Moonshot AI open-weight challenger vs Anthropic flagship · research snapshot July 18, 2026

Kimi K3 vs Claude Opus 4.8

Kimi K3 is the largest open-weight model ever released at 2.8 trillion parameters. Claude Opus 4.8 is Anthropic's production-hardened flagship. One leads on vendor-reported coding benchmarks and price; the other leads on maturity, controls, and independent verification. This comparison uses Moonshot's official launch table, Anthropic's published scores, and independent leaderboard data to help you pick the right model for real work.

Last updated July 18, 2026 · K3 weights promised by July 27 Opus 4.8: stable, shipping since May 28

88.3%
K3 Terminal-Bench 2.1 (KimiCode harness)
78.9%
Opus 4.8 Terminal-Bench 2.1 (Claude Code, independent)
$3 / $15
K3 input / output per 1M
$5 / $25
Opus 4.8 input / output per 1M
Short answer: pick Kimi K3 for high-volume coding, frontend work, and cost-sensitive agent pipelines where open weights matter. Pick Claude Opus 4.8 for production workflows that need mature controls, adjustable reasoning effort, fast mode, and a proven reliability track record. K3 wins the headline benchmarks; Opus wins the operational story.

Model cards side by side

SpecificationKimi K3Claude Opus 4.8
Provider / releaseMoonshot AI · July 16, 2026Anthropic · May 28, 2026
API model IDkimi-k3claude-opus-4-8
ArchitectureMoE · 2.8T total / 16 of 896 experts activeProprietary dense (undisclosed)
Attention mechanismKimi Delta Attention + Attention ResidualsNot publicly detailed
Context window1,048,576 tokens (1M)1M tokens
Maximum outputNot stated on launch page128K tokens
Reasoning modesMax effort at launch; lower modes plannedLow, Medium, High, Extra, Max + Fast mode (2.5× speed)
Input / outputText and native vision / textText and images / text
ToolsFunction calling, structured outputsTool calls, JSON output, MCP, computer use
Weights / licenseOpen weights (Modified MIT), promised by July 27Closed, proprietary, API-only
Speed optionNone announcedFast mode: 2.5× speed at $10/$50 per 1M

K3's architectural claim is genuinely unusual: at 2.8 trillion parameters with 16 active experts out of 896, it's the largest open-weight model ever. Its Kimi Delta Attention is a hybrid linear attention mechanism designed to keep long-context inference efficient. Opus 4.8 is a more mature, closed product — Anthropic doesn't disclose its architecture, but the model ships with adjustable reasoning effort, fast mode, and a proven deployment record since late May.

Benchmarks: K3 leads Moonshot's coding table, Opus holds on independent verification

The table below uses Moonshot's official K3 launch comparison, which reports K3 at max effort and Opus 4.8 at max effort. Harnesses differ per benchmark row — Moonshot's footnotes specify which agent scaffold was used for each model. Rows where the harness differs should not be treated as clean head-to-head results.

BenchmarkKimi K3 (max)Opus 4.8 (max)Interpretation
Terminal-Bench 2.188.384.6K3 +3.7 · K3 uses KimiCode, Opus uses Terminus 2
FrontierSWE81.266.7K3 +14.5 · K3 uses KimiCode; large gap on repo-level engineering
DeepSWE67.559.0K3 +8.5 · K3 uses KimiCode harness
Program Bench77.871.9K3 +5.9 · raw pass rate
SWE Marathon42.040.0K3 +2.0 · both use Claude Code harness; close on sustained autonomy
Kimi Code Bench 2.072.971.7K3 +1.2 · near tie on internal coding-agent eval
GDPval-AA v2 (Elo)1,6681,600K3 +68 Elo · agentic knowledge work
PostTrain Bench36.634.1K3 +2.5
MLS Bench48.342.8K3 +5.5
Selected coding benchmarks: K3 vs Opus 4.8
Higher is better. All scores from Moonshot's official K3 launch table. Harness differences noted in footnotes.
Terminal-Bench · K3
88.3
Terminal-Bench · Opus 4.8
84.6
FrontierSWE · K3
81.2
FrontierSWE · Opus 4.8
66.7
DeepSWE · K3
67.5
DeepSWE · Opus 4.8
59.0
Harness hygiene: Moonshot's table uses KimiCode for K3 on most coding rows, while Opus 4.8 scores come from Claude Code, Terminus 2, or other harnesses depending on the benchmark. An agent benchmark measures the full scaffold — prompt, tools, retries, and timeout — not just the model. The independent Terminal-Bench 2.1 leaderboard tells a more nuanced story: Opus 4.8 scores 78.9% with Claude Code (verified, public trace), while K3's 88.3% uses the KimiCode harness and has not yet appeared on the public leaderboard. Use these numbers to inform testing, not to declare a universal winner.

