Kimi K3 vs Claude Fable 5: Which AI Model Is Better for Coding and Agentic Work?

Kimi K3 and Claude Fable 5 are built for the same difficult category of work: large codebases, long-running agents, complex research, tool-heavy workflows and tasks that cannot be solved reliably in one short response.

They reach that goal through very different product strategies.

Anthropic offers Claude Fable 5 as its most capable widely released model. It is proprietary, expensive, available across several major cloud platforms and surrounded by a mature agent ecosystem led by Claude Code. Its adaptive reasoning is always enabled, and Anthropic positions it for the most demanding coding and knowledge-work tasks.

Moonshot AI offers Kimi K3 as a 2.8-trillion-parameter Mixture-of-Experts model with a one-million-token context window, native image and video understanding, an OpenAI-compatible API and pricing that is dramatically lower than Fable 5. Moonshot has also promised to release the full model weights by July 27, 2026.

The headline result is not that one model wins everything.

Claude Fable 5 remains the stronger overall model and the safer default when maximum quality, mature tooling and predictable long-horizon behavior matter most. Kimi K3 is the more disruptive choice when cost, long context, visual coding, API compatibility and future control over model weights matter more.

Quick comparison

Category	Kimi K3	Claude Fable 5
Developer	Moonshot AI	Anthropic
API model ID	kimi-k3	claude-fable-5
Public launch	July 2026	June 9, 2026; access restored July 1
Model size	2.8 trillion parameters	Not disclosed
Architecture	MoE, KDA, AttnRes, Stable LatentMoE	Not disclosed
Active experts	16 of 896	Not disclosed
Context window	1 million tokens	1 million tokens
Maximum output	Approximately 131k tokens by default/API configuration	Up to 128k tokens
Text input	Yes	Yes
Image input	Yes	Yes
Video input	Yes	No direct video modality listed
Output	Text	Text
Reasoning	Always on; currently max only	Always-on adaptive thinking; low, medium, high, xhigh, max
Standard input price	$3 / 1M tokens	$10 / 1M tokens
Cached-input price	$0.30 / 1M tokens	$1 / 1M tokens
Output price	$15 / 1M tokens	$50 / 1M tokens
Batch price	Not compared here	$5 input / $25 output per 1M
Tool calling	Yes	Yes
Structured output	Yes	Yes
First-party coding agent	Kimi Code	Claude Code
API style	OpenAI-compatible Chat Completions	Anthropic Messages API
Public weights	Announced, not yet released at verification	No
Self-hosting	Planned, but extremely hardware-intensive	No
Data retention	Depends on Kimi product and contract	30 days for Fable; no ZDR
Safety fallback	No equivalent documented model fallback	Can fall back to Claude Opus 4.8

The table reflects official product documentation available on July 17, 2026. Features with similar names may behave differently, and benchmark results should not be treated as interchangeable across different agent harnesses.

What is Kimi K3?

Kimi K3 is Moonshot AI's most capable model to date. It contains 2.8 trillion parameters and uses a sparse Mixture-of-Experts architecture. According to Moonshot, the model activates 16 of 896 experts during inference rather than using the entire network for every token.

Its architecture combines several technologies developed by Moonshot:

• Kimi Delta Attention, designed to improve attention efficiency over long sequences;
• Attention Residuals, which selectively retrieve representations across model depth;
• Stable LatentMoE, used to scale sparse expert routing;
• quantization-aware training using MXFP4 weights and MXFP8 activations.

Moonshot says these changes produce roughly 2.5 times the scaling efficiency of Kimi K2. That figure is a vendor claim and should be treated as such until the full technical report and reproducible implementation details are available.

K3 is designed for:

• long-running software-engineering tasks;
• navigation and modification of very large repositories;
• terminal and tool orchestration;
• visual development using screenshots;
• frontend, game-development and CAD workflows;
• research that combines literature, code and data;
• document, image and video analysis;
• multi-agent knowledge work.

The model is available through Kimi, Kimi Work, Kimi Code and the Kimi API.

What is Claude Fable 5?

Claude Fable 5 is Anthropic's most capable widely released model. Anthropic describes it as a Mythos-class model built for demanding reasoning, software engineering and long-horizon agentic work.

Fable 5 and the limited-access Claude Mythos 5 share the same underlying capabilities and pricing. The major difference is that Fable includes production safety classifiers. When a request triggers one of those classifiers, the API may return a refusal or route the task to Claude Opus 4.8 when fallback is configured.

