Kimi K3 ranks #5 on Artificial Analysis as engineers dispute coding cost

• Kimi K3 shipped as a 2.8T-parameter, 1M-context, native-vision flagship, and Kimi says full weights arrive by July 27 in its launch post.
• The coding story split immediately: Artificial Analysis ranked Kimi K3 joint #5 on its Coding Agent Index, while AlphaSignalAI put it last of seven models on a bug-fix repair harness.
• Frontend is K3's cleanest win: Arena put it #1 in Frontend Code Arena at 1679 points, with a 17-place jump over Kimi K2.6.
• The cost debate is about tokens and time, not list price: Theo argued GPT-5.6 Sol uses about half as many tokens and runs about 2x faster, while Artificial Analysis measured K3 at $0.94 per Intelligence Index task.
• Several eval gaps remain: ARC Prize said it had not tested Kimi K3 because it lacked a zero-data-retention agreement, and one cyber-risk thread noted there was no CyberGym score in the launch benchmarks.

Moonshot's tech blog has the fun stuff: K3 built MiniTriton, redesigned a production-scale training kernel in a 15-hour run, and claims a 48-hour chip-design case study. The Kimi API quickstart also hides a practical gotcha: reasoning_effort currently supports only max, with more levels coming later. The K3 pricing docs list $0.30 cache-hit input, $3 input, and $15 output per 1M tokens.

Kimi K3 is live on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, with open weights promised by July 27. Moonshot says the model uses max thinking effort by default at launch, with low- and high-effort modes coming later.

The useful API facts:

• Model slug: kimi-k3, per the Kimi API quickstart.
• Context: 1,048,576 tokens, per the K3 pricing page.
• Pricing: $0.30 per 1M cached input tokens, $3 per 1M input tokens, $15 per 1M output tokens.
• Reasoning: always on, currently max only.
• Output cap: max_completion_tokens defaults to 131,072 and can be set up to 1,048,576, according to the quickstart.
• Sampling knobs: temperature, top-p, penalties, and n are fixed in the docs.
• Vision: public image URLs are not supported in the API docs; images use base64 or ms:// file references.

On coding agents, Kimi K3 landed near the frontier but below the top closed models. Artificial Analysis scored Kimi K3 in Kimi Code CLI at 57, joint #5 on its Coding Agent Index, behind GPT-5.6 Sol max at 61 and Fable 5 max at 59.

Artificial Analysis broke the coding score into three evals:

• Terminal-Bench v2: 84%.
• DeepSWE: 64%.
• SWE-Atlas-QnA: 23%.
• Average task cost: $3.18.
• Cost comparison: 55% cheaper than GPT-5.6 Sol max, 73% cheaper than Fable 5 max, and 59% cheaper than Opus 4.8 max, per the same report.

Outside AA's coding index, DataCurve put Kimi K3 at #3 on DeepSWE and called it the first open-weights model with frontier-level software-engineering performance. Together AI reported roughly Fable 5-level DeepSWE performance at about 35% of the price.

Frontend is where Kimi K3 looked most dominant. Arena ranked it #1 in Frontend Code Arena at 1679 points, up from Kimi K2.6's #18 position, and said K3 ranked #1 in six of seven frontend domains.

Arena's pairwise result added the cleaner comparison: Kimi K3 outputs won 76% of pairwise frontend matchups, compared with 63% for Claude Fable 5 and 58% for GPT-5.6 Sol.

The repair harness told a colder story. AlphaSignalAI ran K3 against GPT-5.6 Sol, Fable 5, Grok 4.5, Opus 4.8, GLM-5.2, and Gemini 3.1 Pro, then reported:

• Rank: last of 7.
• Result: 53 successful attempts out of 67, or 79%.
• Cost: $0.186 per successful fix.
• Wall time: 702 seconds average.
• GPT-5.6 Sol: 70 of 70.
• Grok 4.5: 99% and 46 seconds average.

AlphaSignal's own reply separated the two tasks: 1679 is frontend generation, while its 79% number is bug-fix repair in one clarification.

Kimi's $3/$15 list price makes it cheaper per token than GPT-5.6 Sol and Fable 5, but the working argument moved to completed-task cost within hours.