Where Opus 4.8 fights back: Anthropic's own published scores

Anthropic published a different set of benchmarks at Opus 4.8's launch — and on those, Opus shows clear strengths that Moonshot's table doesn't capture:

BenchmarkOpus 4.8Context
SWE-bench Verified88.6%+1.0 vs Opus 4.7; near ceiling on this benchmark
SWE-bench Pro69.2%+4.9 vs Opus 4.7; the harder, less-saturated coding set
SWE-bench Multilingual84.4%New for 4.8; strong multilingual code performance
MCP-Atlas82.2%+4.9 vs Opus 4.7; tool-calling over MCP
BrowseComp (single-agent)84.3%+5.0 vs Opus 4.7
HLE (with tools)57.9%+3.2 vs Opus 4.7
HLE (without tools)49.8%+2.9 vs Opus 4.7
GPQA Diamond93.6%-0.6 vs Opus 4.7; near-saturated, variance expected
USAMO 202696.7%+27.4 vs Opus 4.7; largest single jump
GDPval-AA (Elo)1,890Clean lead over GPT-5.5 (1,769) and Gemini 3.1 Pro (1,314)
GraphWalks BFS 1M68.1vs 40.3 on Opus 4.7; massive long-context retrieval gain

Anthropic's GDPval-AA score of 1,890 Elo is notably higher than K3's 1,668 — suggesting Opus still leads on knowledge-work quality. And Opus' 69.2% on SWE-bench Pro (the harder variant) and 88.6% on SWE-bench Verified are independently verified scores that K3 hasn't yet matched in public.

Radar: four shared evidence points

Term-BenchFrontierSWEDeepSWEGDPval-AA

● Kimi K3 ● Opus 4.8

The radar uses four rows where both models have scores in Moonshot's comparison table. Scores are normalized: Term-Bench and FrontierSWE to 100, DeepSWE to 100, GDPval-AA to 2000. It is a visual aid, not a new benchmark.

  • K3 leads on all four selected rows in Moonshot's table.
  • The gap is widest on FrontierSWE (81.2 vs 66.7) — repository-level engineering.
  • GDPval-AA shows a meaningful agentic quality gap in K3's favor within this table, though Anthropic's own GDPval-AA score for Opus 4.8 is 1,890.
  • Opus 4.8's advantages lie outside this polygon: adjustable effort, fast mode, mature tooling, independent verification, and SWE-bench Pro leadership.

Independent verification: what the public leaderboards show

LeaderboardKimi K3Claude Opus 4.8Notes
Artificial Analysis Intelligence Index~57~56K3 edges ahead; both behind Fable 5 (~60) and GPT-5.6 Sol (~59)
Terminal-Bench 2.1 (public)Not listed78.9% (Claude Code, #5)K3's 88.3% uses KimiCode harness, not yet on public board
MCP-Atlas (public)Not listed82.2% (#3)K3 not yet on BenchLM MCP Atlas leaderboard
SWE-bench Verified (public)Not listed88.6% (#3)K3 not yet on public SWE-bench leaderboards
Frontend Code Arena#1Not in top ranksK3 jumped 17 places from K2.6's #18; leads 6 of 7 domains
GPQA Diamond (open-weight)93.5%93.6% (closed)K3: strongest open-weight GPQA result at launch

The independent picture is mixed. K3 leads the Frontend Code Arena and posts the best open-weight GPQA Diamond score. But on established public leaderboards — Terminal-Bench, MCP-Atlas, SWE-bench — Opus 4.8 has verified, traceable scores while K3 has not yet appeared. This doesn't mean K3's scores are wrong; it means they haven't been independently reproduced yet.

