Kimi K3 is Now the Largest Open-Source AI Model at 2.8 Trillion Parameters: What Heavy Users Need to Know About API Pricing

On July 28, 2026, Moonshot AI released Kimi K3, the first open-source model to reach 2.8 trillion parameters. Now available via the Telnyx Inference API, it marks a turning point: open-source models have caught up to closed-source frontier labs on coding, reasoning, and agentic benchmarks. For heavy AI users spending thousands per month on API calls, this changes the cost calculus.

What Kimi K3 brings to the table

Kimi K3 is built on Kimi Delta Attention and Attention Residuals architectures, with a 1 million token context window and native vision capabilities that accept text, images, and video without a separate adapter. It supports configurable reasoning effort (low, high, max), tool calling, structured JSON output, and automatic prompt caching.

The benchmark results are the headline: K3 competes with closed-source frontier models from Anthropic and OpenAI on coding, reasoning, and agentic knowledge work. For heavy AI users, the critical question is whether open-source APIs have finally reached the point where you can substitute them for Claude Opus 5 or GPT-5.6 Terra without sacrificing quality.

Kimi K3 API pricing breakdown

On Telnyx Inference, Kimi K3 pricing is set at:

• Cached input: $0.27 per million tokens
• Input: $2.70 per million tokens
• Output: $13.50 per million tokens

Compare this to the current landscape for heavy AI users:

Model	Input (per MTok)	Output (per MTok)
Kimi K3 (Telnyx)	$2.70	$13.50
Claude Opus 5 (Anthropic)	$5.00	$25.00
GPT-5.6 Terra (OpenAI)	$2.50	$15.00
Gemini 3.6 Flash (Google)	$1.50	$7.50
Claude Sonnet 5 (intro pricing)	$2.00	$10.00
GPT-5.6 Luna (OpenAI)	$1.00	$6.00

Kimi K3 slots in above the budget tiers but well below the top-tier frontier pricing. At $2.70 input and $13.50 output, it costs roughly half of Claude Opus 5 ($5/$25) and about the same as GPT-5.6 Terra ($2.50/$15). For heavy users running large-scale agentic workloads, the savings versus Opus 5 amount to 46% on input and 46% on output.

But raw per-token pricing is only part of the story. The real cost advantage depends on three factors: prompt caching effectiveness, reasoning effort configuration, and whether the model’s quality meets your threshold.

Where Kimi K3 saves heavy users real money

Prompt caching by default. Telnyx Inference applies automatic prefix caching for repeated prompt prefixes across requests. For heavy AI users running batch processing, automated testing pipelines, or agent loops that reuse large system prompts, this effectively cuts input costs to the cached rate of $0.27 per million tokens — a 90% discount versus uncached input. In practice, teams running Claude Opus 5 with long system prompts could see their per-request costs drop by 40-60% by switching to K3 with caching enabled.

Configurable reasoning effort. The model ships with three reasoning levels (low, high, max) that let you trade compute for depth. For simple extraction tasks or classification, the low setting delivers adequate quality at a fraction of the token cost. For complex code generation or multi-step reasoning, max effort matches frontier quality. This is analogous to OpenAI’s reasoning effort parameter but available on an open-weight model — meaning you can self-host and bypass API markup entirely.

No data retention premium. Unlike Claude Fable 5 which imposes a 30-day data retention requirement for certain use cases, K3 on Telnyx carries no such contractual overhead. For teams processing sensitive codebases or proprietary data, this removes a compliance cost that can add thousands in legal and audit fees annually.

The catch: inference infrastructure costs

Kimi K3 is not a model you can run on consumer hardware. At 2.8 trillion parameters, it requires substantial GPU infrastructure. The Telnyx Inference API abstracts this, but the $2.70 per MTok input price reflects the real compute cost of serving a model of this scale. For comparison, a 70B parameter model running on the same infrastructure would cost roughly 40x less to serve.

Heavy users should also factor in that Telnyx Inference is a relatively new platform compared to Anthropic’s or OpenAI’s battle-tested APIs. Rate limits, availability guarantees, and latency SLAs are still evolving. For production workloads that require 99.9% uptime, K3 via Telnyx may work best as a secondary provider for cost-sensitive traffic rather than a primary inference backbone.

Open-source vs closed-source: the pricing gap narrows

The open-source AI model landscape has shifted dramatically in 2026. Fermion’s Neutrino-1 8B (released July 27) showed that small, efficient models can run locally at a fraction of API cost. Kimi K3 now demonstrates that open-source models at the absolute frontier are also viable through third-party API providers at competitive rates.

For heavy AI users, this means the pricing gap between open-source and closed-source APIs is narrowing. Where open-source APIs were once 5-10x cheaper but significantly less capable, K3 closes the quality gap to within striking distance of Opus 5 and GPT-5.6 Terra. The remaining advantages of closed-source providers are reliability guarantees, mature tooling, and ecosystem integration — not raw model quality.

What this means for your AI budget

For a heavy AI user spending $5,000 per month on Claude Opus 5 API calls, switching primary workloads to Kimi K3 on Telnyx could reduce costs to roughly $2,700 per month — a 46% savings. Adding prompt caching pushes that further to around $1,800.

The practical strategy for most heavy users will be a tiered approach: use Kimi K3 via Telnyx for high-volume, cache-friendly workloads where output quality requirements are moderate; reserve Claude Opus 5 or GPT-5.6 Sol for the hardest reasoning tasks where every percentage point of benchmark improvement matters.

Telnyx makes this easy with its OpenAI-compatible API, meaning existing code that targets the OpenAI format can switch providers with a single endpoint change. No SDK migration, no prompt reformatting, no workflow disruption.

The bottom line for heavy AI users

Kimi K3 at $2.70/$13.50 per million tokens is the strongest argument yet that open-source AI models can compete on both quality and price with closed-source frontier providers. For heavy AI users, it opens a new middle tier: near-frontier intelligence at roughly half the price of top-tier APIs.

The decision to adopt K3 comes down to your quality threshold and your tolerance for working with a newer inference provider. For teams that can accept a minor quality gap in exchange for significant cost reduction, the math is compelling. For mission-critical production systems that demand every point of benchmark performance, the premium for Opus 5 or GPT-5.6 Sol remains worth paying — but the gap has never been smaller.