July 26, 2026ai-infrastructurellmfinopsopen-weight-modelsmodel-selection

Open-Weight Isn't Cheap and Closed Isn't Doomed: The AI Price War by the Numbers

Kimi K3 priced itself like a frontier model, two leaderboards disagree by twenty ranks on the same product, and committed GPU capacity got more expensive while spot prices fell. What the token economics actually say about build vs buy.

Eight days. That is how long it took Anthropic to answer Kimi K3.

On July 16, Moonshot AI dropped Kimi K3, a 2.8-trillion-parameter model, live in the API within hours. On July 24, Anthropic shipped Claude Opus 5, pitched as close to frontier intelligence at half the price of its own top tier. Anthropic's marketing never mentioned Moonshot. The press did it for them.

I spent a few days pulling primary sources (official pricing pages, Hugging Face model cards, arXiv papers, Artificial Analysis, LMArena, METR) to see what the economics say rather than what the headlines imply. Some of it confirmed the obvious story. A lot of it did not.

The pricing spread is wider than most people think

Blended input and output, per million tokens, roughly:

Claude Opus 5: $5 / $25 ($3.85 blended)
Claude Sonnet 5: $3 / $15 standard, discounted to $2 / $10 through August 31
Kimi K3: $3 / $15 ($2.31 blended), a real jump from its predecessor K2.6 at $0.95 / $4
DeepSeek V4 Pro: $0.435 / $0.87 ($0.18 blended)
Qwen3.5-397B: around $0.45 / $3.00, depending on host

DeepSeek V4 Pro is not a rounding error cheaper than Opus 5. It is about 21 times cheaper per blended token, for a model that Artificial Analysis scores about 17 points lower on its intelligence composite (44 vs. 61). Divide points by dollars and DeepSeek comes out roughly 15 times more cost-efficient. That is a rough proxy, not a published metric, but it frames the trade honestly.

Open-weight and cheap are not the same axis

Kimi K3 breaks the pattern. Moonshot priced it at about 60% of Claude Opus 4.8's rate, comparable to half of GPT-5.6's. By the same points-per-dollar proxy, K3 is less cost-efficient than Claude Sonnet 5 and less efficient than its own predecessor. It is a top-10 model on raw intelligence and one of the more expensive open-weight releases on the market. Check the per-token price, not the license.

There is a second wrinkle. At launch, K3's weights were not fully public: API first, full weights promised days later. That is the closed-model playbook wearing an open-weight label. Categories are getting blurry, and procurement processes built on the labels will mislead you.

Benchmarks disagree with each other, and that is the finding

Claude Sonnet 5 sits at #16 on Artificial Analysis's Intelligence Index. On LMArena's head-to-head Elo leaderboard it ranks 38th, well behind Kimi K3 and Qwen3.7. Two reputable leaderboards, same week, overlapping models, disagreeing by more than twenty ranks on the same product.

That is not a gotcha. It is the lesson. Elo captures human preference in chat, composites weight agentic and reasoning tasks, METR measures autonomous task duration. None of them is "the" score. If your vendor selection picks a model off one leaderboard number, you are one benchmark's idiosyncrasy away from a bad decision. Run your own eval on your own workload. It is the only number that applies to you.

The fine-tuning lever the pricing table hides

You can tune Llama, Mistral, Qwen, or DeepSeek. You cannot touch Claude's or GPT's weights. On a narrow task, tuned often beats big. Meta's Code Llama 70B-Instruct hit 67.8% pass@1 on HumanEval, edging GPT-4's reported 67.0%. DeepSeek-Coder-V2-Instruct scored 90.2% against GPT-4-Turbo's 88.2%. Predibase's LoRA Land project tuned 25 Mistral-7B adapters across 31 tasks and averaged 0.756 versus GPT-4's 0.661. A 7-billion-parameter model, tuned narrow, beat a much larger generalist.

None of this makes open-weight models smarter in general. It makes them adjustable. For a well-defined task, adjustable beats bigger.

The GPU story LinkedIn gets wrong

The comfortable narrative: GPU prices are crashing, so self-hosting is a free lunch. Half true.

Spot and on-demand H100 pricing has genuinely fallen 64 to 75% since 2023. But committed capacity, the kind production inference actually needs, moved the other way: SemiAnalysis's H100 rental index shows one-year contract pricing up about 40% between October 2025 and March 2026, from $1.70 to $2.35 per GPU-hour. The chips are not the constraint. Allocated capacity is.

Sticker price was never the number that mattered anyway. One TCO analysis puts a 100-H100 cluster's real five-year cost at $8.6M against a $3M hardware price tag, a 2.9x multiplier for power, cooling, and the MLOps staff who keep it alive. Breakeven analyses converge on a practical floor: around 60% sustained utilization and roughly 5 to 7 billion tokens a month through a single 8-GPU node before self-hosting beats a frontier API. Below that volume, "we saved money self-hosting" quietly becomes "we paid to run a smaller data center."

What holds up for a build-vs-buy decision

Frontier closed models stay the default for volatile, quality-sensitive workloads. You are paying for years of RLHF and tool-use training, and for someone else owning the capacity-planning headache.
Open-weight models earn their keep under specific conditions: predictable high-volume traffic, data-residency rules, or in-house MLOps deep enough to beat the 60% utilization floor.
Open-weight does not mean cheap. K3 proved that. Price the token, not the license.
Do not trust a single leaderboard. Pull several, expect them to disagree, then test on your own workload anyway.
Watch the geopolitics. Export controls, distillation disputes, and sovereign-AI mandates can change your vendor's roadmap faster than any benchmark moves. This is supply-chain risk, not dinner conversation.

The frontier-vs-open-weight debate gets framed as ideology. The economics underneath are more useful: a utilization problem, a benchmark-methodology problem, and a supply-chain problem, wearing an AI costume. Model your workload before you pick a side.

Figures verified against official pricing pages, Hugging Face model cards, and arXiv papers, cross-checked against Artificial Analysis, LMArena, METR, and SemiAnalysis as of July 26, 2026. A snapshot of a fast-moving situation, not a settled record.