness.

Selected vendor-reported coding results

These numbers need context. K3 uses KimiCode on some tests, Claude models use Claude Code or Terminus, and GPT-5.6 uses Codex. A model-plus-harness score measures the system developers actually use, but it is not a laboratory-isolated comparison of model weights. Moonshot also notes fallback behavior in some Fable 5 runs. Treat the table as strong evidence that K3 is competitive, then run the same repository, tools, acceptance tests, and time budget across all three systems.

K3’s agent results may matter more than its coding wins

The most consequential K3 results are outside conventional code repair. BrowseComp measures difficult web research where the agent must find obscure facts across many sources. Moonshot reports 91.2 for K3 with its context-compaction setup, compared with 90.4 for GPT-5.6 Sol and 88.0 for Claude Fable 5 in Moonshot's table. OpenAI separately reports Sol at 92.2 under its own multi-agent configuration. Those figures are not directly interchangeable: context management, agent count, tool harness, and reasoning budget all change the system being measured. The defensible conclusion is that K3 is competitive at the frontier of browsing, not that one vendor table proves a universal winner. On the public 600-task subset of AutomationBench, Moonshot reports K3 at 30.8, Sol at 29.7, and Fable at 29.1.

MCP Atlas is even tighter: K3 scores 84.2, Fable 84.7, and Sol 83.6. That is exactly what frontier competition looks like—small differences across a benchmark that tests tool use through Model Context Protocol servers. K3 also leads SpreadsheetBench 2 by a hair at 34.8 versus Fable’s 34.7 and Sol’s 32.4, while Fable retains a clearer advantage on OfficeQA Pro at 69.9 versus K3’s 63.3.

This profile makes K3 unusually interesting for research agents, coding agents, spreadsheet work, browser automation, and multimodal production. It is not simply a reasoning model with a cheap endpoint. Moonshot trained and productized it around long-running, tool-using work—the exact category where model choice increasingly affects whether an agent completes the task or merely discusses it.

Where Fable 5 and GPT-5.6 Sol are still better

Claude Fable 5 remains the strongest option in several demanding categories. It leads K3 on FrontierSWE, PostTrain Bench, MLS Bench, OfficeQA Pro, Toolathlon-Verified, Job Bench, APEX-Agents, and HLE with tools. On HLE-Full, Fable posts 53.3 versus K3’s 43.5 and Sol’s 44.5. That is a material gap on hard, broad knowledge and reasoning—not benchmark noise.

GPT-5.6 Sol has its own clear advantages. Moonshot’s comparison lists Sol at 73.0 on DeepSWE and 88.8 on Terminal-Bench 2.1, both above K3; OpenAI’s own table reports 72.7 and 88.8 respectively under its published setup. Sol also leads GPQA Diamond at 94.1 versus K3’s 93.5 in Moonshot’s comparison and wins several multimodal tests, including MMMU-Pro and MathVision. OpenAI’s broader release data also shows strengths K3’s launch table does not fully answer, including computer use, cybersecurity, scientific work, long-context retrieval, and production efficiency across the Sol, Terra, and Luna tiers.

Moonshot is unusually direct about K3’s current limitations. The company says quality can become unstable if a harness does not preserve thinking history, warns that the model may act too proactively when intent is ambiguous, and acknowledges a noticeable user-experience gap versus Fable 5 and GPT-5.6 Sol. Those are not footnotes for agent builders. They affect approvals, state handling, prompts, rollback behavior, and whether a system can safely operate with limited supervision.

Price and deployment can outweigh a two-point benchmark gap

At published API rates, Kimi K3 costs $3 per million cache-miss input tokens and $15 per million output tokens, with cache-hit input at $0.30. GPT-5.6 Sol is $5 input and $30 output. Claude Fable 5 is $10 input and $50 output. On list price alone, K3 is materially cheaper than both frontier proprietary options.

List price is not cost per completed task. A model that needs more retries, longer reasoning traces, extra agents, or heavier human correction can erase a lower token rate. Conversely, K3’s official API says more than 90% of coding-workload input receives the cache-hit rate, which could matter for large repositories with repeated context. A serious evaluation should record total input and output tokens, wall-clock time, tool calls, retries, reviewer minutes, and successful completions. That is the economic benchmark that survives contact with production.

The open-weight promise is the real disruption

If Moonshot releases the complete weights as promised, K3 changes the decision from “which API is smartest?” to “which layer do we want to own?” Downloadable weights can let organizations choose infrastructure, isolate sensitive workloads, inspect behavior, adapt the model, and avoid making every production workflow dependent on one vendor endpoint.

That does not mean K3 will be easy or cheap to self-host. A 2.8-trillion-parameter model that activates 16 of 896 experts is an infrastructure project. Moonshot recommends supernode configurations with 64 or more accelerators, and its efficient official API benefits from Mooncake caching and a disaggregated serving architecture. “Open weights” creates control and competition; it does not magically turn frontier inference into a single-GPU workload.

The distinction between open source and open weights matters too. A truly useful release includes accessible weights, a workable license, technical documentation, compatible inference code, and enough operational detail for independent reproduction. As of July 17, the weights and technical report are still forthcoming. Buyers should celebrate the direction without treating a dated promise as a completed artifact.

What teams should do next

Put Kimi K3 into the same bake-off as Fable 5 and GPT-5.6 Sol, but evaluate the workflow rather than copying the launch leaderboard. Use a private set of representative repository changes, research questions, spreadsheets, browser tasks, and multimodal inputs. Give every system equivalent tools, time limits, acceptance tests, and human-review rules. Measure completion rate, correction burden, latency, token cost, unsafe actions, and recovery after tool failure.

If K3 stays within a few points of the proprietary leaders on your work, the open-weight roadmap may outweigh a small quality difference. If your tasks depend on the hardest broad reasoning, polished interaction, conservative autonomy, or vendor-supported enterprise controls today, Fable 5 or GPT-5.6 Sol may still be the safer deployment choice.

The headline is not that Kimi K3 has permanently won the model race one day after launch. It is that the frontier now has a credible open-weight challenger before the weights are even available. That is an extraordinary result—and if Moonshot delivers the complete release on July 27, it may be the most important model-economics event of 2026.