Intelligence per Dollar: The Metric Quietly Reshaping AI Benchmarks

TL;DR

For the first time this year, buyers aren’t asking which AI model scores highest — they’re asking which one delivers the most capability per dollar. The clearest proof: Zhipu’s open-weights GLM-5.2 lands within a point of Anthropic’s Claude Opus 4.8 on several agentic benchmarks while costing roughly a fifth as much per token, and it shipped in the same week the U.S. government temporarily cut off access to two of Anthropic’s own frontier models.

The launch that started the conversation

Z.ai — the international brand of Beijing-based Zhipu AI — released GLM-5.2 on June 13, 2026, with full open weights following under an MIT license days later. It’s a 744-billion-parameter mixture-of-experts model that activates roughly 40 billion parameters per token, and anyone can download it and run it on their own servers.

The price is the headline. GLM-5.2’s API runs $1.40 per million input tokens and $4.40 per million output tokens, against Claude Opus 4.8’s $5 and $25. That’s roughly 3.6x cheaper on input and 5.7x cheaper on output.

The benchmark gap barely justifies that price gap. GLM-5.2 trails Opus 4.8 by about a point on FrontierSWE and nearly ties it on the MCP-Atlas tool-use test. It’s about four points behind on Terminal-Bench 2.1, and it beats Opus outright on olympiad-style math tests like AIME 2026. On Artificial Analysis’s Intelligence Index, GLM-5.2 scored 51 — the top open-weights model on the board, and among the very top scores overall, behind only closed frontier systems that cost several multiples more.

Braintrust ran its own production test: an eval built with inference partner Baseten that measured exact long-context retrieval over the CPython standard library. Opus 4.8 came out slightly more accurate, but GLM-5.2’s cost per trace ran 76–78% lower across both context tiers tested — meaning Opus cost roughly four to four-and-a-half times as much per trace for that small accuracy edge.

Pricing and index scores as reported by Artificial Analysis, Braintrust, and vendor announcements in June–July 2026; figures shift as providers update rates.

Why the timing wasn’t a coincidence

GLM-5.2 shipped one day after a U.S. export-control order forced Anthropic to cut off access to its newest models, Claude Fable 5 and Claude Mythos 5, for every foreign national — including its own non-U.S. employees — starting June 12, 2026. OpenAI separately limited access to its GPT-5.6 models around the same time, citing a similar government request.

Both situations have since eased. The Department of Commerce lifted the export controls on June 30, and Anthropic restored global access to both Fable 5 and Mythos 5 on July 1.

But the three-week disruption left a mark. Harvey co-founder Gabe Pereyra told CNBC he’d “been consistently surprised by how quickly the open source has caught up.” The point industry voices keep repeating since: a model that can be revoked by a single regulatory order carries a risk a self-hosted, open-weights alternative doesn’t. For enterprises building multi-year AI roadmaps, that’s now a line item next to the price tag — not a footnote below it.

The pattern goes beyond one model

Microsoft made a similar point with its own numbers in June. As analyst Tomasz Tunguz noted, Microsoft’s MAI-Code-1-Flash announcement included a metric rarely seen on a model card before: average tokens burned per task. The model hit 71.6 on SWE-Bench Verified using roughly a third of the tokens Claude Haiku 4.5 needs for similar results, plus a 16-point lead on SWE-Bench Pro.

Artificial Analysis’s own data tells the same story at the top of the market. GPT-5.5 and Claude Opus 4.8 land within a point of each other near 60 on the Intelligence Index — but running the full evaluation costs $3,357 on GPT-5.5 against $4,685 on Opus 4.8, a 40% price gap for essentially the same output.

That’s the calculation buyers are now running on every purchase: is the extra point of intelligence worth the extra dollar.

Which model, for which job

• Solo developers and small teams running high-volume agent loops are drifting toward cheap, high-context options like DeepSeek V4 Pro, which keeps costs down by activating only a fraction of its parameters per token.
• Enterprises running firm-wide deployments, where consistency matters more than marginal savings, are still defaulting to GPT-5.5 Pro or Claude Opus 4.7 — because one inconsistent output in a regulated workflow can cost more than the tokens it would have saved.
• Teams building latency-sensitive, real-time products are choosing on raw speed: Mercury 2 currently leads throughput at up to 859 tokens per second, intelligence score aside.
• Simple classification and routing tasks inside a larger pipeline are increasingly handled by near-free models like Gemma 3n and Amazon Nova Micro — jobs frontier models were never needed for in the first place.

The takeaway that outlasts this news cycle

GLM-5.2 will get replaced by something else within months. So will Opus 4.8, GPT-5.6, and every other model named here — that part is guaranteed. What won’t change is the question buyers now ask before every AI purchase: not “which model is smartest,” but what a unit of finished work actually costs, end to end.