Claude Fable 5 supports:

• a one-million-token context window;
• up to 128,000 output tokens;
• text and image input;
• adaptive thinking that is always enabled;
• effort control from low to max;
• tool use and programmatic tool calling;
• code execution;
• memory;
• task budgets;
• context editing and compaction;
• vision;
• server-side and client-side fallback handling.

It is available through the Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry.

Benchmark warning: the agent harness is part of the result

A coding benchmark score is not produced by the model alone.

The final result also depends on:

• the coding agent or harness;
• the tools available to the model;
• how repository context is selected;
• whether previous reasoning blocks are preserved;
• the time and token budget;
• retry rules;
• fallback behavior;
• the number of evaluation runs;
• the grader and test environment.

Moonshot's K3 launch table mixes results obtained through Kimi Code, Claude Code, Codex, Terminus and other official benchmark harnesses. Moonshot documents these differences, but they still mean that the table is not a perfectly controlled model-only comparison.

Fable adds another complication: some published results include potential fallback to Claude Opus 4.8 when a safety classifier refuses a request. In those cases, the measured product is effectively Fable plus its fallback policy, not always Fable alone.

The correct interpretation is therefore:

Benchmark results show how a model-and-agent configuration performed under a stated methodology. They do not guarantee that the same model will produce the same result in your repository, toolchain or security environment.

Independent overall evaluations

Independent evaluations currently place Fable slightly ahead overall, but Kimi is close enough to be a credible frontier competitor.

Evaluation	Kimi K3	Claude Fable 5	Leader
Artificial Analysis Intelligence Index	57	60	Fable
Vals Index	74.7%	75.1%	Fable
Arena Text Overall, July 16 snapshot	Below Fable	Ranked #1	Fable
Arena Instruction Following, July 16 snapshot	Below Fable	Ranked #1	Fable
Arena WebDev Overall, July 16 snapshot	Preliminary rank #1	Close behind	Kimi

Artificial Analysis also reports that Kimi generated approximately 130 million output tokens across its Intelligence Index evaluation, compared with approximately 87 million for Fable. That matters because Kimi's lower per-token price can be partly offset when it uses substantially more tokens to solve the same class of tasks.

The Vals Index is especially notable because the difference is tiny: roughly four-tenths of a percentage point. This suggests that Kimi is not merely a low-cost alternative. On some economically relevant finance and coding tasks, it is operating very close to Fable.

Arena results show a useful split. Fable performs better in broad text preference and instruction-following evaluations, while Kimi reached the top of the WebDev leaderboard in a preliminary July 16 snapshot. This supports a practical pattern seen in the vendor results: Fable is more consistently strong overall, while Kimi can be exceptional in visual and implementation-heavy development workflows.

Coding benchmarks

Moonshot published the following direct comparison in the Kimi K3 launch materials. All models were run at their highest stated reasoning effort, but the harnesses differed by benchmark.

Coding benchmark	Kimi K3	Claude Fable 5	Leader
DeepSWE	67.5	70.0	Fable
ProgramBench	77.8	76.8	Kimi
Terminal-Bench 2.1	88.3	84.6	Kimi
FrontierSWE	81.2	86.6	Fable
SWE Marathon	42.0	35.0	Kimi
PostTrain Bench	36.6	41.4	Fable
MLS Bench	48.3	49.9	Fable
Kimi Code Bench 2.0, internal	72.9	76.9	Fable

This is not a one-sided result.

Where Fable looks stronger

Fable leads in:

• DeepSWE, focused on difficult repository-level software tasks;
• FrontierSWE, where the gap is more than five points;
• PostTrain Bench;
• MLS Bench;
• Moonshot's own internal Kimi Code Bench 2.0.

The internal Kimi Code Bench result is particularly interesting because Moonshot reports Fable ahead on a benchmark developed by the Kimi team. It is evidence against interpreting the launch material as a simple marketing table designed to make K3 win every row.

Anthropic separately reports very strong results for Fable on SWE-bench Pro and Terminal-Bench. Its launch table lists 80.3% on SWE-bench Pro and 88.0% on Terminal-Bench 2.1. Those figures should not be directly merged with Moonshot's table because the methodologies differ, but they reinforce the conclusion that Fable is an exceptionally capable repository and terminal agent.

Where Kimi looks stronger

Kimi leads in:

• Terminal-Bench 2.1 within Moonshot's comparison;
• ProgramBench by a small margin;
• SWE Marathon by seven points.

SWE Marathon is relevant because it is intended to capture sustained software-engineering work rather than one isolated patch. Kimi's lead supports Moonshot's positioning of K3 as a model trained for long, difficult trajectories.