Pricing: K3 costs 40% less at list price

Per 1M tokensKimi K3Claude Opus 4.8
Fresh input / cache miss$3.00$5.00
Cached input / cache hit$0.30$0.50
Output$15.00$25.00
10M fresh input + 1M output$45.00$75.00
10M cached input + 1M output$18.00$30.00
Standard fresh-token example: 10M input + 1M output
K3 official pricing page, Anthropic official pricing page. Tool fees and retries excluded.
K3
$45.00
Opus 4.8
$75.00

K3 is cheaper on every token class. But there's a catch: K3 always runs at max reasoning effort, which means reasoning tokens are billed as output at $15/M. Opus lets you dial effort down (or up) and offers fast mode at $10/$50 for latency-sensitive work. The real cost comparison is cost per accepted task, not cost per token — if Opus succeeds in one shot while K3 needs three, the rate card advantage evaporates.

Artificial Analysis measured per-task cost: K3 at $0.94, Opus 4.8 at $1.80, and GPT-5.6 Sol at $1.04. K3's token efficiency (fewer tokens per correct answer) partially offsets its higher per-token price versus its predecessor K2.6.

Open weights: the strategic advantage (with practical caveats)

K3: open weights by July 27

Moonshot has committed to releasing full weights under a Modified MIT license by July 27, 2026. At 2.8T parameters with 16 active experts, self-hosting requires serious infrastructure — Moonshot recommends at least 64 accelerators. Most teams will use an inference provider.

The MIT license enables fine-tuning, air-gapped deployment, auditability, and sovereignty that closed models cannot match. For regulated industries or organizations that cannot send data to external APIs, this is a decisive advantage.

Opus 4.8: API-only, but battle-tested

Opus 4.8 is closed and API-only — no weights, no self-hosting, no fine-tuning. What you get instead: mature SDK support across every major language, prompt caching, adjustable effort controls, fast mode, computer use, and a production record spanning nearly two months.

Anthropic's tooling ecosystem (Claude Code, MCP, dynamic workflows with parallel sub-agents) is tightly integrated with Opus. For teams that want a managed, reliable API without infrastructure headaches, this is the pragmatic choice.

Which model should you use?

Choose Kimi K3 if…

  • You need the strongest open-weight model available and can wait for weights (or use the API now).
  • Your primary workload is frontend coding — K3 is #1 on the Frontend Code Arena.
  • API cost is a first-order constraint and you can validate K3's output quality on your tasks.
  • You need air-gapped deployment, fine-tuning, or data sovereignty that closed APIs cannot provide.
  • You want to run high-volume, measurable coding tasks with clear pass/fail criteria.

Choose Claude Opus 4.8 if…

  • You need a proven, independently verified model with public leaderboard traces.
  • Your workflow benefits from adjustable reasoning effort — dial up for hard problems, down for routine ones.
  • You rely on Anthropic's tooling: Claude Code, MCP, computer use, dynamic workflows with parallel sub-agents.
  • Judgment, recovery from errors, and production reliability matter more than raw benchmark scores.
  • You want fast mode (2.5× speed) for latency-sensitive applications.
Recommended routing: use K3 as the high-volume default for bounded, verifiable coding tasks where its benchmark lead and 40% price advantage translate to real savings. Route ambiguous, high-stakes, or judgment-heavy work to Opus 4.8. Measure cost per accepted task, not cost per token — and run an identical real-world test on both before committing to either.

Compare both models on real code

Run the same repository prompt, tool schema, and acceptance tests rather than trusting a leaderboard headline.

Try both in CodingFleet →

Final verdict

Kimi K3 wins the benchmark headline contest. On Moonshot's launch table, it leads every coding row against Opus 4.8, sometimes by wide margins. It's #1 on the Frontend Code Arena, posts the strongest open-weight GPQA Diamond score, and costs 40% less at list price. For teams that can validate its output on their own tasks, it's an extraordinarily compelling option — especially once the weights land.

Claude Opus 4.8 wins the operational maturity contest. Its scores are independently verified on public leaderboards. It offers adjustable reasoning effort, a 2.5× fast mode, mature SDK support, and nearly two months of production hardening. On SWE-bench Pro (69.2%), MCP-Atlas (82.2%), and GDPval-AA (1,890 Elo), it posts strong, traceable results that K3 hasn't yet matched in public.

The honest assessment: K3 is the most exciting open-weight release of 2026 and a genuine threat to closed-model dominance. But "exciting" and "production-ready" are different standards. Test both on your actual workload. The model that finishes your work with the least total friction is the right one.

Sources and methodology