Kimi K3 Replaces GPT-5.6 Sol in helloai's Frontier Set
Moonshot's 2.8T Kimi K3 lands on the public API at $3/$15 with 1M context. helloai drops GPT-5.6 Sol and adds Moonshot under the six-model cap.
On July 16, Moonshot AI shipped Kimi K3 — a 2.8-trillion-parameter mixture-of-experts model with native multimodal input, a 1,048,576-token context window, and a public Moonshot API priced at $3 per million cache-miss input tokens and $15 per million output ($0.30 on cache hits). helloai admits Kimi K3 as its sixth frontier model and drops GPT-5.6 Sol to stay inside the six-model cap. The swap is pure set-size math: Sol was the lowest-Elo tracked entry at 1484; K3 posts about 1486 on LMArena text overall with a new lab and a different cost-quality curve.
The capability pitch is long-horizon coding and end-to-end knowledge work, not a pure arena top-five story. Official docs describe Kimi Delta Attention and Attention Residuals, with only 16 of 896 experts active per token, plus vision and video inputs on the same model ID (kimi-k3). Thinking is always on, with reasoning_effort currently limited to max. Open weights are promised by July 27, so today's helloai entry is the metered API — not a local-run open-weight card. That distinction matters: K3 may eventually show up in open_weight_models.json after a single-GPU story exists; for now it competes with closed frontier APIs.
Economics sit between Muse Spark 1.1 and the Western flagships. Muse still undercuts everyone at $1.25/$4.25. Kimi K3's $3/$15 list rate is cheaper on input than Opus 4.8 and GPT-5.6 Sol, and cache-hit traffic at $0.30 per million tokens is the real lever for agent loops that resend long prefixes. Output remains expensive relative to Grok 4.5's $6, so teams that burn reasoning tokens should profile before treating K3 as a default cost win. Flat pricing across the full 1M window — no long-context surcharge tier — is a quieter advantage for document and repo-scale jobs.
Dropping GPT-5.6 Sol is the hard part of the decision. Sol still leads many agentic coding slices, including Terminal-Bench 2.1 numbers OpenAI published at launch, and OpenAI remains the default enterprise contract for a large share of teams. helloai is not saying Sol is weak; it is saying the curated comparison set only holds six rows, and Moonshot's first clear frontier API entry is more informative this week than keeping a second Western coding flagship that already has Fable, Opus, and Grok covering adjacent ground. Teams standardized on OpenAI should keep Sol in production routing regardless of the helloai table.
For /api/recommend, the stack is now Anthropic (Fable 5 and Opus 4.8), Google, xAI, Meta, and Moonshot. Muse Spark keeps the budget-agentic floor; Kimi K3 fills the mid-price multimodal long-context slot that neither Muse nor Grok fully owns. The next checkpoint is July 27 weight release and whether arena votes hold once the preliminary K3 sample matures. Until then, helloai tracks the public API Moonshot is selling today — and leaves OpenAI off the six-model board for the first time since the directory launched.