Moonshot also reports that K3 performed competitively with Fable in a 24-hour GPU-kernel optimization test. The task involved profiling, rewriting and benchmarking kernels across several architectures and hardware targets. This was a vendor-designed case study rather than an independent public leaderboard, but it demonstrates the kind of long-horizon engineering workload K3 is designed to pursue.

What the coding results mean in practice

Fable appears more dependable when the task is:

• a difficult repository-level issue;
• a broad refactor with many hidden constraints;
• a migration that requires careful verification;
• a long project where instruction following and judgment matter more than raw throughput;
• work already embedded in Claude Code.

Kimi becomes particularly attractive when the task is:

• terminal-heavy and iterative;
• visually evaluated through screenshots;
• frontend, game or CAD development;
• a long experiment with many retries;
• too expensive to run repeatedly on Fable;
• performed through an existing OpenAI-compatible integration.

Agentic and knowledge-work benchmarks

The same Moonshot table gives a mixed but generally Fable-leaning result for broader agentic work.

Agent benchmark	Kimi K3	Claude Fable 5	Leader
GDPval-AA v2, Elo	1668	1760	Fable
BrowseComp	91.2	88.0	Kimi
Toolathlon-Verified	73.2	77.9	Fable
MCP Atlas	84.2	84.7	Fable
AutomationBench	30.8	29.1	Kimi
JobBench	52.9	57.4	Fable
AA-Briefcase, Elo	1548	1583	Fable
APEX-Agents	37.6	43.3	Fable
SpreadsheetBench 2	34.8	34.7	Practical tie

Fable wins more of these rows, especially on professional work represented by GDPval-AA, JobBench, AA-Briefcase and APEX-Agents.

Kimi performs very well in BrowseComp and AutomationBench. BrowseComp measures difficult information retrieval where agents must search, combine evidence and persist through long chains of exploration. Kimi's result is consistent with Moonshot's demonstrations of large research workflows involving thousands of searches and document operations.

The overall conclusion is:

Fable has the stronger general agent profile. Kimi is highly competitive and can lead in search-intensive, automation-heavy and implementation-oriented workflows.

Reasoning control

Both models reason on every request, but Fable gives developers much more control.

Kimi K3

Kimi K3 currently supports:

reasoning_effort="max"

Thinking is always enabled. Moonshot says low and high modes will be added later, but they were not available at the time of verification.

This makes K3 simple to configure for difficult work, but inefficient for routine tasks. A short extraction or classification request may still receive the behavior of a maximum-effort reasoning model.

K3 also requires the complete assistant message, including reasoning-related fields, to be passed back in multi-turn conversations. Moonshot warns that quality can become unstable when a harness drops that history or switches to K3 in the middle of a session.

Claude Fable 5

Fable uses adaptive thinking on every request and does not support disabling it. Developers can control the total depth and token use with:

low
medium
high
xhigh
max

The default is high. Anthropic recommends:

• low for routine, high-volume or subagent work;
• medium for a balance of cost and capability;
• high for difficult coding and reasoning;
• xhigh for long-running agentic work;
• max for the hardest capability-sensitive tasks.

This is a meaningful production advantage. A team can use one Fable endpoint for both routine subagents and frontier tasks, adjusting effort dynamically without changing models.

Tool use and agent ecosystems

Kimi K3 and Kimi Code

Kimi Code can:

• read and edit files;
• run shell commands;
• search a codebase;
• fetch web pages;
• use tools;
• spawn subagents for parallel work.

The Kimi API supports:

• function calling;
• tool_choice;
• dynamic tool loading;
• structured output through JSON Schema;
• streaming;
• automatic context caching;
• partial completion;
• OpenAI SDK compatibility.

The OpenAI-compatible API is a major practical advantage. Existing applications can often test Kimi by changing the API key, base URL and model name rather than rewriting their entire client layer.

Moonshot does warn that its official web-search functionality was being updated at the time of verification and was not recommended for near-term production use.

Claude Fable 5 and Claude Code

Claude's platform currently offers a broader documented set of production tools and context-management features:

• Claude Code;
• code execution;
• web search and web fetch;
• memory;
• programmatic tool calling;
• task budgets;
• compaction;
• context editing and tool-result clearing;
• strict tool use;
• MCP support;
• server-side fallback;
• managed-agent and orchestration patterns.

Claude Code also exposes multi-agent workflows. Its ultracode mode is not a separate API effort level: Anthropic describes it as xhigh effort combined with standing permission to launch multi-agent workflows.

