GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Kimi K2.7-Code Ships Open-Source: 30% Fewer Thinking Tokens, 81% MCP Mark Verified, $4/M Output

Moonshot AI released Kimi K2.7-Code on June 12, 2026 as an open-source successor to K2.6, the model that held open-source SWE-bench Pro SOTA at 58.6% from April. Weights are on Hugging Face under a Modified MIT license. A high-speed variant with 5-6x faster output, capped at approximately 180 tokens per second, launched June 15 at double the standard API price.

Benchmark Numbers

All figures are Moonshot’s own benchmarks. K2.7-Code ran in Kimi Code CLI; comparisons against GPT-5.5 used Codex xhigh and Claude Opus 4.8 used Claude Code xhigh.

BenchmarkK2.6K2.7-CodeGPT-5.5Opus 4.8
Kimi Code Bench v250.962.069.067.4
Program Bench48.353.669.163.8
MLS Bench Lite26.735.135.542.8
Kimi Claw 24/7 Bench42.946.952.850.4
MCP Atlas69.476.079.481.3
MCP Mark Verified72.881.192.976.4

K2.7-Code beats Opus 4.8 on MCP Mark Verified, 81.1 vs 76.4. MCP Mark Verified tests agents across five real-world software environments: Notion, GitHub, file systems, Postgres, and Playwright browser automation. GPT-5.5 leads at 92.9.

On Program Bench — where agents must reproduce a compiled binary’s behavior without source code — K2.7-Code trails GPT-5.5 by 15.5 points. That is the starkest gap in the table.

The Operational Win

The headline that matters most for production workloads is not the benchmark delta. It is the 30% reduction in reasoning token usage versus K2.6. For teams running K2.7-Code in agentic loops with extended thinking enabled, 30% fewer thinking tokens translates directly to lower inference cost and lower latency per turn. Moonshot’s number — unverified independently — is credible enough to test before committing.

Pricing

VariantInputOutputCache input
Standard$0.95/M$4.00/M$0.19/M
High-speed (June 15)~$1.90/M~$8.00/M—

Output at $4/M is 12x cheaper than Claude Fable 5 at $50/M, and 6x cheaper than Opus 4.8 at $25/M. The standard pricing is unchanged from K2.6.

The Caveat

Every benchmark row above is from Moonshot’s own test suites. K2.6 scored 24% on DeepSWE, an independent benchmark with a 70-point spread across models — tied with GPT-5.4-mini. That is a 35-point gap versus Moonshot’s vendor-run Kimi Code Bench v2 score for K2.6. Developer Sugumaran Balasubramaniyan challenged Moonshot publicly on X to submit K2.7-Code to DeepSWE; as of publication, Moonshot has not done so. No third-party SWE-bench Pro run for K2.7-Code exists yet.

That gap does not mean the model is useless — K2.6 still ran production workloads for teams that tested it. It means the benchmark numbers should be treated as directional until independent results land.

The token efficiency improvement and the open-weight, low-cost positioning are the real reasons to evaluate K2.7-Code against your own agentic workloads. Run the high-speed variant if throughput matters; use standard if you want the cheaper thinking budget.