Artificial Analysis measured K3 at $0.94 per Intelligence Index task, close to GPT-5.6 Sol at $1.04 and about half of Opus 4.8 at $1.80 in its cost comparison. On the broader Intelligence Index, Artificial Analysis scored K3 at 57, with a 13-point gain over K2.6 and 21% fewer output tokens than K2.6.

Theo's hands-on cost model was harsher:

• K3 is roughly half the price of GPT-5.6 Sol per token.
• GPT-5.6 uses roughly half as many tokens.
• GPT-5.6 is about 2x faster in tokens per second.
• The combined loop can be about 4x faster at roughly the same price.

Theo's reply compressed the point further: Kimi uses roughly 2x the tokens per request that GPT-5.6 uses, while Fable uses about 50% more beyond Kimi.

The optimization counterargument is still alive. Elie Bakouch pointed to reasoning tokens per task as the likely bottleneck and said K3.1 may be needed for a bigger efficiency jump, while Emad Mostaque argued K3's cache behavior and future large-RAM deployments could change the serving-cost curve.

The first day exposed a serving gap. BridgeMind measured 24 tokens per second on average and nearly 6 seconds before generation started, then framed the problem as iteration latency for vibe-coding loops.

Other users hit capacity issues. Pawel Skalski said he could not get two successful requests in a row because of 429s, while synthwavedd said Moonshot did not have enough capacity and that throughput was poor even when requests succeeded.

Moonshot's own ecosystem work points at why serving K3 is nontrivial. vLLM said the Kimi team contributed a KDA prefix-caching implementation upstream because KDA breaks assumptions behind conventional prefix caching. a technical-blog excerpt said Moonshot recommends deploying K3 on supernode configurations with 64 or more accelerators for inference efficiency.

K3's technical pitch is not just scale. Kimi says the 2.8T model combines:

• Kimi Delta Attention, a hybrid linear-attention mechanism.
• Attention Residuals.
• Stable LatentMoE.
• 16 activated experts out of 896.
• About 2.5x better scaling efficiency than Kimi K2.

The launch blog's most coder-brained claim was K3 optimizing its own training stack. Kimi said K3 redesigned an AttnRes training kernel over a 15-hour run, cutting forward-plus-backward time from 283.6 ms to 114.4 ms.

Moonshot's tech blog also says K3 built MiniTriton, a compact Triton-like compiler with a tile-level IR, optimization passes, PTX code generation, and enough runtime support to train nanoGPT end to end.

Kimi K3 was available through several developer surfaces on day one or immediately after launch:

• Vercel AI Gateway lists moonshotai/kimi-k3, with 1M context and native vision in Vercel's post.
• Kimi Code added a Vercel plugin with Next.js, AI SDK, and Vercel Functions skills in Vercel's follow-up.
• Cline exposed moonshotai/kimi-k3 for CLI, VS Code, and JetBrains users in its setup post.
• OpenCode made K3 available to OpenCode Go users, with a higher limit burn rate until pricing improves in its rollout note.
• ClinePass added K3 under a subscription for discounted access to open-weight models in Cline's ClinePass post.
• OpenCode auth can select Moonshot AI and Kimi K3 from /models, according to one setup note.
• Nous added Kimi to Hermes Agent through Nous Portal, Kimi Direct, and OpenRouter in Teknium's post.

The unanswered evals are unusually important because K3's strongest public wins are jagged. ARC Prize said it could not verify Kimi K3 yet because Moonshot's API lacked a zero-data-retention agreement, and it would wait for another open-source provider.

Cyber is also unresolved. one security-focused thread noted that the launch benchmarks did not include CyberGym, while Ethan Mollick pointed to upcoming UK AISI testing after weights are out.

ProgramBench has a metric caveat. Ofir Press, a ProgramBench author, said Kimi used an average implementation percentage rather than the recommended percentage of fully implemented programs in his note, and his follow-up explained that 90% partial implementation across all programs can still mean zero fully working programs.

Artificial Analysis found one regression inside an otherwise stronger model: its AA-Omniscience note said accuracy rose from 33% to 46%, but hallucination rate also rose from 39% to 51% compared with K2.6.