For teams that need a mature, documented agent platform rather than only a strong model endpoint, Claude currently has the advantage.

Context, caching and multimodality

Both models support approximately one million tokens of context.

That is enough to hold:

• a large repository snapshot;
• hundreds of pages of documentation;
• a long agent history;
• multiple reports and datasets;
• a large collection of tool results.

Neither model should be assumed to use the entire context perfectly. Retrieval accuracy, instruction retention and resistance to conflicting information still need to be tested on the actual workload.

Kimi

Kimi provides automatic prefix caching. Developers do not need to create a cache identifier or manually set a TTL. Keeping the long prefix unchanged allows later requests to attempt a cache hit automatically.

The official API price for a cache-hit input token is $0.30 per million.

Kimi supports direct image and video input. This makes it especially useful for:

• screenshot-driven frontend correction;
• video analysis;
• visual QA;
• game-development feedback loops;
• reviewing animations and motion design.

Fable

Fable supports prompt caching with explicit or automatic cache control. Official pricing is:

• $12.50 per million tokens for a five-minute cache write;
• $20 per million for a one-hour cache write;
• $1 per million for cache reads.

Fable accepts text and images. Anthropic does not list direct video input for the model. Video workflows therefore normally require extracted frames, transcripts or another preprocessing layer.

API pricing

Token category	Kimi K3	Claude Fable 5
Uncached input	$3 / 1M	$10 / 1M
Cached input read	$0.30 / 1M	$1 / 1M
Five-minute cache write	Automatic caching; no separate K3 write price listed	$12.50 / 1M
One-hour cache write	Not listed as an equivalent manual tier	$20 / 1M
Output	$15 / 1M	$50 / 1M
Batch input	Not included in this comparison	$5 / 1M
Batch output	Not included in this comparison	$25 / 1M

At list price, Kimi is 70% cheaper for uncached input and 70% cheaper for output.

Cost example 1: repository task

Assumptions:

• 100,000 uncached input tokens;
• 10,000 output tokens.

Model	Input	Output	Total
Kimi K3	$0.30	$0.15	$0.45
Claude Fable 5	$1.00	$0.50	$1.50

Kimi is 70% cheaper in this example.

Cost example 2: large repeated context with a cache hit

Assumptions:

• 500,000 cached input tokens;
• 20,000 output tokens;
• cache-write cost excluded because it depends on lifecycle and reuse.

Model	Cached input	Output	Total
Kimi K3	$0.15	$0.30	$0.45
Claude Fable 5	$0.50	$1.00	$1.50

Again, Kimi is 70% cheaper at the listed read and output rates.

Price per token is not cost per solved task

Artificial Analysis observed that Kimi generated around 130 million output tokens during its Intelligence Index evaluation, versus around 87 million for Fable.

This does not mean Kimi will always use 49% more tokens in production. It does mean that a realistic cost evaluation should measure:

• total input tokens;
• cache-hit rate;
• reasoning and output tokens;
• retries;
• tool calls;
• failed runs;
• human review time;
• success rate.

A model that costs one-third as much per token may not cost one-third as much per accepted result if it needs longer trajectories or more retries.

Open weights and self-hosting

Kimi's planned weight release is one of the largest strategic differences between the models.

Moonshot says the full K3 weights will be released by July 27, 2026. At the time this article was verified, the weights, final license and full technical report were not yet publicly available.

The eventual release could allow organizations to:

• inspect and modify the model;
• run it in controlled infrastructure;
• avoid permanent dependence on one API provider;
• fine-tune or quantize it;
• integrate it with private inference stacks.

However, "open weights" does not mean "easy to run."

Moonshot recommends supernode deployments with 64 or more accelerators. A 2.8-trillion-parameter sparse model remains an enormous infrastructure project even when only part of the network is active for each token.

For most companies, Kimi will remain an API product. The open-weight option is more relevant to hyperscalers, national infrastructure providers, major AI labs and organizations with specialized inference hardware.

Fable is fully proprietary. Anthropic has not disclosed its parameter count or architecture, and there is no self-hosting path.

Safety, refusals and data retention

Claude Fable 5

Fable includes safety classifiers for sensitive areas such as cybersecurity and biology. Anthropic says they trigger in fewer than 5% of sessions on average, but also acknowledges that harmless requests can be caught.

