Kimi K3 Shares a Four-Way Lead in Vercel’s Moving Next.js Eval

• Kimi K3 with OpenCode passed 22 of 24 base tasks, tying three other model-agent pairings at a displayed 92% rather than winning outright.
• Bundled Next.js documentation lifted K3 to 23 passes, but 13 of the table’s 25 pairings reached the same 96% mark with that material.
• K3’s API is cheaper per token than the two proprietary models Moonshot identifies as stronger overall, but Vercel publishes no token or dollar cost per successful task.
• Moonshot promises downloadable weights by July 27 and recommends at least 64 accelerators for efficient inference, making independent deployment an infrastructure decision.

Kimi K3 has joined the leaders in a practical Next.js coding test. The result strengthens Moonshot AI’s claim to frontier-level coding performance, but it measures a particular model-agent stack on a small, changing suite—not K3 in isolation or the economics of using it in production.

Vercel’s company-reported average durations for the four model-agent pairings displayed at 92% success in its July 17 Next.js evaluation snapshot. Source: Next.js by Vercel.

Four stacks share 22 passes

The July 17 evaluation snapshot gives Kimi K3 with OpenCode a 92% success rate on Next.js code-generation and migration tasks. Claude Fable 5 with Claude Code, Cursor Composer 2.5 with Cursor, and GPT 5.6 Sol with Codex also show 92%. The corresponding result commits record 22 passes from 24 tasks for each pairing.

The pass counts for the four leaders come from the retained result commits; the public page displays rounded percentages. K3 was neither the sole success-rate leader nor the fastest co-leader. Cursor’s pairing had the lowest published average duration.

The comparison does not separate model capability from the surrounding agent. Each leader used a different harness, and the K3 commit says its configuration used four runs, early exit and a 1,200-second timeout. The public table provides no variance, distribution of run times or rule for translating those repeated attempts into a purchasing estimate.

An apparent generational gain is also hard to quantify cleanly. Kimi K2.7 Code appears at 75% with the same OpenCode agent, but that result was committed a month earlier against an earlier pinned suite. Vercel says reworked tests retain a model’s previous measurement until it is rerun. Subtracting the two rounded rows would therefore mix model generations, test states and dates.

Documentation compresses much of the field

The commit record says K3 failed the base proxy task and the instant task. Supplying an AGENTS.md file with bundled Next.js documentation cured the proxy failure, raising K3 from 22 to 23 passes; the instant task remained unsolved.

That one additional pass moves the rounded display from 92% to 96%. It does not create separation: 13 of the 25 rows on the public table show 96% with the documentation, including models whose base results range from 58% to 92%. GPT 5.6 Sol is a useful counterexample to any claim that the documentation always helps: its commit says the material fixed one task but another task regressed, leaving it at 22 of 24.

The result is evidence that current framework instructions can change outcomes on this suite. It is not evidence that documentation erases general model differences elsewhere, because the table tests only Next.js work and preserves some older measurements.

The open repository improves auditability. Each task is a self-contained project; the agent receives a prompt and editable source files, while the test assertions are withheld until the run ends. The runner memoizes completed model-task pairs and ordinarily runs only missing pairs. Non-model failures such as infrastructure errors and timeouts are deleted during evaluation runs before results are exported.

The repository also exposes why the leaderboard should be read as a maintained engineering dashboard. Its scraped README enumerates 20 “current evals,” while the result commits score 24. The history records judge conversions, keeps older measurements for changed tests, and shows that earlier GPT 5.6 Sol entries were removed after invalid model identifiers fell back to other metadata. A replacement run then produced the current 22-of-24 row. Those corrections make the board more credible over time, but less like a simultaneous frozen tournament.

The API discount is not yet a cost-per-task result

Moonshot says in its technical post that it charges K3 API users $3 per million cache-miss input tokens, $15 per million output tokens and $0.30 per million cache-hit input tokens. The company says coding workloads on its own service achieve cache-hit rates above 90%, but that is a company claim about its serving system, not a measurement from Vercel’s test.

A retained market and pricing analysis lists Claude Fable 5 at $10 per million input tokens and $50 per million output tokens, and GPT 5.6 Sol at $5 and $30. On those rate cards, K3 output is 70% cheaper than Fable and 50% cheaper than Sol. K3 is not the low-price option across open models, however: the same analysis lists Kimi K2.7 Code at $4 per million output tokens and GLM 5.2 at $4.40.

None of those sticker-price comparisons establishes the cheapest completed task. Models can consume different numbers of input, reasoning and output tokens, use caches differently, retry at different rates and produce different failure costs. Vercel reports duration but not token use, tool calls, cache hits or dollars per pass. The current evidence supports “lower API price per token than two frontier rivals,” not “lower production cost.”

Moonshot also has more financing capacity than the image of a small challenger suggests. The same analysis describes the Beijing company as founded in 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin, with Alibaba and Tencent among its backers. Citing May reporting, it says Moonshot raised about $2 billion at a roughly $20 billion valuation and had annual recurring revenue above $200 million by April. That context makes aggressive pricing strategically plausible, but it does not reveal K3’s serving margin or prove the API is subsidized.

A narrow coding result cannot settle reliability

Moonshot describes K3 as a 2.8-trillion-parameter mixture-of-experts model with a one-million-token context window and native vision. The company also explicitly says K3 still trails Claude Fable 5 and GPT 5.6 Sol overall and has a noticeable user-experience gap. Its benchmark notes mix KimiCode, Claude Code and Codex harnesses depending on the model and test, reinforcing the need to preserve pairing details.

Moonshot identifies two operational limitations. K3’s quality can become unstable if a harness does not return its full preserved thinking history or if a session switches to K3 midway. It also says the model can make unexpected decisions when instructions are ambiguous and recommends explicit behavioral constraints. The successful OpenCode result shows that pairing worked on Vercel’s suite; it does not establish compatibility or safe autonomy across other agents and workloads.

A launch report gives a further reason not to generalize from coding passes: it cites a 51% hallucination rate for K3 on AA-Omniscience, up from 39% for K2.6. That single reported benchmark is not a complete reliability assessment, but it directly contradicts the idea that higher capability necessarily means fewer fabricated answers.

The word “open” also remains prospective at the snapshot. K3 is available through Moonshot’s hosted products and API, while the company says full weights will arrive by July 27. Moonshot recommends supernode configurations with 64 or more accelerators for efficient inference. Releasing weights would expand control over customization and deployment; it would not make a 2.8-trillion-parameter model economical for ordinary local hardware.

What would turn the tie into a buying signal

The next decision is not which logo sits first on a rounded table. It is whether the K3 stack delivers reliable work at a lower total cost under reproducible conditions.

Vercel could make that decision easier by freezing one 24-task version, rerunning every contender at the same revision and publishing the complete model, agent, effort, timeout and repeated-run settings. A cross-over design—testing more than one agent with each compatible model—would show how much of the result belongs to the model and how much to the harness. Token use, cache hits, tool calls, retries and dollars per successful task would turn rate cards into comparable deployment economics.

For K3, the July 27 weight release is the next control point. Independent operators would need to verify the artifact and license, disclose hardware and serving configurations, reproduce the coding results, and test the reliability limitations Moonshot and the launch reporting identify. Until those checks exist, K3 has earned a place among the Next.js leaders. The evidence supports a four-way tie between stacks, not a durable crown or a proven cost advantage for one model.