If you only read one piece on AI economics this quarter, this is the one to bookmark.
The state of LLM pricing as of May 2026 is not a price. It's a 442× spread across more than 325 production models, with quality scores 79–100 all available at radically different price points. The spread is not a temporary market dislocation — it is the new structure of the market.
gpt-5.2-pro ($168/M, quality 96) and deepseek-v3.2 ($0.38/M, quality 79). Two models, one market.Live snapshot from CostGoat's LLM pricing tracker, updated May 7, 2026 (325+ models tracked). Selected entries by quality tier:
| Model | Quality | Input ($/M) | Output ($/M) | Value Score |
|---|---|---|---|---|
claude-opus-4.6 | 100 | $5.00 | $25.00 | 4.0 |
gpt-5.2 (standard) | 96 | $1.75 | $14.00 | 6.9 |
gpt-5.2-pro | 96 | $21.00 | $168.00 | 0.6 |
glm-5 (Z-AI) | 94 | $0.60 | $1.92 | 49.0 |
kimi-k2.5 | 89 | $0.44 | $2.00 | 44.5 |
gemini-3-flash-preview | 87 | $0.50 | $3.00 | 29.0 |
claude-sonnet-4.5 | 81 | $3.00 | $15.00 | 5.4 |
deepseek-v3.2 | 79 | $0.25 | $0.38 | 209.0 |
minimax-m2.5 | 79 | $0.15 | $1.15 | 68.7 |
Quality scores are normalized benchmarks; value scores are inverse-cost-per-quality-point. Higher value score = more output quality per dollar.
$168/M (gpt-5.2-pro) → $0.38/M (deepseek-v3.2) = 442× output price spread.
The two models sit at quality scores 96 and 79 respectively. That's a meaningful difference for a complex reasoning workload — but for the majority of production traffic (summarization, classification, extraction, formatting, reformatting), quality 79 is operationally indistinguishable from quality 96. And the spread is 442×.
gpt-5.2 ($14/M, Q96) vs glm-5 ($1.92/M, Q94) = 7.3x spread at functionally equivalent quality. That's $4,400/year saved per million daily output tokens — multiply by your real volume.The more actionable number for most readers is the spread between models that are genuinely interchangeable:
That 7.3× is the headline number for the "same job, different vendor" routing case. A workload running 1,000,000 output tokens per day on GPT-5.2 costs $5,110/year. The same workload on GLM-5 costs $700/year — a $4,400/year savings per million tokens of daily output. Multiply by your actual production volume.
Looking at the full catalog, the market has settled into three clear bands:
Nano tier (under $1/M output). MiMo-V2-Flash ($0.29/M, quality 77), gpt-oss-120b ($0.18/M, quality 62), DeepSeek V3.2 ($0.38/M, quality 79), MiniMax-M2.5 ($1.15/M, quality 79). Open-weight or near-open-weight. Quality range 45-79. Excellent for high-volume routine work.
Mid tier ($1-5/M output). GLM-5, Kimi K2.5, Gemini 3 Flash, Claude Sonnet 4.5. Quality 81-94. The workhorse band — most enterprise traffic should land here.
Premium tier ($5-30/M output). GPT-5.2 standard, Claude Opus 4.6, GPT-5.5. Quality 96-100. Reserve for genuinely complex reasoning, agentic chains where retries are expensive, and high-stakes user-facing output.
Gated tier ($25-180/M output). GPT-5.2 Pro, GPT-5.5 Pro, specialized vendors. Quality 96-100, but value scores below 1.0. Use only when no alternative meets the latency or context-window constraint.
glm-5, kimi-k2.5, minimax-m2.5, mimo-v2-flash, and OpenAI's own gpt-oss family — widening the floor, not raising it. The cheap end now includes frontier-quality options.Six weeks of new entries on the tracker:
There are now ~10 models in the catalog at quality 80+ that cost less than $2/M output. There are ~5 models at quality 90+ that cost less than $3/M output. The "we have to use the most expensive model" argument is no longer defensible without specific evidence of why the cheaper alternatives fail for that specific workload.
The procurement question is not "which AI provider should we standardize on?" — it's "for which workloads does the premium tier actually justify itself, and how do we route everything else automatically?"
If your AI strategy treats LLM pricing as a single number, you are operating with a 2023 mental model in a 2026 market. The 442× spread is not going to compress. New entrants from China (GLM-5, Kimi, MiniMax, MiMo) and OpenAI's own open-weight push (gpt-oss) are widening the floor, not raising it.
The math has been done. The infrastructure to capture it has been built. The only question is whether your organization is using it.
Trimio is the LLM API gateway built for AI cost governance. We track the full LLM pricing catalog in real time and route every request to the right tier automatically. See how it works.