When a request is refused:

• the API can return HTTP 200 with stop_reason: "refusal";
• the response identifies the classifier involved;
• applications can retry through server-side, client-side or manual fallback;
• a configured fallback can use Claude Opus 4.8;
• requests refused before output generation are not billed;
• Anthropic provides a fallback credit mechanism to avoid paying the prompt-cache switching cost twice.

This architecture is safer for broad public deployment, but it can reduce reproducibility. Two apparently similar coding tasks may be handled by different underlying models when one triggers a classifier.

Fable requires 30-day data retention and is not available under Zero Data Retention. That can be a decisive limitation for regulated or highly confidential workloads.

Kimi K3

Kimi does not document an equivalent automatic fallback to another model in the K3 API.

Moonshot does document two practical behavioral limitations:

1. Sensitivity to thinking history. If the agent does not preserve the complete assistant message, or switches models mid-session, quality can become unstable.
2. Excessive proactiveness. K3 may make unexpected decisions when instructions are ambiguous because it is trained to pursue long, difficult tasks autonomously.

Moonshot recommends explicit constraints in the system prompt or AGENTS.md when an agent must remain within strict boundaries.

Data handling for Kimi should be evaluated against the exact product and contract being used. Teams processing confidential information should obtain written terms covering retention, training use, data location, deletion and incident handling.

Which model should you choose?

Choose Kimi K3 when:

• API cost is a primary constraint;
• you need to run many long experiments or retries;
• the workload uses very large contexts;
• the agent works with screenshots or video;
• frontend, game development or visual iteration is central;
• you already use an OpenAI-compatible client;
• future access to open weights matters;
• you can validate agent behavior with tests and sandboxes;
• you are prepared to manage a younger ecosystem.

Choose Claude Fable 5 when:

• maximum general capability matters more than token price;
• difficult repository-level work is the main workload;
• instruction following and judgment must remain strong over long sessions;
• your team already uses Claude Code;
• you need mature context-management and agent tools;
• multi-cloud availability matters;
• you want to tune effort dynamically from low to max;
• safety fallback is acceptable or desirable;
• 30-day data retention does not block the project.

Use both when:

A hybrid routing strategy may be the best engineering decision.

Kimi can handle:

• initial implementation attempts;
• visual frontend work;
• broad repository exploration;
• high-volume code generation;
• long-context preprocessing;
• parallel candidate solutions.

Fable can handle:

• final architectural review;
• difficult bug escalation;
• verification of risky changes;
• complex migrations;
• instruction-sensitive work;
• the hardest long-horizon tasks.

This approach uses Kimi's price advantage without treating it as a universal replacement for Fable.

Verdict

Claude Fable 5 is the stronger overall model for coding and agentic work.

It leads the independent Artificial Analysis Intelligence Index, narrowly leads the Vals Index, ranks at the top of broad text and instruction-following arenas, and wins more of Moonshot's published general-agent benchmarks. Its reasoning controls, Claude Code ecosystem, context-management features and multi-cloud availability make it the safer default for teams that prioritize capability and operational maturity.

Kimi K3 is the stronger value proposition.

It reaches near-frontier performance at roughly 30% of Fable's standard token price, leads several coding and automation benchmarks, supports direct video input, integrates through the OpenAI SDK and may soon offer full downloadable weights. It is especially compelling for visual coding, long experiments, large contexts and workloads where Fable's cost would prevent enough iteration.

The honest conclusion is:

Choose Claude Fable 5 for the highest probability of success on the hardest, most instruction-sensitive engineering tasks. Choose Kimi K3 when you need frontier-level capability at a dramatically lower price and can compensate with evaluation, sandboxing and stronger workflow controls.

Do not select either model from one leaderboard row. Run both against a private evaluation set built from real repository issues, expected tool calls, security constraints and acceptable cost per completed task.

Sources

• Moonshot AI, Kimi K3: Open Frontier Intelligence
• Kimi API Platform, Kimi K3 quickstart and API features
• Kimi API Platform, Kimi K3 pricing
• Moonshot AI, Kimi Code
• Anthropic, Claude Fable 5 and Claude Mythos 5
• Claude Platform Docs, Introducing Claude Fable 5 and Claude Mythos 5
• Claude Platform Docs, Effort controls
• Claude Platform Docs, API pricing
• Claude Platform Docs, Prompt caching
• Anthropic, Claude Fable 5 and Claude Mythos 5 System Card
• Artificial Analysis, Kimi K3 evaluation
• Artificial Analysis, Claude Fable 5 evaluation
• Vals AI, Vals Index
• Arena, WebDev leaderboard
• Arena, Text leaderboard